Machine-Learning Regression for Concrete Compressive Strength: Model Comparison and Mix Optimisation on Open Data

CS-Opt is a declared-protocol study that benchmarks machine-learning regression models for concrete compressive strength on the UCI Concrete Compressive Strength dataset (1,030 mixes, Yeh, DOI 10.24432/C5PK67) — and then uses the winning model, carefully fenced, to propose trial mixes under Indian codal constraints.

Cement is India's construction backbone, and supplementary cementitious materials such as fly ash — available cheaply from thermal power plants — are central to both cost and sustainability targets. Mix design under IS 10262 and durability rules under IS 456 remain largely chart-and-iteration practice, while published machine-learning papers on concrete strength typically report a single model, a single random split and a single R², with no confidence intervals, no baseline comparison and no honest statement that a model trained on one country's materials may not transfer to another's aggregates and cements.

This study fixes all three defects. Ordinary least squares and ridge regression serve as honest baselines; random forest and gradient boosting are tuned by nested cross-validation with fixed seeds; results are reported as mean ± standard deviation across repeated folds; variable importance is measured by permutation, not by intuition; and a mix-optimisation stage searches for minimum-binder mixes meeting a target strength inside declared IS 456-consistent constraint envelopes, with an explicit trust region — the optimiser only proposes mixes inside the data's support, and says so.

The deliverable is a paper-style study with a reproducible Python pipeline: leakage-safe evaluation, model-comparison tables, importance analyses and a constraint-checked mix shortlist a lab could batch-test.

Concrete mix design in Indian practice remains chart-and-iteration under IS 10262, and data-driven strength models could support it — but the published machine-learning record on concrete strength is statistically weak: single models, single splits, no baseline anchoring, tuning that leaks into test data, no confidence intervals, and optimisation claims that ignore the model's valid operating range and codal durability constraints. There is no reproducible, declared-protocol study that (a) benchmarks linear and ensemble regressors on the standard public concrete-strength dataset under nested, repeated cross-validation with an untouched holdout, (b) interprets variable importance with fold-level uncertainty rather than intuition, and (c) proposes trial mixes that respect IS 456-consistent constraint envelopes strictly inside the model's trust region with prediction intervals attached. This dissertation fills that gap with a fully scripted pipeline, seed-fixed reproducibility, and honest reporting of where data-driven mix proposals stop being trustworthy.

Abstract

This dissertation benchmarks four regression families — ordinary least squares, ridge, random forest and gradient boosting — for predicting concrete compressive strength from mix proportions and age, on the public UCI Concrete Compressive Strength dataset under a frozen evaluation protocol: repeated 10-fold cross-validation with fixed seeds, nested cross-validation for hyperparameter tuning, an untouched holdout scored once at the end, and metrics reported as mean ± standard deviation (RMSE, MAE, R²). Permutation importance with confidence intervals is checked against water–cement-ratio expectations. A final stage optimises trial mixes for minimum binder at target strength inside declared IS 456-consistent constraint envelopes and the model's trust region, with prediction intervals per proposal. The study quantifies the ensemble advantage over honest baselines, the signal variables, and where mix proposals stop being trustworthy.

Introduction

India's infrastructure programme makes concrete strength prediction a practical economics question: every MPa of over-design is cement cost and embodied carbon, every under-design is structural risk. Mix design under IS 10262 is chart-and-iteration practice, while materials vary by source, so data-driven strength models appeal. Machine-learning regression offers that, but its published record on concrete is weakened by single-split evaluation habits.

Existing system and literature gap

Since Yeh's 1998 study on this very dataset, dozens of papers have applied regression to concrete strength, but accessible ones typically report one model on one split, tune on the test set implicitly, and never state the model's valid operating range. What is missing is a baseline-anchored, leakage-safe, seed-fixed comparison on the field's standard public dataset, with permutation-based interpretation and a codally constrained, trust-region optimisation stage — a study whose numbers regenerate exactly and whose claims carry their limits.

Proposed work

• Ingest the UCI dataset with full provenance (DOI, feature definitions, units); audit ranges, units and the uneven age distribution before modelling.
• Implement the four model families in leakage-safe scikit-learn pipelines (preprocessing inside every fold); tune the ensembles by nested cross-validation; freeze seeds and grids in configuration.
• Evaluate by repeated 10-fold CV plus a single untouched holdout; report RMSE, MAE and R² as mean ± standard deviation; test the ensemble-versus-linear gap across folds.
• Interpret with permutation importance (with fold-level variability) and partial dependence for cement, water and age; contrast against water–cement-ratio expectations.
• Optimise trial mixes: minimise binder subject to target strength and declared IS 456-consistent constraints (w/c bounds, cementitious minimums, volume closure), inside the training data's trust region, with prediction intervals per proposal.

