Answers are grounded in this paper and cite their sources.
Machine learning models for self-compacting mortar strength are trained on pooled data with unknown noise and never compared against the response surface methodology (RSM) their data was designed for. This study runs that comparison on two published central composite design campaigns of self-compacting high early strength mortar (124 mixes, replicate readings retained): whether machine learning beats the response surface once referenced to noise, whether either transfers between campaigns, and whether symbolic regression finds missed structure. Five models, RSM, LASSO, Gaussian process regression (GPR), gradient boosting, and symbolic regression on the RSM residual, were evaluated across grouped cross validation, an author defined holdout, and bidirectional transfer with bootstrap intervals. Inside the wider, well centred design, RSM is close to the noise floor for 24-hour compressive strength, and only GPR separates from it, narrowly; the other three remain tied. Pooling the second campaign's two superplasticiser designs into one surface collapses RSM's fit toward a null model on two of four targets, while GPR and LASSO remain informative: regularisation compensates for the orthogonality pooling destroys. Under transfer, behaviour is target specific: the unclipped quadratic fails severely for strength, over fourteen times a naive baseline, but beats it for slump flow; LASSO is the only model whose strength interval excludes zero. Symbolic regression recovers essentially no residual structure. The conclusion is conditional: a response surface is trustworthy up to its noise floor inside the design it was fitted on, and a regularised alternative is safer once pooling or extrapolation breaks that condition.
No takes yet. Share an insight, caveat, or question.
Pedram et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: