ComMat: Datasets and Benchmarks from complex materials for graph machine learning

ICLR 2026 Conference Submission21554 Authors

Keywords: graph machine learning, complex materials, graph neural networks

TL;DR: we present datasets of complex materials and benchmarks for graph machine learning

Abstract: Recent research has demonstrated the efficacy of graph learning over a wide spectrum of materials, including molecular graphs, crystals, mechanical metamaterials, and strongly disordered systems. In this work, we draw attention to the broad class of \textit{complex materials}, which combine order and disorder and fall outside the above categories, yet have shown superior properties throughout the materials science literature. We present a Complex Materials Benchmark (ComMat), including three graph datasets of complex materials from experimental and computational research studies, unifying distinctly developed data-to-graph pipelines under a standardized graph-based representation. We then quantitatively show that these graphs are fundamentally different from existing materials datasets. We design various predictive tasks to advance machine learning (ML) methods, including experimentally measured properties, simulated mechanical response, and structural awareness. Extensive benchmark experiments are conducted over popular graph learning models, revealing their limitations and the need for further development in handling complex material networks. ComMat is openly released to accelerate ML research and innovation in complex material design.

Primary Area: datasets and benchmarks

Submission Number: 21554