Feasibility

• Technical: everything runs in Python (pandas, NumPy, scikit-learn, SciPy) on a 16 GB laptop; the full nested-CV campaign completes in about an hour.
• Data: the UCI dataset is public, citable and the field's standard benchmark; no lab time is required for the modelling stages.
• Economic: zero licence cost; the optimiser's output is a shortlist for a later lab phase.
• Operational (research): the protocol, seeds and constraint envelopes are frozen before evaluation; every table regenerates from two commands — the audit trail an M.Tech viva demands.

In scope

• The UCI Concrete Compressive Strength dataset (1,030 mixes, 8 mix/age features, strength in MPa) as the modelling corpus, with documented provenance.
• Four regression families under one frozen, leakage-safe evaluation protocol; permutation importance and partial-dependence interpretation.
• Mix optimisation inside declared constraint envelopes consistent with IS 456 durability provisions and IS 10262 mix-design logic, restricted to the training data's trust region.
• Honest transferability discussion: the dataset's materials are not Indian aggregates and cements, and the report says so throughout.

Out of scope

• Laboratory casting and testing of proposed mixes — the optimiser's output is a shortlist for an optional later phase, not validated mixes.
• New-material regimes (geo-polymer concretes, recycled aggregates) absent from the dataset; data collection is future scope.
• Deep-learning architectures; the model families chosen span the practical spectrum and keep the study interpretable.
• Long-term strength evolution beyond the dataset's age feature, and creep/shrinkage behaviour entirely.

The work follows a quantitative, comparative-experimental research design in six phases over 24 weeks, with the evaluation protocol (folds, repeats, seeds, grids, constraint envelopes) frozen at the end of Phase 1 — before any model is scored.

Validation inside the study: unit tests for metric implementations against hand-computed examples and for the constraint checker against boundary mixes; a leakage test asserting that no scaling statistic crosses fold boundaries; a trust-region test asserting the optimiser rejects any proposal outside the training support; and a full re-run reproducing every table from fixed seeds.

The pipeline is organised around one rule — nothing computed from evaluation data may enter training — enforced structurally rather than by good intentions.

flowchart TD
  A["UCI Concrete dataset\nDOI 10.24432/C5PK67"] --> B["Ingestion and audit\nranges, units, age distribution"]
  B --> C[("Versioned data store\nprovenance recorded")]
  C --> D["Leakage-safe pipelines\nscaler inside every fold"]
  D --> E["Model families\nOLS, ridge, forest, boosting"]
  E --> F["Nested CV tuning\nfrozen grids, fixed seeds"]
  F --> G["Repeated 10-fold evaluation\nRMSE, MAE, R² as mean ± std"]
  G --> H["Untouched holdout\nscored once, at the end"]
  G --> I["Permutation importance\nand partial dependence"]
  E --> J["Trust-region fence\ntraining-data support only"]
  C --> K["Constraint envelopes\nIS 456-consistent bounds"]
  J --> L["Mix optimiser\nminimum binder at target strength"]
  K --> L
  L --> M["Mix shortlist with\nprediction intervals"]
  I --> N["Results tables and figures"]
  H --> N
  M --> N

Three design decisions matter at the viva. First, every preprocessing step lives inside the scikit-learn pipeline object, so cross-validation refits it per fold and a unit test asserts no statistic leaks across fold boundaries — the single most common defect in published strength-modelling work. Second, the holdout is scored exactly once, after the final configurations are frozen, and the scoring run is logged; anything tuned against it would make the headline number decorative. Third, the mix optimiser sits behind an explicit trust-region fence: it rejects any proposal outside the training data's support, because an extrapolated prediction for a mix combination the model has never seen is not a design recommendation, it is a guess — and the shortlist ships with prediction intervals and this caveat printed beside it.

• Cast and test the optimiser's shortlist in a college structures lab, closing the modelling-to-laboratory loop with measured 7/28-day strengths.
• Augment the corpus with published Indian-material datasets (fly-ash blends from thermal-plant sources) under documented provenance, and test transfer.
• Extend to supplementary cementitious materials beyond fly ash (GGBS, silica fume) as more open data becomes available.
• Add cost and embodied-carbon objectives to the optimiser for a multi-objective sustainable-mix study.
• Study probabilistic strength targets: specify mixes for a required characteristic strength with a declared safety-margin policy instead of a point target.