AMS-LPP: adaptive model selection for lightfastness property prediction in textile materials using similarity-guided transfer learning and AutoML

Abstract Accurate material property prediction can accelerate materials discovery, but conventional experimental methods remain costly and time-intensive, and machine learning alternatives are often degraded by domain shifts across heterogeneous datasets. Transfer Learning (TL) is effective only when source and target domains are similar, while AutoML ignores inter-domain relationships entirely, so neither approach alone adapts to varying domain similarity. This paper proposes AMS-LPP, a similarity-guided framework that dynamically switches between TL and AutoML based on domain similarity quantified via Mahalanobis distance: a pretrained TabNet model is fine-tuned with Bayesian hyperparameter optimization under high similarity, while AutoML is invoked under low similarity to ensure robust generalization. Experiments on multi-domain textile datasets for lightfastness prediction show that TL achieves $$R^2 > 0.99$$ with up to 35% less training time under high and moderate similarity, while AutoML attains the best generalization ( $$R^2 = 0.998$$ ) under low similarity. SHAP-based explainability further shows predictions are driven by physically meaningful optical and color-related features. These results demonstrate that AMS-LPP provides a scalable, efficient, and interpretable solution for similarity-aware model selection in heterogeneous material property prediction.