Learning

Model concrete properties using machine learning

Screen 28-day concrete strength before committing to laboratory trials

This surrogate model estimates 28-day compressive strength from seven constituent quantities. It helps concrete technologists compare candidate mixtures and focus physical testing on the most promising, feasible formulations.

42528-day laboratory specimens

7mixture-component inputs

85independent testing rows

4.88 MPatesting MAE

Try the model Download concrete_properties.csv

Compressive-strength testing is essential, but each result arrives only after batching, curing, and destructive laboratory testing. A data-driven surrogate can provide rapid estimates for screening and scenario analysis, while qualified mix design, trial batches and standards-based testing remain the basis for production and structural decisions.

Industrial challenge
Data set
Model
Training
Model selection
Testing
Deployment
Limitations
References

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.

Screen candidate mixturesEstimate whether a formulation is likely to approach the required strength before scheduling a complete trial.

Study material trade-offs: Explore how cement, supplementary cementitious materials, water, admixture, and aggregates jointly affect the response.

Prioritize laboratory workUse predictions to rank feasible candidates while retaining physical testing for verification and acceptance.

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

Concrete technologyReady-mix productionPrecast manufacturingMaterials laboratoriesQuality assuranceSustainable construction

Model role. This is a 28-day strength-screening surrogate. It supports engineering exploration but does not issue an approved mix design, verify code compliance, or replace batch trials and destructive tests.

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.

Download dataset: concrete_properties.csv

CSV variable	Engineering meaning	Role	Unit	Observed range
`cement`	Cement	Input	kg/m³	102 to 540
`blast_furnace_slag`	Blast-furnace slag	Input	kg/m³	0 to 359.4
`fly_ash`	Fly ash	Input	kg/m³	0 to 200.1
`water`	Water	Input	kg/m³	121.8 to 247
`superplasticizer`	Superplasticizer	Input	kg/m³	0 to 32.2
`coarse_aggregate`	Coarse aggregate	Input	kg/m³	801 to 1145
`fine_aggregate`	Fine aggregate	Input	kg/m³	594 to 992.6
`compressive_strength`	28-day compressive strength	Target	MPa	8.54 to 81.75

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

Distribution of 28-day concrete compressive strength across 425 specimens

Pearson correlations between concrete constituents and 28-day compressive strength — **Marginal relationships.** Cement has the strongest positive linear correlation, while water has the strongest negative one. These one-variable relationships do not describe mixture interactions.

Scatter chart of cement content and 28-day concrete compressive strength — **Cement alone is insufficient.** Strength varies substantially at similar cement contents because the complete formulation matters.

Validation note. The split is random. The file contains 416 unique input formulations and five duplicated formulation groups; two duplicate groups cross subset boundaries. A stronger evaluation should group identical or related mixtures so that one formulation cannot contribute rows to both model development and 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.

Initial concrete strength neural network with seven inputs, three tanh neurons and one output

The initial network provides a compact starting point. Model selection then changes the hidden-layer size; Section 5 shows the final testing and deployment model.

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 it reaches the minimum loss decrease criterion. The recorded training and selection errors are 0.098 NSE and 0.049 NSE.

Quasi-Newton training and selection error history for the initial concrete strength model

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, 0.042 NSE, occurs at 10 neurons; the corresponding training error is 0.083 NSE.

Training and selection errors for concrete models with one to ten hidden neurons

Because the best result occurs at the configured maximum, 10 neurons is the best tested size, not 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.

Final concrete strength neural network with seven inputs, ten tanh neurons and one output

6. Testing analysis

We recalculated the exported final model 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.

Testing rows	R²	MAE	RMSE	Bias	95th-percentile absolute error	Maximum absolute error
85	0.8214	4.88 MPa	6.06 MPa	+0.68 MPa	10.70 MPa	19.34 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.

Predicted versus measured 28-day concrete compressive strength on the testing subset

Example testing specimen. Row 421 contains 276.4 kg/m³ cement, 116.0 kg/m³ slag, 90.3 kg/m³ fly ash, 179.6 kg/m³ water, 8.9 kg/m³ superplasticizer, 870.1 kg/m³ coarse aggregate, and 768.3 kg/m³ fine aggregate. The measured strength is 44.28 MPa, and the model predicts 44.74 MPa, an absolute error of 0.46 MPa.

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 apply engineering constraints before inference, and do not use predictions as production specifications without validation.

Candidate formulation and material data

Range, mass-balance and feasibility checks

28-day strength surrogate

Trial batch, testing and engineering approval

Optimization case 1: reduce cement for a 45 MPa requirement

Operational question: what is the lowest-cement candidate that provides a modeled safety margin above a required 28-day strength of 45 MPa? The executed task minimizes cement, keeps the other constituents inside defined operating intervals, and requires a predicted strength of at least 55 MPa. The additional 10 MPa is an illustrative screening margin informed by the model error; it is not a certified characteristic-strength calculation or prediction interval.

Variable	Condition	Lower bound	Upper bound	Optimized value	Unit
Cement	Minimize	—	—	126.513	kg/m³
Blast-furnace slag	Between	50	220	219.710	kg/m³
Fly ash	Between	0	150	139.026	kg/m³
Water	Between	160	195	172.402	kg/m³
Superplasticizer	Between	0	15	14.277	kg/m³
Coarse aggregate	Between	850	1,075	1,060.860	kg/m³
Fine aggregate	Between	650	875	869.364	kg/m³
Compressive strength	≥	55	—	55.000	MPa

Result and engineering interpretation

Check	Candidate	Observed data domain	Assessment
Total binder	485.249 kg/m³	200.0–640.0 kg/m³	Inside observed interval
Water-to-binder ratio	0.355	0.235–0.900	Inside observed interval
Total constituent mass	2,602.152 kg/m³	2,194.6–2,551.0 kg/m³	51.2 kg/m³ above observed maximum

The optimizer reduces cement by moving slag close to its upper bound and using substantial fly ash, superplasticizer, and aggregate quantities. The model reproduces the requested 55 MPa threshold, but the constituent total is approximately 2.0% above the largest formulation in the dataset. The candidate therefore demonstrates a credible low-cement objective, yet it still requires a mass-balance or yield constraint and laboratory validation before it can be considered a feasible mix.

Optimization case 2: maximize strength with cement capped at 300 kg/m³

Operational question: what is the highest 28-day strength predicted inside the declared constituent limits when cement cannot exceed 300 kg/m³? This scenario maximizes the model output while bounding cement and every other constituent. It represents a producer seeking higher strength without moving to a high-cement formulation.

Variable	Condition	Lower bound	Upper bound	Optimized value	Unit
Cement	Between	200	300	299.998	kg/m³
Blast-furnace slag	Between	0	220	220.000	kg/m³
Fly ash	Between	0	150	149.499	kg/m³
Water	Between	155	190	178.930	kg/m³
Superplasticizer	Between	0	15	13.390	kg/m³
Coarse aggregate	Between	850	1,075	1,075.000	kg/m³
Fine aggregate	Between	650	875	874.735	kg/m³
Compressive strength	Maximize	—	—	80.156	MPa

Result and engineering interpretation

Check	Candidate	Observed data domain	Assessment
Total binder	669.497 kg/m³	200.0–640.0 kg/m³	29.5 kg/m³ above observed maximum
Water-to-binder ratio	0.267	0.235–0.900	Inside observed interval
Total constituent mass	2,811.552 kg/m³	2,194.6–2,551.0 kg/m³	260.6 kg/m³ above observed maximum

The solution reaches 80.156 MPa by driving cement, slag, fly ash, and both aggregates to, or very close to, their upper bounds. This is a typical boundary-seeking optimizer response. Although the individual values satisfy the declared limits, binder content and total mass lie outside every mixture represented in the data. The result is best interpreted as a diagnostic upper-bound scenario: it shows why a real maximum-strength optimization also needs total-binder, yield, workability, cost and durability constraints.

Comparison of the two executed cases

Scenario	Primary objective	Cement	Predicted strength	Total mass	Current status
Low-cement target-strength case	Minimize cement with predicted strength ≥55 MPa	126.513 kg/m³	55.000 MPa	2,602.152 kg/m³	Promising objective; revise mass balance
Cement-capped maximum-strength case	Maximize strength with cement ≤300 kg/m³	299.998 kg/m³	80.156 MPa	2,811.552 kg/m³	Boundary result; reformulate constraints

Decision rule. Response optimization identifies model-supported candidates, not approved concrete recipes. A candidate should advance only after joint-domain checks, mixture-volume and material-compatibility review, trial batching, curing, and standards-based strength testing.

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.

Predicted 28-day compressive strength as cement content changes at a fixed reference formulation

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.

Try the 28-day strength model

Enter one mixture or load a verified example. The browser evaluates the exact scaling, weights, and output transformation from the exported Python model.

Example mixture

Predicted 28-day compressive strength—Model estimate

Total constituent mass—Joint-domain check

Screening model — not a certified mix design, conformity assessment, or structural acceptance method.

Download and reproduce

Download Python model (ZIP)Download Neural Designer project + CSV (ZIP)Download concrete_properties.csv

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 not included in 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; the model does not capture workability, slump, density, durability, heat, shrinkage, cost, or embodied carbon.
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.