{"id":3408,"date":"2026-04-28T01:20:26","date_gmt":"2026-04-27T23:20:26","guid":{"rendered":"https:\/\/neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/"},"modified":"2026-10-01T13:49:32","modified_gmt":"2026-10-01T11:49:32","slug":"neural-designer-in-scientific-research","status":"publish","type":"blog","link":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/","title":{"rendered":"Neural Designer in scientific research"},"content":{"rendered":"<style>\/* nd-blog-width:start *\/.nd-native-article__content{width:min(100% - 72px,1200px)}\/* nd-blog-width:end *\/<\/style>\n<p>The scientific and research community has used Neural Designer widely.<\/p>\n<p>Many articles use this program to support or even fully develop their research.<\/p>\n<p>This page collects peer-reviewed articles published in first-quartile (Q1) journals, according to the <a href=\"https:\/\/www.scimagojr.com\/\" target=\"_blank\" rel=\"noopener\">SCImago Journal Rank (SJR)<\/a>, in which the authors used <a href=\"https:\/\/www.neuraldesigner.com\/\">Neural Designer<\/a> to build, train, or analyze their models and data.<\/p>\n<p>You can download Neural Designer <a href=\"https:\/\/www.neuraldesigner.com\/downloads\/\">here<\/a>.<\/p>\n<section>\n<h3>Contents<\/h3>\n<\/section>\n<h3><strong>Healthcare &amp; Life Sciences<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo20\">Application of machine learning for the prediction of Bacillus cereus growth: A study on the integration of mathematical models<\/a>.<\/li>\n<li><a href=\"#Articulo21\">Integrating Genomic Classifiers and Nonsuspicious Magnetic Resonance Imaging Findings in Predictive Modelling for Lymph Node Metastasis in Patients With Localized Prostate Cancer<\/a>.<\/li>\n<li><a href=\"#Articulo11\">Machine learning modeling for solubility prediction of recombinant antibody fragment in four different E. coli strains<\/a>.<\/li>\n<li><a href=\"#Articulo49\">Pneumonectomy and broncho-pleural fistula: predicting factors and stratification of the risk<\/a>.<\/li>\n<li><a href=\"#Articulo51\">Internet of things-inspired healthcare system for urine-based diabetes prediction<\/a>.<\/li>\n<li><a href=\"#Articulo22\">Enhanced release of acid sphingomyelinase-enriched exosomes generates a lipidomics signature in CSF of Multiple Sclerosis patients<\/a>.<\/li>\n<li><a href=\"#Articulo23\">Predictors of in-hospital mortality following major lower extremity amputations in type 2 diabetic patients using artificial neural networks<\/a>.<\/li>\n<\/ol>\n<h3><strong>Pharmaceutics &amp; Drug Delivery<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo24\">Building Artificial Neural Networks for the Optimization of Sustained-Release Kinetics of Metronidazole from Colonic Hydrophilic Matrices<\/a>.<\/li>\n<li><a href=\"#Articulo25\">Integrating Artificial Intelligence with Quality by Design in the Formulation of Lecithin\/Chitosan Nanoparticles of a Poorly Water-Soluble Drug<\/a>.<\/li>\n<\/ol>\n<h3><strong>Energy &amp; Environment<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo26\">Optimization of Biomass Delivery Through Artificial Intelligence Techniques<\/a>.<\/li>\n<li><a href=\"#Articulo52\">Artificial neural network assisted numerical analysis on performance enhancement of Sb2(S,Se)3 solar cell with SnS as HTL<\/a>.<\/li>\n<li><a href=\"#Articulo27\">Deep learning-based evaluation of photovoltaic power generation<\/a>.<\/li>\n<li><a href=\"#Articulo28\">Design and simulation of a highly efficient CuBi2O4 thin-film solar cell with hole transport layer<\/a>.<\/li>\n<li><a href=\"#Articulo29\">Forecasting meteorological impacts on the environmental sustainability of a large-scale solar plant via artificial intelligence-based life cycle assessment<\/a>.<\/li>\n<li><a href=\"#Articulo30\">Performance analysis and optimization of SnSe thin-film solar cell with Cu2O HTL through a combination of SCAPS-1D and machine learning approaches<\/a>.<\/li>\n<li><a href=\"#Articulo31\">Utilizing machine learning to enhance performance of thin-film solar cells based on Sb2(SxSe1\u2212x)3: investigating the influence of material properties<\/a>.<\/li>\n<li><a href=\"#Articulo32\">AI-coherent data-driven forecasting model for a combined cycle power plant<\/a>.<\/li>\n<li><a href=\"#Articulo14\">N,S co-doped biocarbon for supercapacitor application: Effect of electrolytes concentration and modelling with artificial neural network<\/a>.<\/li>\n<li><a href=\"#Articulo33\">Mineralogical composition and total organic carbon quantification using x-ray fluorescence data from the Upper Cretaceous Eagle Ford Group in southern Texas<\/a>.<\/li>\n<li><a href=\"#Articulo34\">Using a Neural Network to Improve the Optical Absorption in Halide Perovskite Layers Containing Core-Shells Silver Nanoparticles<\/a>.<\/li>\n<\/ol>\n<h3><strong>Chemistry &amp; Materials Science<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo53\">Low-pressure reverse osmosis membrane made of cellulose nanofiber and carbon nanotube polyamide nano-nanocomposite for high purity water production<\/a>.<\/li>\n<li><a href=\"#Articulo12\">Predicting the Properties of High-Performance Epoxy Resin by Machine Learning Using Molecular Dynamics Simulations<\/a>.<\/li>\n<li><a href=\"#Articulo35\">Prediction of Lap Shear Strength and Impact Peel Strength of Epoxy Adhesive by Machine Learning Approach<\/a>.<\/li>\n<li><a href=\"#Articulo19\">Statistical and Machine Learning-Driven Optimization of Mechanical Properties in Designing Durable HDPE Nanobiocomposites<\/a>.<\/li>\n<li><a href=\"#Articulo50\">Metal recovery prediction of elements from anode slime<\/a>.<\/li>\n<li><a href=\"#Articulo36\">Application of DFT-based machine learning for developing molecular electrode materials in Li-ion batteries<\/a>.<\/li>\n<\/ol>\n<h3><strong>Civil Engineering &amp; Construction<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo37\">Machine learning evaluation of PI control effects on neutral equilibrium in bridge virtual pier systems<\/a>.<\/li>\n<li><a href=\"#Articulo38\">Optimization of concrete with human hair using experimental study and artificial neural network via response surface methodology and ANOVA<\/a>.<\/li>\n<li><a href=\"#Articulo39\">Optimization of a coal mine roof characterization model using machine learning<\/a>.<\/li>\n<li><a href=\"#Articulo40\">Enhancing of uniaxial compressive strength of travertine rock prediction through machine learning and multivariate analysis<\/a>.<\/li>\n<li><a href=\"#Articulo41\">Neural network modeling of rutting performance for sustainable asphalt mixtures modified by industrial waste alumina<\/a>.<\/li>\n<li><a href=\"#Articulo42\">Environmental impacts cost assessment model of residential building using an artificial neural network<\/a>.<\/li>\n<li><a href=\"#Articulo43\">Predicting the contribution of recycled aggregate concrete to the shear capacity of beams without transverse reinforcement using artificial neural networks<\/a>.<\/li>\n<li><a href=\"#Articulo16\">Modeling resilient modulus of fine-grained materials using different statistical techniques<\/a>.<\/li>\n<\/ol>\n<h3><strong>Engineering &amp; Manufacturing<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo44\">Tool life prognostics in CNC turning of AISI 4140 steel using neural network based on computer vision<\/a>.<\/li>\n<\/ol>\n<h3><strong>Transportation &amp; Urban Systems<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo45\">A neural network-based model for estimating the delivery time of oxygen gas cylinders during COVID-19 pandemic<\/a>.<\/li>\n<li><a href=\"#Articulo46\">Exploring the Injury Severity Risk Factors in Fatal Crashes with Neural Network<\/a>.<\/li>\n<li><a href=\"#Articulo4\">Predicting Travel Times of Bus Transit in Washington, D.C. Using Artificial Neural Networks<\/a>.<\/li>\n<\/ol>\n<h3><strong>Social Sciences &amp; Education<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo47\">Effective Integration of Distance Courses Through Project-Based Learning<\/a>.<\/li>\n<\/ol>\n<h3><strong>Other Scientific Applications<\/strong><\/h3>\n<ol>\n<li><a href=\"#Articulo48\">Economic Land Utilization Optimization Model<\/a>.<\/li>\n<\/ol>\n<section id=\"Articulo20\">\n<h2>Application of machine learning for the prediction of Bacillus cereus growth: A study on the integration of mathematical models<\/h2>\n<p><b>Summary:<\/b><br \/>\nBacillus cereus is a foodborne pathogen whose growth depends strongly on storage conditions such as temperature, water activity and pH. This study used growth records from the ComBase database to build machine learning models that estimate how much B. cereus grows in different foods and assign a risk level, and it examined how outlier filtering and classic microbial growth equations can make these models more reliable.<\/p>\n<p>Neural Designer was one of the two platforms used to build the models, alongside Google Colab. The authors loaded the culture-medium dataset (about 9,500 records) with temperature, water activity, pH, initial count and time as inputs, and trained an approximation network for growth and a classification network for four danger levels, from safe to high risk. The data were split into 60% training, 20% selection and 20% testing samples, the number of neurons was set with Neural Designer&#8217;s model selection, and the models were evaluated with its testing analysis (goodness-of-fit for growth, confusion matrix for risk class).<\/p>\n<p>On the database test data, the Neural Designer classifier correctly classified about 64% of samples, while the coded Colab models scored somewhat higher. Against new laboratory growth experiments, however, the Neural Designer classifier gave the highest share of correct risk levels (about 61%) and the fewest severe misclassifications (2.7%), so the authors considered it the more suitable framework for risk classification. For predicting growth, the best results came from Colab models that combined PCA-DBSCAN outlier removal with the Gompertz model.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Lee, H.-Y., Her, J., Kim, W.-J., Kim, S.-O., &amp; Kim, S.-S. (2026). <a href=\"https:\/\/doi.org\/10.1016\/j.ijfoodmicro.2026.111901\" target=\"_blank\" rel=\"noopener\">Application of machine learning for the prediction of Bacillus cereus growth: A study on the integration of mathematical models<\/a>. International Journal of Food Microbiology, 459, 111901.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo21\">\n<h2>Integrating Genomic Classifiers and Nonsuspicious Magnetic Resonance Imaging Findings in Predictive Modelling for Lymph Node Metastasis in Patients With Localized Prostate Cancer<\/h2>\n<p><b>Summary:<\/b><br \/>\nExtended pelvic lymph node dissection during radical prostatectomy gives accurate nodal staging, but it lengthens surgery and causes complications, so surgeons need better ways to decide who really needs it. This study developed models to predict lymph node involvement in men with localized prostate cancer, including those with nonsuspicious MRI and those with genomic classifier results, and tested whether a neural network could match a standard logistic regression nomogram.<\/p>\n<p>The retrospective data covered patients operated on at a single center between 2014 and 2024, grouped into three cohorts: suspicious MRI (2,172 patients), nonsuspicious MRI (1,233) and genomic classifier testing (1,003). Neural Designer was used to build a binary classification model of lymph node involvement from the variables that were significant in the multivariable analysis, such as PSA, biopsy grade group, the proportion of positive cores and MRI nodal findings, plus cohort-specific predictors like MRI lesion size, MRI stage, prostate volume or the genomic score. The software split each cohort into training (60%), selection (20%) and testing (20%) samples, and discrimination was assessed with ROC curves.<\/p>\n<p>The neural networks reached areas under the ROC curve of 0.90, 0.82 and 0.91 in the three cohorts, close to the 0.84-0.92 obtained with the logistic regression nomograms. The authors conclude that neural networks are a feasible way to stratify lymph node risk with accuracy comparable to conventional multivariable models, although external validation is needed before clinical use.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Wagaskar, V. G., Maheshwari, A., Zaytoun, O., Agarwal, Y., Kolanukuduru, K. P., Tillu, N., et al. (2025). <a href=\"https:\/\/doi.org\/10.1016\/j.clgc.2025.102364\" target=\"_blank\" rel=\"noopener\">Integrating genomic classifiers and nonsuspicious magnetic resonance imaging findings in predictive modelling for lymph node metastasis in patients with localized prostate cancer<\/a>. Clinical Genitourinary Cancer, 23(4), 102364.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo11\">\n<h2>Machine learning modeling for solubility prediction of recombinant antibody fragment in four different E. coli strains<\/h2>\n<p><b>Summary:<\/b><br \/>\nThis article compares artificial neural networks (ANN) and response surface methodology (RSM) for the statistical optimization of E. coli fermentation conditions, with the aim of producing soluble recombinant single-chain antibody fragments (scFv).<\/p>\n<p>Neural Designer was used to train and optimize the artificial neural network. The inputs come from densitometry analysis, and the output is the soluble expression of anti-EpEX-scFv.<\/p>\n<p>The results show that post-induction time, inducer concentration, and the E. coli strain have a significant linear effect, with post-induction time being the most critical parameter.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Hashemi, A., Basafa, M., &amp; Behravan, A. (2022). <a href=\"https:\/\/doi.org\/10.1038\/s41598-022-09500-6\" target=\"_blank\" rel=\"noopener\">Machine learning modeling for solubility prediction of recombinant antibody fragment in four different E. coli strains<\/a>. Scientific Reports, 12, 5463.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo49\">\n<h2>Pneumonectomy and broncho-pleural fistula: predicting factors and stratification of the risk<\/h2>\n<p><b>Summary:<\/b><br \/>\nBronchopleural fistula (BPF) is one of the most serious complications after pneumonectomy for lung cancer. This study reviewed 366 consecutive patients who underwent pneumonectomy between 2009 and 2019 to identify the risk factors for BPF and to stratify patients into risk classes before surgery.<\/p>\n<p>Alongside log-rank tests and multivariable Cox regression, the authors ran a non-linear multiple regression with an artificial neural network, using Neural Designer together with SPSS and RStudio for the data analysis. The feedforward network had a scaling layer of 32 inputs and two perceptron layers, was trained with the quasi-Newton method, and was validated with leave-one-out cross-validation because the series was short.<\/p>\n<p>Fifty-one patients (13.9%) developed BPF. The Cox model identified male sex, right-sided surgery, postoperative pulmonary complications, and adjuvant treatment as independent predictors, while the neural network confirmed right side (p = 0.043) and adjuvant treatment (p = 0.032). The authors also observed a trend, not statistically significant, toward a higher risk of early fistula across the four preoperative risk classes.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Mazzella, A., Bertolaccini, L., Sedda, G., Prisciandaro, E., Loi, M., Lo Iacono, G., et al. (2022). <a href=\"https:\/\/doi.org\/10.1007\/s13304-022-01290-w\" target=\"_blank\" rel=\"noopener\">Pneumonectomy and broncho-pleural fistula: predicting factors and stratification of the risk<\/a>. Updates in Surgery, 74(4), 1471\u20131478.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo51\">\n<h2>Internet of things-inspired healthcare system for urine-based diabetes prediction<\/h2>\n<p><b>Summary:<\/b><br \/>\nHome monitoring with Internet of Things (IoT) sensors can help people with diabetes detect urine-related complications early. This study proposed a four-layer IoT system that acquires urine-based diabetes data, classifies it, computes a Diabetes Infection Measure, and predicts it so that preventive steps can be taken in time.<\/p>\n<p>The prediction layer is a recurrent neural network, which the authors implemented in Neural Designer. The network had six input units (the urine-based diabetes parameters), three recurrent hidden layers of nine neurons each, and one output, and it was evaluated with the coefficient of determination and RMSE on 8,569 records collected from four individuals.<\/p>\n<p>The system reported prediction accuracies of 97.3% and 97.6% for the first two individuals, higher than the baseline and ensemble methods it was compared with. Its reliability reached 94.86%, against 90.25% for k-NN, 92.45% for a classical neural network, and 93.02% for SVM.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Bhatia, M., Kaur, S., Sood, S. K., &amp; Behal, V. (2020). <a href=\"https:\/\/doi.org\/10.1016\/j.artmed.2020.101913\" target=\"_blank\" rel=\"noopener\">Internet of things-inspired healthcare system for urine-based diabetes prediction<\/a>. Artificial Intelligence in Medicine, 107, 101913.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo22\">\n<h2>Enhanced release of acid sphingomyelinase-enriched exosomes generates a lipidomics signature in CSF of Multiple Sclerosis patients<\/h2>\n<p><b>Summary:<\/b><br \/>\nMultiple sclerosis has no simple biomarkers for early, definite diagnosis, and the sphingomyelin pathway has been linked to its nerve damage and myelin repair. This study profiled cerebrospinal fluid (CSF) lipids in multiple sclerosis patients and in patients with other neurological diseases. It then looked at acid sphingomyelinase activity and exosomes as possible explanations for the differences.<\/p>\n<p>Neural Designer (version 2.9.5) carried out the first multivariate analysis of the lipidomics data. A neural network was fitted to 421 phosphatidylcholine and sphingomyelin signals from the CSF of 20 multiple sclerosis and 17 control patients, with diagnosis as the target. The correlation of each input with the diagnosis highlighted 96 discriminating lipid signals, and the final network had 3 hidden neurons. The authors cross-checked these results with PLS-DA and volcano-plot analyses.<\/p>\n<p>Thirty-two lipid signals were selected by all three methods, and several sphingomyelins were lower in the CSF of multiple sclerosis patients. A combined ROC model built on the selected lipids reached an AUC of 0.942. Follow-up assays showed higher acid sphingomyelinase activity in these patients (AUC of 0.88 against peripheral neurological diseases) and more exosomes enriched in the enzyme, which tracked enzyme activity and disease severity.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Pieragostino, D., Cicalini, I., Lanuti, P., Ercolino, E., di Ioia, M., Zucchelli, M., et al. (2018). <a href=\"https:\/\/doi.org\/10.1038\/s41598-018-21497-5\" target=\"_blank\" rel=\"noopener\">Enhanced release of acid sphingomyelinase-enriched exosomes generates a lipidomics signature in CSF of Multiple Sclerosis patients<\/a>. Scientific Reports, 8(1), 3071.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo23\">\n<h2>Predictors of in-hospital mortality following major lower extremity amputations in type 2 diabetic patients using artificial neural networks<\/h2>\n<p><b>Summary:<\/b><br \/>\nPeople with type 2 diabetes who undergo a major lower-limb amputation face a high risk of dying in hospital. Clinicians need better ways to tell which patients are most at risk. This study applied artificial neural networks to Spain&#8217;s national hospital discharge database to find predictors of in-hospital death and to compare which comorbidity index gives the best predictions.<\/p>\n<p>The data covered 40,857 major amputations in patients with type 2 diabetes from 2003 to 2013. It was split into training (60%), selection (20%) and testing (20%) subsets. The authors used Neural Designer to design four classification networks predicting in-hospital death. Inputs were age, sex, length of stay and comorbidities, with the comorbidities defined four ways: all discharge diagnoses, all diagnoses except infections, the Charlson index conditions, or the Elixhauser index conditions. The networks were trained with the quasi-Newton method, and model selection set the number of hidden neurons. A sensitivity analysis then ranked the inputs by importance.<\/p>\n<p>The Elixhauser-based model performed best, with an area under the ROC curve of 91.7% versus 88.9% for the Charlson model. On the testing data it reached an accuracy of 0.861, a specificity of 0.960 and a precision of 0.951. Age, female sex, congestive heart failure, renal failure and chronic pulmonary disease were the strongest predictors of in-hospital mortality.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Lopez-de-Andres, A., Hernandez-Barrera, V., Lopez, R., Martin-Junco, P., Jimenez-Trujillo, I., Alvaro-Meca, A., et al. (2016). <a href=\"https:\/\/doi.org\/10.1186\/s12874-016-0265-5\" target=\"_blank\" rel=\"noopener\">Predictors of in-hospital mortality following major lower extremity amputations in type 2 diabetic patients using artificial neural networks<\/a>. BMC Medical Research Methodology, 16(1), 160.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo24\">\n<h2>Building Artificial Neural Networks for the Optimization of Sustained-Release Kinetics of Metronidazole from Colonic Hydrophilic Matrices<\/h2>\n<p><b>Summary:<\/b><br \/>\nTreating local colonic diseases such as diverticulitis with metronidazole calls for coated matrix tablets whose release depends on many interacting formulation and process factors. This study set out to build an artificial neural network that links those technological factors to the in vitro release of colonic hydrophilic matrices developed under a quality-by-design framework.<\/p>\n<p>The data came from 28 batches of coated HPMC\/chitosan matrix tablets manufactured in an earlier fractional factorial design. Six inputs (HPMC viscosity grade, HPMC and chitosan percentages, mixing time, coating agent and coating weight gain) were used to predict five targets: the percentage of drug released at 1, 6, 12 and 24 h and the Weibull mean dissolution time. The network was built in Neural Designer with a 60\/20\/20 training, selection and testing split; an incremental order selection settled on four hidden neurons (a 6-4-5 architecture), and the weights were trained with a quasi-Newton method, L2 regularization and the software&#8217;s built-in stopping criteria.