<ul data-eligibleForWebStory="true"><li>Graph foundation models using graph neural networks are being used for atomistic modeling to handle multi-source, multi-fidelity data during pre-training.</li><li>Recent studies employ multi-task learning where shared layers process atomistic structures initially regardless of the source, routing them to different decoding heads for data-specific predictions.</li><li>A new multi-task parallelism method is proposed to distribute each head across computing resources with GPU acceleration, implemented in the open-source HydraGNN architecture.</li><li>The method was trained on over 24 million structures from five datasets and tested on supercomputers like Perlmutter, Aurora, and Frontier, showing efficient scaling on heterogeneous computing architectures.</li></ul>

Multi-task parallelism for robust pre-training of graph foundation models on multi-source, multi-fidelity atomistic modeling data

Discover more