Materials graph neural networks reveal efficiency in parameter use
Measuring trainable degrees of freedom in materials graph neural networks: a random-subspace intrinsic dimension analysis
Artificial Intelligence
Summary
Materials graph neural networks (GNNs) are tools that predict properties of materials based on their structure. The authors show that looking only at final accuracy doesn’t tell us how efficiently these networks use their adjustable parts. They test several GNNs by limiting the number of adjustable parameters and see how well the networks recover performance as more parameters are available. This method uncovers which tasks and models depend more on having a lot of adjustable parts and which do well even with fewer.
What this means in practice
- •For materials scientists: Evaluate how effectively GNN models use trainable parameters for different materials property predictions before choosing a model.
- •For machine learning engineers: Design more efficient GNN architectures by testing sensitivity to trainable parameter subspaces to optimize model size and training resources.
Authors
Shehroz Ahmad Shoaib, Kangming Li
Abstract
Final predictive accuracy is the standard basis for comparing graph neural networks (GNNs) in materials-property prediction, but it does not show how strongly performance depends on access to trainable parameter-space directions. Here, we introduce trainable-degree dependence as a complementary characterization of materials GNN learning. Using random-subspace intrinsic-dimension analysis, we train CGCNN, ALIGNN, and DimeNet++ in randomly oriented parameter subspaces across six prediction tasks and measure how performance recovers as independent trainable degrees of freedom are restored. The resulting recovery curves separate endpoint accuracy from the trainable-dimensional demand required to recover it. They reveal distinctions that final errors alone miss: metallic classification and log-bulk-modulus regression recover near-reference performance from small fractional subspaces, formation-energy and band-gap prediction show stronger architecture dependence, and phonon prediction is most sensitive to dimensional restriction. Dataset-size sweeps show that band-gap models require larger fractional subspaces as training data grows, whereas formation-energy and bulk-modulus responses are more stable. A width sweep shows that fractional thresholds can remain stable while absolute threshold dimensions increase with model size. Random-subspace analysis therefore provides a targeted stress test for how materials GNNs use their optimization space.