<\/p>\n<p>Correlations between predicted and observed values were 0.971, 0.964 and 0.953 for release at 1, 6 and 12 h, and 0.876 and 0.789 for release at 24 h and mean dissolution time, with mean percentage errors below 1% for every output except release at 1 h (5.5%). The network predicted better than the Weibull model at early times and better than multiple linear regression at later times, and it pointed to the coating agent and coating weight gain as the main drivers of early release and to HPMC viscosity as key at 12 and 24 h.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Maderuelo, C., Ar\u00e9valo-P\u00e9rez, R., &amp; Lanao, J. M. (2025). <a href=\"https:\/\/doi.org\/10.3390\/pharmaceutics17111451\" target=\"_blank\" rel=\"noopener\">Building artificial neural networks for the optimization of sustained-release kinetics of metronidazole from colonic hydrophilic matrices<\/a>. Pharmaceutics, 17(11), 1451.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo25\">\n<h2>Integrating Artificial Intelligence with Quality by Design in the Formulation of Lecithin\/Chitosan Nanoparticles of a Poorly Water-Soluble Drug<\/h2>\n<p><b>Summary:<\/b><br \/>\nSilymarin shows promising activity against liver cancer, but its very low water solubility limits how much of it the body absorbs. This study combined a quality-by-design approach with artificial intelligence to develop silymarin-loaded lecithin\/chitosan nanoparticles and to predict how fast drugs are released from this kind of carrier.<\/p>\n<p>For the modelling part, the authors gathered 23 lecithin\/chitosan nanoparticle formulations from the literature that used similar excipients and processing but different drugs. Each drug was described by eight molecular descriptors (molecular weight, LogP, hydrogen bond donor and acceptor counts, rotatable bonds, topological polar surface area, heavy atom count and complexity), which served as inputs to predict the percentage released at 2, 8 and 12 h. Three deep neural networks, one per time point, were built in Neural Designer 4.2.0 with a 60\/20\/20 training, selection and testing split, a bounding layer on the output, Levenberg-Marquardt training, and early stopping with L2 regularization against overfitting.<\/p>\n<p>The release models achieved coefficients of determination of 0.981, 0.991 and 0.937 at 2, 8 and 12 h, and their predictions for the optimized silymarin formulation did not differ significantly from the measured release. Separately, the experimental design produced optimized nanoparticles of about 161 nm with 97% entrapment efficiency, which lowered the IC50 of silymarin on HepG2 liver cancer cells from 140 to 90.2 \u00b5M.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Dawoud, M. H. S., Mannaa, I. S., Abdel-Daim, A., &amp; Sweed, N. M. (2023). <a href=\"https:\/\/doi.org\/10.1208\/s12249-023-02609-5\" target=\"_blank\" rel=\"noopener\">Integrating artificial intelligence with quality by design in the formulation of lecithin\/chitosan nanoparticles of a poorly water-soluble drug<\/a>. AAPS PharmSciTech, 24(6), 169.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo26\">\n<h2>Optimization of Biomass Delivery Through Artificial Intelligence Techniques<\/h2>\n<p><b>Summary:<\/b><br \/>\nSupplying biomass to a power plant is a hard logistics problem: fuel quality varies, suppliers are scattered, and market data are often incomplete, which limits classical optimization methods. This study set out to build an artificial neural network (ANN) model for Biomass Delivery Management (BDM) that recommends suitable suppliers for a fluidized-bed combined heat and power (CHP) plant in Poland.<\/p>\n<p>The BDM model was developed in Neural Designer, with biomass type, unit price (PLN\/GJ) and annual demand (Mg\/year) as inputs and a supplier identifier as the target. The dataset comprised 39 records of real operating data from the plant (2018-2019), with 10% of the records reserved for testing and 10% for validation. After comparing several architectures, the authors selected a network with scaling and unscaling layers and two hidden layers of three hyperbolic-tangent neurons each, trained with the quasi-Newton method.<\/p>\n<p>On the test data the model reached MAE = 0.16, MSE = 0.02 and R2 = 0.99. In three example supply scenarios with different biomass types, prices and quantities, the rounded model output pointed to the correct supplier in the knowledge base. The authors present the model as a proof of concept, noting that the small dataset limits how far the results can be generalized.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Wesolowska, M., \u017belazna-Jochim, D., Wisniewski, K., Krzywanski, J., Sosnowski, M., &amp; Nowak, W. (2025). <a href=\"https:\/\/doi.org\/10.3390\/en18185028\" target=\"_blank\" rel=\"noopener\">Optimization of biomass delivery through artificial intelligence techniques<\/a>. Energies, 18(18), 5028.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo52\">\n<h2>Artificial neural network assisted numerical analysis on performance enhancement of Sb2(S,Se)3 solar cell with SnS as HTL<\/h2>\n<p><b>Summary:<\/b><br \/>\nAntimony sulfide-selenide, Sb2(S,Se)3, is a promising absorber for thin-film solar cells. The authors used the SCAPS-1D simulator to design a new Ni\/SnS\/Sb2(S,Se)3\/ZnS\/FTO\/Al cell, with tin sulfide as the hole transport layer and zinc sulfide replacing the toxic CdS buffer, and studied how layer thickness, doping, defects, temperature, and resistances affect its performance.<\/p>\n<p>Neural Designer was used to build an artificial neural network, without programming, that predicts the power conversion efficiency from four absorber properties: thickness, band gap, electron affinity, and dielectric permittivity. The 256 simulated devices were split into 154 training, 51 selection, and 51 testing samples, and the importance of each input was assessed through the partial derivatives of the output.<\/p>\n<p>The network reproduced the simulated efficiencies almost exactly, with an R\u00b2 of 0.9994 on the training, selection, and testing data and a testing RMSE of 0.073. Electron affinity was the most influential property, and the optimized cell reached a simulated efficiency of 28.20% (Voc of 0.94 V, Jsc of 34.65 mA\/cm\u00b2, and fill factor of 86.22%).<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Khan, T. M., Hosen, A., Saidani, O., &amp; Ahmed, S. R. A. (2024). <a href=\"https:\/\/doi.org\/10.1016\/j.mtcomm.2024.109639\" target=\"_blank\" rel=\"noopener\">Artificial neural network assisted numerical analysis on performance enhancement of Sb2(S,Se)3 solar cell with SnS as HTL<\/a>. Materials Today Communications, 40, 109639.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo27\">\n<h2>Deep learning-based evaluation of photovoltaic power generation<\/h2>\n<p><b>Summary:<\/b><br \/>\nThe output of photovoltaic (PV) systems rises and falls with the weather, which makes it hard for grid operators to anticipate voltage and current conditions as more solar capacity is connected. This study set out to build a short-term forecasting model of PV power that could be plugged into existing utility management software to support grid planning and stability.<\/p>\n<p>The modelling was carried out in Neural Designer using about a year of minute-level measurements from the NIST Net-Zero Energy Residential Test Facility, covering the power of four PV arrays, total and average PV power, AC total power, inverter efficiency and PV-to-DC efficiency. The data were split into 26,783 training, 8,927 selection and 8,927 testing samples, and the network combined an LSTM layer with perceptron, scaling, unscaling and bounding layers. The authors applied growing-inputs selection, a normalized squared error loss with L2 regularization and the quasi-Newton optimizer, and then checked the model with correlation analysis, linear regression on the test subset, error statistics and time-series plots.<\/p>\n<p>Training stopped after 178 epochs, with final training and selection errors of about 0.0018. On the testing data the mean percentage error was about 5.9%, and the regression between predicted and measured total DC power gave a correlation of 0.96. The authors conclude that their model forecasts short-term PV output more accurately than a deep recurrent neural network reported earlier in the literature.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Diaba, S. Y., Alola, A. A., Simoes, M. G., &amp; Elmusrati, M. (2024). <a href=\"https:\/\/doi.org\/10.1016\/j.egyr.2024.08.007\" target=\"_blank\" rel=\"noopener\">Deep learning-based evaluation of photovoltaic power generation<\/a>. Energy Reports, 12, 2077\u20132085.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo28\">\n<h2>Design and simulation of a highly efficient CuBi2O4 thin-film solar cell with hole transport layer<\/h2>\n<p><b>Summary:<\/b><br \/>\nCopper bismuth oxide (CuBi2O4) is a non-toxic, earth-abundant absorber for low-cost thin-film solar cells, but carrier recombination at the rear contact limits its efficiency. This study used the SCAPS-1D simulator to compare several hole transport layers (HTLs) and to optimize a Mo\/Cu2O\/CuBi2O4\/CdS\/FTO cell, introducing Cu2O as the HTL in this type of device for the first time.<\/p>\n<p>From the SCAPS-1D simulations the authors generated a dataset of photovoltaic outputs, and used Neural Designer to train an artificial neural network linking material and device input parameters to performance outputs such as efficiency. Neural Designer was also used to assess the errors and accuracy of the model on the training and test sets.<\/p>\n<p>The authors report high accuracy and low error metrics for both the training and the test data. The optimized cell, with a 1.0 um absorber and a Cu2O HTL, reached a simulated efficiency of 29.2%, with an open-circuit voltage of 1.02 V, a short-circuit current density of 32.49 mA\/cm2 and a fill factor of 87.91%.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Hosen, A., Yeasmin, S., Bin Rahmotullah, K. M. S., Rahman, M. F., &amp; Ahmed, S. R. A. (2024). <a href=\"https:\/\/doi.org\/10.1016\/j.optlastec.2023.110073\" target=\"_blank\" rel=\"noopener\">Design and simulation of a highly efficient CuBi2O4 thin-film solar cell with hole transport layer<\/a>. Optics &amp; Laser Technology, 169, 110073.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo29\">\n<h2>Forecasting meteorological impacts on the environmental sustainability of a large-scale solar plant via artificial intelligence-based life cycle assessment<\/h2>\n<p><b>Summary:<\/b><br \/>\nLarge-scale solar farms avoid emissions while they run, but making the panels and later dismantling the plant have environmental costs. The electricity they deliver also varies with the weather. This study combined life cycle assessment with an artificial neural network to see how weather variability changes the overall environmental balance of a large solar plant in Malaysia.