Abstract
Machine learning accelerates composite-material design by predicting mechanical properties from composition, reducing costly physical testing.Conventional models return a single value with no indication of its trustworthiness, and the reliability tools that could supply this, namely conformal intervals and domain-of-applicability detectors, are seldom tested for transfer to a new material.This paper presents CALIBER, a reliability-aware system that augments a gradient-boosting core with conformal prediction intervals and a domain-of-applicability detector, and introduces a novelty-error coupling criterion: a pre-deployment test that determines, from calibration data alone, whether a domain detector will be informative for a dataset.CALIBER is evaluated on two cementitious composite datasets, 1,005 concrete mixtures and 307 steel-fibre-reinforced concrete specimens, under a leakage-free protocol.Consistent with the criterion, the conformal interval transfers between the two datasets whereas the domain detector does not, because the steel-fibre-reinforced concrete error is governed by one physical variable rather than featurespace novelty.