Answers are grounded in this paper and cite their sources.
Machine-learning (ML) surrogates for composite properties are often validated with random train/test splits, although experimental composite datasets are small and heavily replicated. We quantify the consequences on an open dataset of 420 fibre-reinforced polymer records that reduce to 28 material configurations (fabric × × forming technology) and only 42 distinct rows, predicting warp-direction tensile strength from constituent, textile and processing descriptors. A random record-level split gives R² = 1.00 R 2 = 1.00 for a random forest, whereas leave-one-configuration-out (LOCO) validation gives R² = 0.71 R 2 = 0.71 (RMSE 274 MPa) at record level and 0.77 at configuration level. Simple weave-aware predictors match or outperform every learned model: a weave-class mean reaches R² = 0.91 R 2 = 0.91 , and a rule of mixtures with a weave-dependent knock-down factor calibrated within each training fold reaches 0.96. Single 20% group hold-outs are unstable (random-forest median R² = 0.65 R 2 = 0.65 , interquartile range 0.06-0.80) because the result depends on whether a unidirectional carbon configuration is withheld. Tree-ensemble spread and quantile-forest intervals under-cover a nominal 90% level, and group-weighted conformal intervals reach it only at widths of about 1500 MPa. Validation should be carried out at the level of the true unit of replication, here the material configuration, and learned surrogates should be benchmarked against physics-informed categorical baselines before being embedded in additive-manufacturing or AI-assisted design workflows.
No takes yet. Share an insight, caveat, or question.
Byron Oviedo (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: