Researchers have developed MatUQ, a new dataset and evaluation framework for predicting materials properties using graph neural networks (GNNs). This work addresses the challenge of reliability in GNN predictions, especially when encountering out-of-distribution (OOD) data—materials with structures or compositions significantly different from those seen during training. MatUQ focuses on quantifying the uncertainty of these predictions, a crucial aspect for the safe application of artificial intelligence in the discovery and design of new materials.

The study highlights that while GNNs have shown great potential in predicting materials properties, their performance and reliability drop sharply when applied to OOD materials. MatUQ provides a benchmark for evaluating the ability of GNN models to not only predict properties but also to estimate the uncertainty associated with those predictions in OOD scenarios. This is fundamental for materials scientists to trust AI model recommendations and avoid synthesizing materials with unexpected or undesirable properties.

MatUQ includes a diverse set of materials properties and structures, designed to rigorously test the robustness of models against data variability. The proposed methodology allows for comparing different GNN approaches in their ability to generate accurate predictions and well-calibrated uncertainty estimates. This advance is a significant step towards developing more robust and reliable AI models for materials science, accelerating the discovery of new compounds with optimized properties for various technological applications.