Murat Akdoğan
Glass transition temperature (T g) is a key thermophysical property in polymer informatics, yet many machine learning (ML) studies focus on point prediction accuracy without explicitly evaluating reliability under chemical novelty. Here, we evaluate two established descriptor-based regressors, gradient-boosted trees (XGBoost) and support vector regression (SVR), on a 410-sample simulation-derived polymer data set using stratified, scaffold-based, and fingerprint-clustered validation regimes. We combine learning curves, applicability-domain diagnostics, split conformal prediction (SCP), subgroup coverage analysis, model-specific descriptor-importance analysis, and interval-aware triage metrics. Performance degraded and variability increased under novelty-enforcing splits, with SVR showing more stable point-prediction behavior than XGBoost in several regimes; this trend is interpreted as benchmark-specific, not as general model-class superiority. Conformal intervals maintained near-nominal marginal coverage but were often too wide for fine-grained candidate ranking, and subgroup diagnostics revealed local reliability limitations in low-similarity, high-T g, or chemistry-specific subsets. Thus, conformal intervals are best interpreted as conservative risk indicators for uncertainty-aware triage, not as high-resolution screening tools. Descriptor-importance analyses highlighted chemically plausible feature families related to topology, polarity, surface area, heteroatom content, and electronic-state descriptors, but these attributions are treated as model-level plausibility diagnostics, not physical validation of T g mechanisms. Overall, this work provides a reproducible reliability-assessment workflow for small-data polymer T g prediction against MD-derived labels, with experimental validation required before deployment against measured T g data.