<\/p>\n<p>The neural network was a feedforward model built in Neural Designer 5.9.3. It took monthly weather variables (ambient temperature, direct normal irradiance, PV module temperature and wind speed) and predicted the plant&#8217;s monthly electricity generation. In the study&#8217;s framework, the model was trained with the Levenberg\u2013Marquardt algorithm and judged by R\u00b2 and relative root mean square error. It was then fed Monte Carlo-simulated weather to forecast monthly output from mid-2022 to 2045.<\/p>\n<p>Predicted monthly generation closely followed measured values (R\u00b2 \u2248 0.92 in the paper&#8217;s validation plot). The life cycle results showed PV panel production as the main hotspot, causing about 30% of the human-health and ecosystem damage and 34% of the resource damage. Recycling aluminium cut these burdens by about 9\u201310%. The emissions avoided by the forecasted electricity were 12 to 98 times larger than the plant&#8217;s life-cycle burdens, depending on the damage category.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Wan, M. J., Phuang, Z. X., Hoy, Z. X., Dahlan, N. Y., Azmi, A. M., &amp; Woon, K. S. (2024). <a href=\"https:\/\/doi.org\/10.1016\/j.scitotenv.2023.168779\" target=\"_blank\" rel=\"noopener\">Forecasting meteorological impacts on the environmental sustainability of a large-scale solar plant via artificial intelligence-based life cycle assessment<\/a>. Science of The Total Environment, 912, 168779.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo30\">\n<h2>Performance analysis and optimization of SnSe thin-film solar cell with Cu2O HTL through a combination of SCAPS-1D and machine learning approaches<\/h2>\n<p><b>Summary:<\/b><br \/>\nTin selenide (SnSe) is an earth-abundant absorber that takes in light strongly, but SnSe solar cells have remained inefficient, partly because their transport layers are poorly matched. The authors designed an Al\/FTO\/WS2\/SnSe\/Cu2O\/Ni thin-film cell in SCAPS-1D and compared it with other electron- and hole-transport layers. They also studied how absorber thickness, doping, defects, back-contact work function and temperature affect performance.<\/p>\n<p>The authors used Neural Designer to build an artificial neural network that predicts cell efficiency from absorber properties, without writing code. The dataset had 243 SCAPS-1D simulations, with absorber thickness, band gap, electron affinity, acceptor density and defect density as inputs and conversion efficiency as the target. The samples were split into 146 for training, 24 for selection and 73 for testing, and the study also reports the relative importance and correlation of each input.<\/p>\n<p>The network reached R2 values of 0.910, 0.903 and 0.908 on the training, selection and testing data, with a testing RMSE of 0.586. Defect density was by far the most influential property (about 74% relative importance), followed by electron affinity. The optimized device reached a simulated efficiency of 29.31%, with Voc of 0.89 V, Jsc of 38.23 mA\/cm2 and FF of 86.30%.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Khan, T. M., Islam, B., &amp; Ahmed, S. R. A. (2024). <a href=\"https:\/\/doi.org\/10.1016\/j.mtcomm.2024.110490\" target=\"_blank\" rel=\"noopener\">Performance analysis and optimization of SnSe thin-film solar cell with Cu2O HTL through a combination of SCAPS-1D and machine learning approaches<\/a>. Materials Today Communications, 41, 110490.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo31\">\n<h2>Utilizing machine learning to enhance performance of thin-film solar cells based on Sb2(SxSe1\u2212x)3: investigating the influence of material properties<\/h2>\n<p><b>Summary:<\/b><br \/>\nAntimony sulfoselenide, Sb2(S,Se)3, is a promising cadmium-free absorber for low-cost thin-film solar cells. Using the SCAPS-1D simulator, the authors designed a Ni\/Cu2O\/Sb2(S,Se)3\/WS2\/FTO\/Al cell, tuned the properties of each layer, and then turned to machine learning to see which absorber properties drive its efficiency.<\/p>\n<p>Neural Designer was used to build a neural network that predicts power conversion efficiency from five absorber properties: thickness, band gap, electron affinity, electron mobility and hole mobility. The 1024 simulated devices were split into 616 training, 204 selection and 204 testing samples, and the trained model was used to rank the inputs by the partial derivatives of the output with respect to each one.<\/p>\n<p>Predicted and simulated efficiencies agreed almost perfectly, with R2 of about 0.9993 and RMSE between 0.020 and 0.026 on the three subsets. Band gap, electron affinity and absorber thickness were the most influential properties, and the optimized cell reached a simulated efficiency of 30.18% (Voc 1.02 V, Jsc 33.65 mA\/cm2, fill factor 87.59%).<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Khan, T. M., Saidani, O., &amp; Ahmed, S. R. A. (2024). <a href=\"https:\/\/doi.org\/10.1039\/d4ra03340j\" target=\"_blank\" rel=\"noopener\">Utilizing machine learning to enhance performance of thin-film solar cells based on Sb2(SxSe1\u2212x)3: Investigating the influence of material properties<\/a>. RSC Advances, 14(38), 27749\u201327763.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo32\">\n<h2>AI-coherent data-driven forecasting model for a combined cycle power plant<\/h2>\n<p><b>Summary:<\/b><br \/>\nCombined cycle power plants are among the most efficient ways to generate electricity from fuel, but their output varies with ambient conditions. This study takes a tutorial-style approach to data-driven modelling, using machine learning to forecast the electrical output of such a plant and to find operating conditions that raise it.<\/p>\n<p>The authors built the prediction model in Neural Designer from 9,568 hourly records collected over six years at a plant running at full load. Ambient temperature, ambient pressure, relative humidity and exhaust vacuum were the inputs, and net hourly electrical output was the target. The data were split into training, selection and testing subsets, and the network was trained with the quasi-Newton (BFGS) method. The growing-neurons and growing-inputs algorithms chose the network size and inputs, and testing analysis evaluated the final model.<\/p>\n<p>The model reached a coefficient of determination of 0.945 between predicted and actual output, with final training and selection errors of about 0.063 and 0.059. By simulating operating scenarios and adjusting the input variables, the authors report that output could rise from 452 MW to 462.1 MW, a gain of 2.23%.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Danish, M. S. S., Nazari, Z., &amp; Senjyu, T. (2023). <a href=\"https:\/\/doi.org\/10.1016\/j.enconman.2023.117063\" target=\"_blank\" rel=\"noopener\">AI-coherent data-driven forecasting model for a combined cycle power plant<\/a>. Energy Conversion and Management, 286, 117063.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo14\">\n<h2>N,S co-doped biocarbon for supercapacitor application: Effect of electrolytes concentration and modelling with artificial neural network<\/h2>\n<p><b>Summary:<\/b><br \/>\nExcessive consumption of fossil fuels has driven research into energy storage devices, and supercapacitors have emerged as competitive and robust energy storage systems. This study prepares nitrogen and sulfur co-doped biocarbon derived from chicken bones for supercapacitor electrodes.<\/p>\n<p>The objective is to predict the capacitance and energy density of the biocarbon. The input variables are sulfur doping, nitrogen doping, electrolyte concentration, electrolyte type, and S\/N co-doping. Neural Designer was used to select the model parameters, namely the number of layers and neurons.<\/p>\n<p>The resulting model achieves an R\u00b2 of 0.964 and an average desirability function of 0.96.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Oladipo, A. A. (2021). <a href=\"https:\/\/doi.org\/10.1016\/j.matchemphys.2020.124129\" target=\"_blank\" rel=\"noopener\">N,S co-doped biocarbon for supercapacitor application: Effect of electrolytes concentration and modelling with artificial neural network<\/a>. Materials Chemistry and Physics, 260, 124129.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo33\">\n<h2>Mineralogical composition and total organic carbon quantification using x-ray fluorescence data from the Upper Cretaceous Eagle Ford Group in southern Texas<\/h2>\n<p><b>Summary:<\/b><br \/>\nOperators need to know the mineral content and total organic carbon (TOC) of shale to decide where to place horizontal wells and how to complete them. Laboratory XRD and TOC measurements, however, are slow, costly and give only scattered points. Alnahwi and Loucks tested whether cheap, non-destructive handheld X-ray fluorescence (XRF) scans of Eagle Ford cores from southern Texas could be turned into continuous mineralogy and TOC logs with neural networks.<\/p>\n<p>In Neural Designer, the authors built a separate network for each mineral target (calcite, dolomite, quartz, feldspar, pyrite and clay minerals). Inputs were XRF major elements (Al, Ca, Fe, K, Mg, S and Si), calibrated against 35 XRD analyses split into 21 training, 7 selection and 7 testing instances. A second model predicted TOC from the trace metals Mo, Ni and Cu, using 678 TOC measurements split 60\/20\/20. Both models used z-score scaling, a normalized squared error loss, hyperbolic tangent activations and the quasi-Newton training algorithm, with the number of hidden layers and neurons tuned to minimise error. The authors checked each model with a testing analysis and then deployed it in the software to produce high-resolution logs for the cores.<\/p>\n<p>On the testing instances, five of the six mineral models reached R\u00b2 values between 0.90 and 0.99; quartz was lowest at 0.73. The TOC model reached R\u00b2 = 0.85 with a slope of 1.00. On two blind-test cores never used in training, predicted TOC still matched laboratory values with R\u00b2 of 0.77 and 0.66. The authors argue this makes XRF a practical route to near-real-time mineral and TOC estimates, for example from drill cuttings.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Alnahwi, A., &amp; Loucks, R. G. (2019). <a href=\"https:\/\/doi.org\/10.1306\/04151918090\" target=\"_blank\" rel=\"noopener\">Mineralogical composition and total organic carbon quantification using x-ray fluorescence data from the Upper Cretaceous Eagle Ford Group in southern Texas<\/a>. AAPG Bulletin, 103(12), 2891\u20132907.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo34\">\n<h2>Using a Neural Network to Improve the Optical Absorption in Halide Perovskite Layers Containing Core-Shells Silver Nanoparticles<\/h2>\n<p><b>Summary:<\/b><br \/>\nPlasmonic core-shell nanoparticles can boost light absorption in thin-film solar cells, but the number of possible designs (particle size, number and thickness of shells, position in the film) is practically endless. This study used a neural network to search for silver core-shell configurations that maximize optical absorption in a methylammonium lead iodide perovskite film.