Model concrete properties using machine learning

1. Industrial challenge

This is an approximation problem: the model maps seven constituent quantities to the continuous target compressive_strength, measured in MPa after 28 days.

Potential users include ready-mix and precast producers, concrete technologists, materials laboratories, civil and structural engineers, quality teams and low-carbon construction specialists.

2. Data set

The source UCI Concrete Compressive Strength dataset contains 1,030 laboratory results with age as an input. This example uses all 425 records tested at 28 days, so age is fixed and omitted from the local file. The semicolon-delimited CSV contains seven numerical inputs, one numerical target and no missing values.

The configured random split assigns 255 rows to training, 85 to selection and 85 to testing.

3. Neural network

The initial model standardizes seven inputs using mean and standard deviation, evaluates three tanh hidden neurons and returns one linear strength estimate before unscaling to MPa. The initial 7–3–1 architecture contains 28 trainable parameters.

The initial network provides a compact starting point. Model selection subsequently changes the hidden-layer size; the final testing and deployment model is shown in section 5.

4. Training strategy

The initial network minimizes normalized squared error with L2 regularization weight 0.01 using the quasi-Newton method. Optimization stops after 78 epochs because the minimum loss decrease criterion is reached. The recorded training and selection errors are 0.098 NSE and 0.049 NSE.

Normalized squared error is useful during optimization, but MPa-based metrics are used later so engineering users can interpret the practical size of prediction errors.

5. Model selection

The growing-neurons task compares hidden-layer sizes from 1 to 10, using three trials per size. The smallest recorded selection error, 0.0429 NSE, occurs at 10 neurons; the corresponding training error is 0.083 NSE.

Because the best result occurs at the configured maximum, 10 neurons is the best tested size rather than proof of a global optimum. The final 7–10–1 network contains 91 trainable parameters and is the model used for testing, optimization and Python export.

6. Testing analysis

The exported final model was recalculated on the 85 rows marked as testing. R² is the squared correlation reported by the Neural Designer goodness-of-fit analysis; MAE and RMSE express errors directly in MPa.

A training-mean baseline has an RMSE of 14.22 MPa, so the neural model materially improves on a constant prediction. Nevertheless, only 56.5% of testing rows fall within ±5 MPa, and the largest error is operationally significant. This supports screening use, not automatic acceptance.

7. Model deployment

A practical workflow starts with a physically feasible formulation and ends with laboratory confirmation. The model can screen or rank candidates, but engineering constraints must be applied before inference and predictions must not become production specifications without validation.

Comparison of the two executed cases

Cement sensitivity at a reference formulation

At 265 kg/m³ cement, 86 kg/m³ slag, 62 kg/m³ fly ash, 183 kg/m³ water, 7 kg/m³ superplasticizer, 956 kg/m³ coarse aggregate and 764 kg/m³ fine aggregate, the model predicts 37.47 MPa. The directional plot varies cement while holding the other quantities fixed.

Interpret the curve only across the observed cement interval of 102–540 kg/m³. The portion beyond 540 kg/m³ is model extrapolation and should not support a design decision.

Download and reproduce

Video walkthrough

The original tutorial remains available as a step-by-step Neural Designer walkthrough. Numerical results on this page reflect the newly exported project.

8. Scope and limitations

• The model represents 28-day strength only. It cannot estimate early-age or long-term strength because age is fixed and absent from the inputs.
• The data comes from a public laboratory collection and does not establish transfer to a specific plant, cement source, aggregate grading, admixture product, curing regime or test procedure.
• Random row splitting and duplicated formulations can produce a more optimistic test than validation by unseen formulation family, production batch or chronological campaign.
• The target is compressive strength only; workability, slump, density, durability, heat, shrinkage, cost and embodied carbon are not modelled.
• Inputs are constituent masses, but the model does not enforce volume balance, total yield, water-to-binder limits, aggregate proportions or material compatibility.
• The observed range of each variable does not define a valid multidimensional mixture domain. Joint feasibility must be checked separately.
• Production deployment requires local data, grouped or time-aware validation, drift monitoring, uncertainty limits and periodic recalibration.
• Predictions support screening and experimentation; qualified personnel and standards-based testing remain responsible for mix approval and structural acceptance.

References

• Yeh, I.-C. Concrete Compressive Strength Dataset, UCI Machine Learning Repository, DOI: 10.24432/C5PK67.
• Yeh, I.-C. Modeling of strength of high-performance concrete using artificial neural networks, Cement and Concrete Research 28(12), 1797–1808 (1998).
• Neural Designer testing analysis: goodness-of-fit and regression metrics.
• Neural Designer model deployment: directional outputs, response optimization and model export.