24 / 2026-11-05 18:25:34

Benchmarking Graph Neural Networks for Metallic-Glass Recommendation

metallic glass,graph neural networks,recommendation system,small data,graph pre-propagation

The discovery of metallic glasses (MGs) is constrained by a vast composition space and a comparatively small body of experimentally validated systems. To guide exploration of this space, statistical learning methods such as Random Forest have been used to predict glass-forming ability from composition-derived descriptors [1]. However, conventional feature-only models do not explicitly encode relationships among elements across known MG systems. Material networks retain these relationships by representing elements as nodes and reported glass-forming combinations as connections [2]. Building on this representation, graph learning combines elemental features with network connectivity to recommend candidate MG systems [3].

Despite progress in graph-based MG recommendation, model selection for small materials graphs remains largely empirical. This study therefore benchmarks 24 graph neural networks (GNNs) to assess architectural choices and the contribution of topology to small-data recommendation.