<\/p>\n<p>A limited set of finite-difference time-domain simulations was run for particles of 20, 30 and 60 nm with up to three silver shells, shell thicknesses between 1.5 and 4.5 nm, and several positions in or on a 200 nm perovskite layer, and the resulting perovskite absorption spectra were used as training data. The networks were built in Neural Designer 3.0 with hyperbolic tangent hidden layers and a linear output layer; three architectures were compared, and order selection was run after training to minimize the selection error. The models were analysed over three wavelength ranges: 500-600 nm, 700-800 nm and 1000-1200 nm.<\/p>\n<p>The three models had average errors of at most 2.5%. The network beat the best simulated design only around 500 nm, close to the silver plasmon resonance, where it proposed a configuration that had not been simulated (a 60 nm particle near the top of the film with a silver core and a 4.5 nm silver outer shell); it could not improve on the simulations in the infrared range. The authors present the work as a proof of concept that neural networks can guide nanophotonic light-management designs for solar cells.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Nelson, M. D., &amp; Di Vece, M. (2019). <a href=\"https:\/\/doi.org\/10.3390\/nano9030437\" target=\"_blank\" rel=\"noopener\">Using a neural network to improve the optical absorption in halide perovskite layers containing core-shells silver nanoparticles<\/a>. Nanomaterials, 9(3), 437.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo53\">\n<h2>Low-pressure reverse osmosis membrane made of cellulose nanofiber and carbon nanotube polyamide nano-nanocomposite for high purity water production<\/h2>\n<p><b>Summary:<\/b><br \/>\nProducing high-purity water by reverse osmosis requires membranes that combine high salt rejection, good permeability, and durability at low pressure. This study synthesized polyamide membranes that incorporate both cellulose nanofibers (CNF) and multiwall carbon nanotubes (CNT) and tested them in a low-pressure reverse osmosis system.<\/p>\n<p>Part of the data analysis was carried out in Neural Designer, which the authors used to fit the data from several membrane synthesis experiments and quantify how each preparation variable influences salt rejection and water permeation. The analysis showed that NaNO2 treatment improves both rejection and permeation, whereas higher TEA concentrations reduce salt rejection and IPA concentrations above 8% lower water permeation.<\/p>\n<p>The PA-CNT\/CNF membrane achieved a NaCl rejection of 99.47% and a water permeability of 1.65 m\u00b3\/m\u00b2\u00b7day at 0.75 MPa, competitive with commercial membranes. In a 2-inch spiral module, its permeability dropped by 24.9% after 100 h of operation, compared with 49.42% for commercial modules, and in high-purity water production it showed about twice the durability of a certified TW30 module.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Fajardo-Diaz, J. L., Morelos-Gomez, A., Cruz-Silva, R., Ishii, K., Yasuike, T., Kawakatsu, T., et al. (2022). <a href=\"https:\/\/doi.org\/10.1016\/j.cej.2022.137359\" target=\"_blank\" rel=\"noopener\">Low-pressure reverse osmosis membrane made of cellulose nanofiber and carbon nanotube polyamide nano-nanocomposite for high purity water production<\/a>. Chemical Engineering Journal, 448, 137359.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo12\">\n<h2>Predicting the Properties of High-Performance Epoxy Resin by Machine Learning Using Molecular Dynamics Simulations<\/h2>\n<p><b>Summary:<\/b><br \/>\nThe performance of epoxy systems varies significantly with the composition of the base resin and the curing agent, and exploring many formulations by trial and error is expensive and slow. The authors combined molecular dynamics (MD) simulations with machine learning to predict the adhesive properties of epoxy resins.<\/p>\n<p>Using MD, the authors built datasets of cohesive energy density (CED), modulus, and glass transition temperature (Tg) for formulations combining four base resins and three curing agents. The entire machine learning procedure, from data analysis to training and optimizing the neural networks with training, selection, and test subsets, was carried out in Neural Designer.<\/p>\n<p>The optimized neural network predicted CED with RMSEs of 1.413 and 1.124, and the modulus with RMSEs of 0.223 and 0.347, on the training and test sets, respectively. The correlation analysis also revealed which constituents drive each property, such as the curing agent DICY for CED.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Choi, J., Kang, H., Lee, J. H., Kwon, S. H., &amp; Lee, S. G. (2022). <a href=\"https:\/\/doi.org\/10.3390\/nano12142353\" target=\"_blank\" rel=\"noopener\">Predicting the properties of high-performance epoxy resin by machine learning using molecular dynamics simulations<\/a>. Nanomaterials, 12(14), 2353.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo35\">\n<h2>Prediction of Lap Shear Strength and Impact Peel Strength of Epoxy Adhesive by Machine Learning Approach<\/h2>\n<p><b>Summary:<\/b><br \/>\nStructural epoxy adhesives used in vehicles blend a resin with many additives, and the number of possible recipes makes their bonding performance hard to predict in advance. This study aimed to estimate two key adhesion properties, lap shear strength at room temperature and impact peel strength at -40 \u00b0C, directly from the formulation using artificial neural networks.<\/p>\n<p>The authors tested 50 formulations and used the amounts of eight components (resin, core-shell rubber, flexibilizer, diluent, filler, promoter, curing agent and catalyst) as inputs, building a separate network for each target. All modelling was done in Neural Designer: the data were min-max scaled and divided into selection, training and testing subsets, the number of hidden layers and neurons was tuned on the selection set (three hidden layers of 4, 3 and 7 neurons for lap shear; four layers of 7, 5, 3 and 2 for impact peel), and the tanh networks were trained with the quasi-Newton method on a normalized squared error loss.<\/p>\n<p>The lap shear model reached RMSEs of 0.053 on training and 0.590 on testing data, while the impact peel model gave 1.730 and 8.218, and linear regression of predicted against measured values yielded R2 of 0.642 and 0.588, respectively. Using the trained networks to vary the most correlated components, the authors found that more catalyst lowered both strengths and that impact peel strength levelled off at its highest values once the flexibilizer weight ratio exceeded about 25.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Kang, H., Lee, J. H., Choe, Y., &amp; Lee, S. G. (2021). <a href=\"https:\/\/doi.org\/10.3390\/nano11040872\" target=\"_blank\" rel=\"noopener\">Prediction of lap shear strength and impact peel strength of epoxy adhesive by machine learning approach<\/a>. Nanomaterials, 11(4), 872.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo19\">\n<h2>Statistical and Machine Learning-Driven Optimization of Mechanical Properties in Designing Durable HDPE Nanobiocomposites<\/h2>\n<p><b>Summary:<\/b><br \/>\nChoosing the nanofillers and compatibilizing agents, and their size and concentration, is crucial to designing durable nanobiocomposites with high fracture strength, yield strength, and Young&#8217;s modulus. Statistical optimization of these design factors reduces the number of experimental runs and their cost.<\/p>\n<p>The authors combined statistical methods (ANOVA and response surface methodology) with artificial neural networks and genetic algorithms to optimize the concentrations of TiO\u2082, cellulose nanocrystals, and compatibilizer in injection-molded HDPE nanocomposites. Neural Designer was used to create and train the neural network, which was then coupled with a genetic algorithm for optimization.<\/p>\n<p>ANOVA showed that TiO\u2082 and cellulose nanocrystal concentrations, and their combination, significantly improved durability, and the ANN-GA approach gave the best prediction of the desired mechanical properties.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Mairpady, A., Mourad, A.-H. I., &amp; Mozumder, M. S. (2021). <a href=\"https:\/\/doi.org\/10.3390\/polym13183100\" target=\"_blank\" rel=\"noopener\">Statistical and machine learning-driven optimization of mechanical properties in designing durable HDPE nanobiocomposites<\/a>. Polymers, 13(18), 3100.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo50\">\n<h2>Metal recovery prediction of elements from anode slime<\/h2>\n<p><b>Summary:<\/b><br \/>\nCopper refineries produce anode slime, a by-product that contains precious and other valuable metals, but predicting how much of each element a leaching process will recover is difficult because experimental data are scarce. This study modeled the recovery of copper, lead, barium, gold, selenium, and tellurium from the anode slime of a Turkish copper refinery leached with an ionic liquid.<\/p>\n<p>All the prediction models were built in Neural Designer. Each network used four inputs (temperature, leaching time, solid-to-liquid ratio, and ionic liquid concentration) to predict the recovery percentage of an element, and was trained with the quasi-Newton method. The authors assessed the models with a linear regression of predicted against measured recoveries, error tables (sum squared, mean squared, and Minkowski errors), and plots of predicted recovery as a function of each input.<\/p>\n<p>The R\u00b2 values ranged from 87.13% to 91.78%, with the best fits for barium, gold, and selenium. The model predicted a maximum gold recovery of 82.11% at a solid-to-liquid ratio of 1\/25, a temperature of 45 \u00b0C, an ionic liquid concentration of 40%, and a leaching time of 0.5 h.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>R\u00fc\u015fen, A., Yildizel, S. A., &amp; Top\u00e7u, M. A. (2019). <a href=\"https:\/\/doi.org\/10.1007\/s13762-019-02224-7\" target=\"_blank\" rel=\"noopener\">Metal recovery prediction of elements from anode slime<\/a>. International Journal of Environmental Science and Technology, 16(11), 6797\u20136804.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo36\">\n<h2>Application of DFT-based machine learning for developing molecular electrode materials in Li-ion batteries<\/h2>\n<p><b>Summary:<\/b><br \/>\nOrganic molecules are attractive candidates for lithium-ion battery electrodes, but screening them by redox potential with quantum chemistry alone is slow. The authors combined density functional theory (DFT) with machine learning to create a high-throughput way of predicting the redox potentials of candidate molecular electrode materials.<\/p>\n<p>DFT supplied electronic properties (electron affinity, HOMO, LUMO and HOMO-LUMO gap) for 114 quinone derivatives and functionalized or boron-doped graphene fragments, along with structural counts such as oxygen and lithium atoms. A network with two hidden layers of five and three neurons was trained in Neural Designer with the quasi-Newton method and the normalized squared error, using 108 cases for training and six literature cases for external checking; a correlation analysis reduced the inputs from ten to six.<\/p>\n<p>The model closely reproduced the DFT redox potentials, with an average error of 3.54% on cases outside the training set. A contribution analysis ranked electron affinity as the most influential input, followed by the number of oxygen atoms, the HOMO-LUMO gap, the number of lithium atoms, LUMO and HOMO.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Allam, O., Cho, B. W., Kim, K. C., &amp; Jang, S. S. (2018). <a href=\"https:\/\/doi.org\/10.1039\/c8ra07112h\" target=\"_blank\" rel=\"noopener\">Application of DFT-based machine learning for developing molecular electrode materials in Li-ion batteries<\/a>. RSC Advances, 8(69), 39414\u201339420.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo37\">\n<h2>Machine learning evaluation of PI control effects on neutral equilibrium in bridge virtual pier systems<\/h2>\n<p><b>Summary:<\/b><br \/>\nNeutral equilibrium mechanisms (NEMs) use pre-tensioned tendons and servo-driven rotating arms to act as &#8220;virtual piers&#8221; that actively counteract the vertical deflection of a bridge deck. How well they work depends on the proportional and integral gains of their PI controller, so this study used machine learning to find the best gain settings for a scaled bridge fitted with two NEMs.<\/p>\n<p>The authors ran dynamic loading tests under four gain combinations and logged displacements at two sensor points every millisecond, producing more than 21 million data points. Neural Designer was used to build a multilayer perceptron relating the control settings, time and displacement records to the bridge response, with the data split 60\/20\/20 into training, selection and testing sets. The software&#8217;s sensitivity analysis ranked the influence of each input, and its optimization module was set up to search for the gains that minimize steady-state displacement; random forest regression and K-means clustering complemented the analysis.<\/p>\n<p>The network predicted displacements with low errors, with RMSE of 0.029 on the training set and 0.038 on the selection and testing sets. Displacement records under the higher gain settings proved the most informative inputs, and together with clustering the study identified GP = 1.0 and GI = 0.010 as the best balance, reducing peak displacements from about 5 mm to under 0.4 mm and the settling time to 9.8 s.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Sung, W.-P., &amp; Shih, M.-H. (2025). <a href=\"https:\/\/doi.org\/10.1038\/s41598-025-20482-z\" target=\"_blank\" rel=\"noopener\">Machine learning evaluation of PI control effects on neutral equilibrium in bridge virtual pier systems<\/a>. Scientific Reports, 15(1), 36630.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo38\">\n<h2>Optimization of concrete with human hair using experimental study and artificial neural network via response surface methodology and ANOVA<\/h2>\n<p><b>Summary:<\/b><br \/>\nHuman hair is an abundant waste fibre that could reinforce concrete instead of ending up in landfills. This study made concretes with hair contents of 1% to 5% of the cement weight across several mix groups. It measured their fresh and mechanical properties, then modelled and optimized the results with artificial neural networks, response surface methodology and ANOVA.<\/p>\n<p>The authors built and optimized the ANN models in Neural Designer 6.0 without writing custom code. Four mix-design ratios were the inputs: water\/cement, (water + cement)\/aggregate, fine\/coarse aggregate and hair content. The outputs were compressive, splitting tensile and flexural strength. The network was a multilayer perceptron with sigmoid activation, trained with the quasi-Newton method and a Brent line search for the learning rate (tolerance 0.001).<\/p>\n<p>The Neural Designer models closely matched the measured strengths, with R2 values of 0.962 for compressive, 0.956 for splitting tensile and 0.947 for flexural strength. RMSE ranged from 1.39 to 1.53 and MAPE from 0.030 to 0.034. In the experiments, strength rose with hair content up to about 3% and then fell, which the authors attribute to poor dispersion and clumping of the fibres.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Y\u0131ld\u0131zel, S. A., Karalar, M., Aksoylu, C., Althaqafi, E., Beskopylny, A. N., Stel&#8217;makh, S. A., et al. (2025). <a href=\"https:\/\/doi.org\/10.1038\/s41598-025-12782-1\" target=\"_blank\" rel=\"noopener\">Optimization of concrete with human hair using experimental study and artificial neural network via response surface methodology and ANOVA<\/a>. Scientific Reports, 15(1), 27215.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo39\">\n<h2>Optimization of a coal mine roof characterization model using machine learning<\/h2>\n<p><b>Summary:<\/b><br \/>\nRoof deformation in underground coal mine roadways results from several interacting factors, such as mining-induced stress, roof lithology, rock mass strength and excavation geometry, yet these are usually assessed in isolation. Working with an underground coal mine in Australia&#8217;s Bowen Basin, this study set out to merge geophysical log data and geotechnical information into one machine learning model that flags intersections prone to roof deformation.<\/p>\n<p>Data from 304 mined intersections were imported into Neural Designer to build a binary classification network whose target indicates whether roof deformation occurred. From 63 candidate inputs (including sonic-derived UCS, porosity, shale volume, depth of cover and roadway orientation), the authors narrowed the set to 20 and compared training algorithms and data splits, settling on the quasi-Newton method, a 60\/20\/20 split, L2 regularization and a single hidden layer of nine neurons.<\/p>\n<p>Checked against field monitoring data, the final model reached 77% classification accuracy and an ROC area of 0.78, and about 90% accuracy over the whole dataset at a 0.5 decision threshold. Its predictions were contoured into a geotechnical hazard map that helps site engineers anticipate roof conditions before mining.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Melville, M., Mondal, S., Nehring, M., &amp; Chen, Z. (2024). <a href=\"https:\/\/doi.org\/10.1016\/j.ijrmms.2024.105835\" target=\"_blank\" rel=\"noopener\">Optimization of a coal mine roof characterization model using machine learning<\/a>. International Journal of Rock Mechanics and Mining Sciences, 181, 105835.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo40\">\n<h2>Enhancing of uniaxial compressive strength of travertine rock prediction through machine learning and multivariate analysis<\/h2>\n<p><b>Summary:<\/b><br \/>\nUniaxial compressive strength (UCS) is a key design property of rock, but measuring it directly requires high-quality cores and costly, destructive testing. This study tested 61 travertine cores from the Dead Sea and Jordan Valley area and aimed to estimate UCS from simpler index tests using regression and machine learning.<\/p>\n<p>Alongside multiple linear regression, an M5 tree and k-nearest neighbors, the authors built a multilayer perceptron in Neural Designer with eight inputs from the physical and index tests and UCS as the target. The software recommended a 60\/20\/20 training\/selection\/testing split, the growing-neurons order selection algorithm set the hidden layer at 12 hyperbolic tangent neurons, and the network was trained with the Levenberg-Marquardt algorithm on the normalized squared error.<\/p>\n<p>Training stopped after 104 epochs, with normalized squared errors of 0.047, 0.035 and 0.086 on the training, selection and testing sets. Point load index and Schmidt hammer rebound were the strongest predictors of UCS; in the overall comparison the M5 tree model performed best, followed by KNN and then the neural network.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Malkawi, D. A., Rabab&#8217;ah, S. R., Sharo, A. A., Aldeeky, H., Al-Souliman, G. K., &amp; Saleh, H. O. (2023). <a href=\"https:\/\/doi.org\/10.1016\/j.rineng.2023.101593\" target=\"_blank\" rel=\"noopener\">Enhancing of uniaxial compressive strength of travertine rock prediction through machine learning and multivariate analysis<\/a>. Results in Engineering, 20, 101593.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo41\">\n<h2>Neural network modeling of rutting performance for sustainable asphalt mixtures modified by industrial waste alumina<\/h2>\n<p><b>Summary:<\/b><br \/>\nUsing recycled concrete aggregate (RCA) from demolition waste instead of natural aggregate makes asphalt paving more sustainable, but the resulting mixtures resist rutting less well. This study added waste alumina at 1%, 2% and 3% by mixture weight to RCA mixtures, evaluated them with Marshall, dynamic creep and wheel-tracking tests, and modeled rut depth with an artificial neural network.<\/p>\n<p>The ANN was built in Neural Designer (version 5.0.3) with five inputs (alumina content, index of plastic deformation, resilient modulus, flow number and load cycles) and rut depth as the target. The data were split 60\/20\/20 into training, selection and testing subsets, and the network, with scaling, perceptron and unscaling layers, was trained with the quasi-Newton method. Over 112 epochs, the training error dropped from 0.88 to 0.012 and the selection error from 1.97 to 0.019.<\/p>\n<p>The testing analysis showed close agreement between predicted and measured rut depth, with R2 = 0.995. In the experiments, 2% alumina was the best dosage: it reduced rut depth by up to 26% and raised the index of plastic deformation, resilient modulus and flow number by 30%, 16% and 23%, respectively.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Ismael, M. Q., Joni, H. H., &amp; Fattah, M. Y. (2023). <a href=\"https:\/\/doi.org\/10.1016\/j.asej.2022.101972\" target=\"_blank\" rel=\"noopener\">Neural network modeling of rutting performance for sustainable asphalt mixtures modified by industrial waste alumina<\/a>. Ain Shams Engineering Journal, 14(5), 101972.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo42\">\n<h2>Environmental impacts cost assessment model of residential building using an artificial neural network<\/h2>\n<p><b>Summary:<\/b><br \/>\nBuildings emit greenhouse gases across their whole life cycle, but the tools used to estimate energy use and environmental impact are slow and need detailed data that is rarely available at early design stages. This study developed EICAM, an Environmental Impacts Cost Assessment Model that lets designers of Saudi residential units estimate energy costs and carbon emissions quickly.<\/p>\n<p>The model was built in Neural Designer from 201 envelope design combinations simulated in DesignBuilder, split 70\/15\/15 for training, testing and validation. Six inputs (wall type, roof type, glazing type, window-to-wall ratio, shading device and orientation) feed a network with 13 hidden neurons that predicts four targets: annual energy cost, 20-year operational carbon, envelope embodied carbon and total carbon per square metre, trained with the conjugate gradient algorithm. The software&#8217;s model deployment tab was then used to validate the model on input combinations.