DISTAL Improves Structure-Agnostic Materials Property Prediction

This paper introduces DISTAL, a novel dual-prior framework designed to improve materials property prediction, especially in low-data environments where labeled samples are scarce and crystal structure information might be unavailable during early-stage screening. Traditional high-accuracy models often depend on crystal structures, limiting their applicability. DISTAL overcomes this by enabling structure-agnostic prediction. The framework operates in two main stages. First, it employs self-supervised compositional pretraining to learn transferable representations from a vast virtual composition space, utilizing 145 composition-derived descriptors. Second, it distills structural knowledge from a pre-trained, structure-aware ALIGNN teacher model into a composition-conditioned student model. This allows the benefits of structural priors to be leveraged during training without needing structural inputs at inference time. By integrating explicit compositional descriptors, pre-trained latent features, and distilled structural features, DISTAL captures complementary signals, leading to superior performance. Across 39 benchmark tasks, the best multimodal configuration of DISTAL improved upon the reference benchmark on 37 tasks, demonstrating its robust capability for composition-only prediction in materials informatics.