講演情報
[10a-N302-1]Improving Out-of-Distribution Generalization in Property Prediction via Transfer Learning
〇(P)Enda Xiao1, Terumasa Tadano1 (1.NIMS)
キーワード：
Transfer learning、Out-of-Distribution、Extrapolation
Machine-learning (ML) models have become an important tool for materials-property prediction, but their performance often deteriorates when applied to compounds outside the distribution represented in the training data. Improving out-of-distribution (OOD) generalization is therefore a critical challenge for data-driven materials discovery. In this work, we investigate a transfer-learning (TL) framework that leverages large-scale pretrained models to improve robustness under distribution shift. This TL approach exhibits superior performance on both in-distribution (ID) and OOD benchmarks. ID performance is evaluated using the MatBench and benchmark provided by Ito et al., while OOD generalization is assessed through leave-one-group-out (LOGO) and leave-one-period-out (LOPO) tests, in which compounds containing selected elemental groups or periods are excluded from training and used exclusively for evaluation. The results show that representations learned from large-scale datasets can be effectively transferred to downstream property-prediction tasks, leading to substantial improvements in extrapolation performance under distribution shifts. These findings demonstrate that transfer learning provides a practical strategy for developing robust and generalizable materials-property prediction models. We further discuss characteristic failure modes of transfer learning and conventional training-from-scratch approaches under large domain shifts. The underlying code is available as open-source package at https://github.com/nims-spin-theory/MLIP_FTL.