<\/p>\n<p>The training error fell from 80.52 to 0.016 after 735 iterations, and regression of predicted against actual outputs gave R2 values from 0.993 to 0.998. Prediction errors were below 0.3% for every output, and for a new design combination the model gave results almost identical to DesignBuilder without the need for a 3D building model.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Hamida, A., Alsudairi, A., Alshaibani, K., &amp; Alshamrani, O. (2021). <a href=\"https:\/\/doi.org\/10.1108\/ecam-06-2020-0450\" target=\"_blank\" rel=\"noopener\">Environmental impacts cost assessment model of residential building using an artificial neural network<\/a>. Engineering, Construction and Architectural Management, 28(10), 3190\u20133215.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo43\">\n<h2>Predicting the contribution of recycled aggregate concrete to the shear capacity of beams without transverse reinforcement using artificial neural networks<\/h2>\n<p><b>Summary:<\/b><br \/>\nRecycled aggregate concrete reuses demolition waste, but design codes have not been adapted to it, partly because its shear behaviour differs from that of conventional concrete. This study set out to show that an artificial neural network can predict how much the concrete itself contributes to the shear capacity of recycled aggregate concrete beams without stirrups.<\/p>\n<p>The authors compiled 231 test results from the literature and used six inputs (beam width, effective depth, longitudinal reinforcement ratio, recycled aggregate replacement ratio, shear span-to-depth ratio and cylinder compressive strength) to predict the concrete shear strength. The model was developed with Neural Designer: inputs and output were scaled with the minimum-maximum method, the hidden layers used hyperbolic tangent activations, and training minimized a normalized squared error with regularization using a quasi-Newton algorithm, with about 80% of the data for training and selection and 20% for testing. Comparing several architectures led to a 6-6-2-1 network, which was then used for parametric and sensitivity studies.<\/p>\n<p>The network reached R2 values of 0.96 for training and 0.95 for testing, with mean relative errors of 2.91% and 3.77%. On 34 additional beams not used in development, the average ratio of measured to predicted strength was 1.01 with a coefficient of variation of about 8%, outperforming code and literature equations such as ACI-318, CSA, NZS and Zsutty. The parametric study confirmed that shear strength rises with reinforcement ratio and concrete strength and falls as the shear span-to-depth ratio and recycled aggregate content increase.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Ababneh, A., Alhassan, M., &amp; Abu-Haifa, M. (2020). <a href=\"https:\/\/doi.org\/10.1016\/j.cscm.2020.e00414\" target=\"_blank\" rel=\"noopener\">Predicting the contribution of recycled aggregate concrete to the shear capacity of beams without transverse reinforcement using artificial neural networks<\/a>. Case Studies in Construction Materials, 13, e00414.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo16\">\n<h2>Modeling resilient modulus of fine-grained materials using different statistical techniques<\/h2>\n<p><b>Summary:<\/b><br \/>\nThis study uses multiple linear regression, nonlinear regression, and backpropagation artificial neural networks to predict the resilient modulus of fine-grained materials, based on 3,709 soil samples from the Long-Term Pavement Performance (LTPP) database.<\/p>\n<p>The inputs are the confining pressure, nominal maximum axial stress, percentages of silt and clay, liquid limit, plasticity index, percent passing the No. 200 sieve, maximum dry density, and optimum moisture content. The neural networks were built in Neural Designer using its order selection and input selection methods, and its directional outputs were used to analyze how each parameter influences the resilient modulus.<\/p>\n<p>The artificial neural network built with Neural Designer gave the best approximation of the three techniques.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Khasawneh, M. A., &amp; Al-jamal, N. F. (2019). <a href=\"https:\/\/doi.org\/10.1016\/j.trgeo.2019.100263\" target=\"_blank\" rel=\"noopener\">Modeling resilient modulus of fine-grained materials using different statistical techniques<\/a>. Transportation Geotechnics, 21, 100263.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo44\">\n<h2>Tool life prognostics in CNC turning of AISI 4140 steel using neural network based on computer vision<\/h2>\n<p><b>Summary:<\/b><br \/>\nUninterrupted, intelligent machining needs low-cost and reliable estimates of tool wear and remaining tool life. This study built a direct, vision-based tool-condition monitoring system for dry CNC turning of AISI 4140 steel with carbide inserts. It combines an industrial camera mounted on the machine, image processing and an artificial neural network.<\/p>\n<p>Each tool image was denoised, contrast-enhanced and thresholded, and the white-pixel count of the worn area was measured. This count, together with cutting speed, feed and depth of cut, formed the four inputs of a neural network built in Neural Designer to predict flank wear and remaining tool life. The model used scaling, perceptron and unscaling layers (two perceptron layers, 11 neurons), L2 regularization, and training, selection and testing subsets, and the authors compared sigmoid and ReLU activation functions.<\/p>\n<p>The ReLU network did better than the sigmoid one. For remaining tool life it gave an RMSE of 0.871 min versus 1.437 min and a MAPE of 6.7% versus 13.5%; for wear, the MAPE was 3.4% versus 5.9%. Tool-life prediction accuracy was about 93% with ReLU and 86.5% with sigmoid, which the authors see as support for image processing plus ANN as a low-cost industrial monitoring method.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Bagga, P. J., Makhesana, M. A., Darji, P. P., Patel, K. M., Pimenov, D. Y., Giasin, K., et al. (2022). <a href=\"https:\/\/doi.org\/10.1007\/s00170-022-10485-9\" target=\"_blank\" rel=\"noopener\">Tool life prognostics in CNC turning of AISI 4140 steel using neural network based on computer vision<\/a>. The International Journal of Advanced Manufacturing Technology, 123(9\u201310), 3553\u20133570.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo45\">\n<h2>A neural network-based model for estimating the delivery time of oxygen gas cylinders during COVID-19 pandemic<\/h2>\n<p><b>Summary:<\/b><br \/>\nDuring the COVID-19 pandemic, demand for medical oxygen soared, and hospitals needed reliable estimates of when oxygen cylinder deliveries would arrive. This study developed a multilayer perceptron model to predict delivery times from the real logistics records of a company supplying hospitals across Saudi Arabia, and compared it with support vector machine (SVM) and multiple linear regression (MLR) models.<\/p>\n<p>Using Neural Designer, the authors modeled total delivery hours from 420 trip records (2019) with inputs such as destination, cylinder quantity, truck, number of trips between cities, number of hospital drops and distance; fuel consumption and driver were removed after correlation analysis. The data were split 60\/20\/20 for training, selection and testing, and the network was trained with the quasi-Newton method. Incremental order selection set the network complexity and growing-inputs selection chose the final input subset, followed by a sensitivity analysis of each numerical input.<\/p>\n<p>The model achieved R2 = 92.44% on the test set and 94.48% on 30 additional unseen validation records. Averaged over six random splits, the test R2 was 91.93% for the neural network, 94.06% for SVM and 90.57% for MLR. Total distance traveled was identified as the main contributor to delivery delays.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Ghaithan, A. M., Alarfaj, I., Mohammed, A., &amp; Qasim, O. (2022). <a href=\"https:\/\/doi.org\/10.1007\/s00521-022-07037-3\" target=\"_blank\" rel=\"noopener\">A neural network-based model for estimating the delivery time of oxygen gas cylinders during COVID-19 pandemic<\/a>. Neural Computing and Applications, 34(13), 11213\u201311231.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo46\">\n<h2>Exploring the Injury Severity Risk Factors in Fatal Crashes with Neural Network<\/h2>\n<p><b>Summary:<\/b><br \/>\nKnowing which conditions make road crashes fatal helps authorities target safety measures. Using three years (2017-2019) of crash records from 15 rural highways in Saudi Arabia, this study built a feed-forward neural network to classify crashes as fatal or non-fatal and to measure how strongly each risk factor affects severity.<\/p>\n<p>The ANN was implemented in Neural Designer on 12,566 crash records, of which only about 7% were fatal. Nine inputs chosen by their correlation with severity (including crash type, weather, traffic volume, vehicle type, number of lanes, road type, damage at the site, average speed and number of vehicles) fed a network with one hidden layer of six neurons and a logistic output. The data were split 80\/10\/10; to handle class imbalance, a weighted squared error penalized missed fatal crashes six times more than non-fatal ones, and the model was trained with the quasi-Newton method. The resulting model expression was then used for a one-factor-at-a-time sensitivity analysis.<\/p>\n<p>On the test data the model reached an overall accuracy of 77.5%, with 56.6% sensitivity and 79.2% specificity. Traffic volume, travel speed, weather, damage at the site, road and vehicle type, and pedestrian involvement were the most influential factors for crash severity.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Jamal, A., &amp; Umer, W. (2020). <a href=\"https:\/\/doi.org\/10.3390\/ijerph17207466\" target=\"_blank\" rel=\"noopener\">Exploring the injury severity risk factors in fatal crashes with neural network<\/a>. International Journal of Environmental Research and Public Health, 17(20), 7466.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo4\">\n<h2>Predicting Travel Times of Bus Transit in Washington, D.C. Using Artificial Neural Networks<\/h2>\n<p><b>Summary:<\/b><br \/>\nThis study aimed to develop travel time prediction models for transit buses to help decision-makers improve service quality and ridership. The authors used six months of Automatic Vehicle Location and Automatic Passenger Counting data from six Washington Metropolitan Area Transit Authority bus routes in Washington, D.C.<\/p>\n<p>The researchers used neural networks to approximate bus travel times, designing and training the different models in Neural Designer.<\/p>\n<p>The models were accurate, with a normalized squared selection error between 0.04 and 0.26 across the different periods of the day.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Arhin, S., Manandhar, B., &amp; Baba-Adam, H. (2020). <a href=\"https:\/\/doi.org\/10.28991\/cej-2020-03091615\" target=\"_blank\" rel=\"noopener\">Predicting travel times of bus transit in Washington, D.C. using artificial neural networks<\/a>. Civil Engineering Journal, 6(11), 2245\u20132261.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo47\">\n<h2>Effective Integration of Distance Courses Through Project-Based Learning<\/h2>\n<p><b>Summary:<\/b><br \/>\nTeaching a foreign language online makes it hard to keep students motivated and to link the language to their main subjects. Malyuga and Petrosyan tested whether project-based learning in a distance Business English course could connect the language with students&#8217; majors in the humanities and in the natural sciences at RUDN University.<\/p>\n<p>The study surveyed 150 first- to third-year students (75 from the humanities and 75 from the natural sciences) before and after a series of online project-based lessons, and compared them with a control group. Answers on a 10-point scale were mapped onto three criteria: motivation to learn Business English, performance in major subjects and performance in minor subjects. Neural Designer handled the statistics for the questionnaires: it computed each group&#8217;s average scores and ran a correlation analysis between the scores and the three criteria. The paper&#8217;s survey charts were also produced with it.<\/p>\n<p>Compared with the control group, students in the project lessons doubled their academic-performance scores and tripled their motivation to study English. The authors conclude that project work built around English is an effective way to connect major and minor subjects in distance learning. They also found that humanities students were more motivated to study related minor subjects.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Malyuga, E. N., &amp; Petrosyan, G. O. (2022). <a href=\"https:\/\/doi.org\/10.3389\/feduc.2021.788829\" target=\"_blank\" rel=\"noopener\">Effective integration of distance courses through project-based learning<\/a>. Frontiers in Education, 6, 788829.<\/li>\n<\/ul>\n<\/section>\n<section id=\"Articulo48\">\n<h2>Economic Land Utilization Optimization Model<\/h2>\n<p><b>Summary:<\/b><br \/>\nFarming in Egypt is under pressure from climate change and scarce water, yet decisions about what to plant on a plot rarely weigh life-cycle cost against water use. The authors developed ELUOM, an automated tool that pairs a database of more than 400 crops and their soil, water and climate needs with filtering steps and a genetic-algorithm optimizer, aiming to propose crop plans that maximize 20-year return while keeping water consumption low.<\/p>\n<p>Neural Designer was used in the model&#8217;s financial module to forecast crop cultivation costs four years ahead. For each crop, a neural network took six World Bank macroeconomic indicators as inputs (total reserves, inflation, exports of goods and services, GDP, official exchange rate, and agriculture, forestry and fishing value added) and predicted cost as the output, with 85% of the cases used for training and 15% for testing.<\/p>\n<p>The cost networks showed an average total error of about 6% on the testing cases, and their forecasts fed the economic optimization. On a 30-feddan farm in Giza, the complete ELUOM model proposed a crop mix that raised the projected net present value by 290% over the owner&#8217;s current practice, and on a 50-feddan vacant plot at the American University in Cairo it suggested six crops with an estimated 20-year NPV of roughly 72 million EGP.<\/p>\n<p><b>Source:<\/b><\/p>\n<ul>\n<li>Hosny, O. A., Dorra, E. M., Tarabieh, K. A., El Eslamboly, A., Abotaleb, I., Amer, M., et al. (2023). <a href=\"https:\/\/doi.org\/10.3390\/su15032594\" target=\"_blank\" rel=\"noopener\">Economic land utilization optimization model<\/a>. Sustainability, 15(3), 2594.<\/li>\n<\/ul>\n<\/section>\n<section>\n<h2>Related posts<\/h2>\n<\/section>\n","protected":false},"author":15,"featured_media":1840,"template":"","categories":[],"tags":[36],"class_list":["post-3408","blog","type-blog","status-publish","has-post-thumbnail","hentry","tag-tutorials"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Neural Designer in scientific research<\/title>\n<meta name=\"description\" content=\"Neural Designer can be used in your research. Here are some examples for different purposes.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Neural Designer in scientific research\" \/>\n<meta property=\"og:description\" content=\"Neural Designer can be used in your research. Here are some examples for different purposes.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/\" \/>\n<meta property=\"og:site_name\" content=\"Neural Designer\" \/>\n<meta property=\"article:modified_time\" content=\"2026-10-01T11:49:32+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp\" \/>\n\t<meta property=\"og:image:width\" content=\"780\" \/>\n\t<meta property=\"og:image:height\" content=\"520\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/webp\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:site\" content=\"@NeuralDesigner\" \/>\n<meta name=\"twitter:label1\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data1\" content=\"17 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/\",\"url\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/\",\"name\":\"Neural Designer in scientific research\",\"isPartOf\":{\"@id\":\"https:\/\/www.neuraldesigner.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp\",\"datePublished\":\"2026-04-27T23:20:26+00:00\",\"dateModified\":\"2026-10-01T11:49:32+00:00\",\"description\":\"Neural Designer can be used in your research. Here are some examples for different purposes.\",\"breadcrumb\":{\"@id\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#primaryimage\",\"url\":\"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp\",\"contentUrl\":\"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp\",\"width\":780,\"height\":520,\"caption\":\"Neural Designer articles\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\/\/www.neuraldesigner.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Blog\",\"item\":\"https:\/\/www.neuraldesigner.com\/blog\/\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"Neural Designer in scientific research\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.neuraldesigner.com\/#website\",\"url\":\"https:\/\/www.neuraldesigner.com\/\",\"name\":\"Neural Designer\",\"description\":\"No-code AI Platform\",\"publisher\":{\"@id\":\"https:\/\/www.neuraldesigner.com\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.neuraldesigner.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/www.neuraldesigner.com\/#organization\",\"name\":\"Neural Designer\",\"url\":\"https:\/\/www.neuraldesigner.com\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\/\/www.neuraldesigner.com\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/05\/logo-neural-1.png\",\"contentUrl\":\"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/05\/logo-neural-1.png\",\"width\":1024,\"height\":223,\"caption\":\"Neural Designer\"},\"image\":{\"@id\":\"https:\/\/www.neuraldesigner.com\/#\/schema\/logo\/image\/\"},\"sameAs\":[\"https:\/\/x.com\/NeuralDesigner\",\"https:\/\/es.linkedin.com\/showcase\/neuraldesigner\/\"]}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Neural Designer in scientific research","description":"Neural Designer can be used in your research. Here are some examples for different purposes.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/","og_locale":"en_US","og_type":"article","og_title":"Neural Designer in scientific research","og_description":"Neural Designer can be used in your research. Here are some examples for different purposes.","og_url":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/","og_site_name":"Neural Designer","article_modified_time":"2026-10-01T11:49:32+00:00","og_image":[{"width":780,"height":520,"url":"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp","type":"image\/webp"}],"twitter_card":"summary_large_image","twitter_site":"@NeuralDesigner","twitter_misc":{"Est. reading time":"17 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/","url":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/","name":"Neural Designer in scientific research","isPartOf":{"@id":"https:\/\/www.neuraldesigner.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#primaryimage"},"image":{"@id":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#primaryimage"},"thumbnailUrl":"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp","datePublished":"2026-04-27T23:20:26+00:00","dateModified":"2026-10-01T11:49:32+00:00","description":"Neural Designer can be used in your research. Here are some examples for different purposes.","breadcrumb":{"@id":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#primaryimage","url":"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp","contentUrl":"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/06\/neural-articles.webp","width":780,"height":520,"caption":"Neural Designer articles"},{"@type":"BreadcrumbList","@id":"https:\/\/www.neuraldesigner.com\/blog\/neural-designer-in-scientific-research\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.neuraldesigner.com\/"},{"@type":"ListItem","position":2,"name":"Blog","item":"https:\/\/www.neuraldesigner.com\/blog\/"},{"@type":"ListItem","position":3,"name":"Neural Designer in scientific research"}]},{"@type":"WebSite","@id":"https:\/\/www.neuraldesigner.com\/#website","url":"https:\/\/www.neuraldesigner.com\/","name":"Neural Designer","description":"No-code AI Platform","publisher":{"@id":"https:\/\/www.neuraldesigner.com\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.neuraldesigner.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/www.neuraldesigner.com\/#organization","name":"Neural Designer","url":"https:\/\/www.neuraldesigner.com\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/www.neuraldesigner.com\/#\/schema\/logo\/image\/","url":"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/05\/logo-neural-1.png","contentUrl":"https:\/\/www.neuraldesigner.com\/wp-content\/uploads\/2023\/05\/logo-neural-1.png","width":1024,"height":223,"caption":"Neural Designer"},"image":{"@id":"https:\/\/www.neuraldesigner.com\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/NeuralDesigner","https:\/\/es.linkedin.com\/showcase\/neuraldesigner\/"]}]}},"_links":{"self":[{"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/blog\/3408","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/blog"}],"about":[{"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/types\/blog"}],"author":[{"embeddable":true,"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/users\/15"}],"version-history":[{"count":11,"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/blog\/3408\/revisions"}],"predecessor-version":[{"id":24179,"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/blog\/3408\/revisions\/24179"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/media\/1840"}],"wp:attachment":[{"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/media?parent=3408"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/categories?post=3408"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.neuraldesigner.com\/api\/wp\/v2\/tags?post=3408"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}