Exascale Multi-Task Graph Foundation Models for Imbalanced, Multi-Fidelity Atomistic Data

Researchers have developed an exascale workflow using graph foundation models trained on 544+ million atomistic structures to accelerate materials discovery. The system can screen 1.1 billion structures in 50 seconds—a task requiring years of traditional computation—and demonstrates strong transfer learning capabilities across diverse chemical applications.

This research represents a major computational breakthrough in accelerating materials discovery through machine learning at unprecedented scale. By training a single foundation model on 16 open first-principles datasets covering 85+ elements, the team created a versatile tool that reduces materials screening from years to seconds. The multi-task architecture with per-dataset heads allows the model to handle imbalanced and multi-fidelity data—a critical challenge in computational chemistry where experimental datasets vary widely in size and quality.

The achievement builds on a decade-long trend of applying deep learning to materials science, where traditional quantum mechanical simulations became the computational bottleneck. Foundation models have recently transformed fields like natural language processing and computer vision by learning transferable representations from massive datasets. This work extends that paradigm to atomistic systems, showing that a single pre-trained model can fine-tune effectively across twelve chemically diverse downstream tasks with minimal retraining.

For the materials discovery industry, this accelerated screening capability dramatically expands the accessible chemical design space. Pharmaceutical companies, battery manufacturers, and semiconductor firms could use billion-scale screening to identify promising candidates before expensive lab validation. The demonstrated scaling across three exascale supercomputers (Frontier, Aurora, Perlmutter) indicates the approach is reproducible and infrastructure-agnostic.

Looking ahead, the key challenge is bridging the gap between computational predictions and real-world validation. While the model achieves impressive speed, practitioners must validate its predictions experimentally. The work suggests foundation models will become standard tools in materials science workflows, potentially commoditizing certain discovery phases and shifting value toward validation and synthesis expertise.

• →A foundation model trained on 544+ million atomistic structures enables screening of 1.1 billion materials in 50 seconds versus years of traditional computation.
• →Multi-task learning across 16 datasets allows effective transfer to twelve chemically diverse downstream tasks with minimal fine-tuning.
• →The workflow demonstrates strong scaling properties across three major supercomputers, suggesting broad adoption potential in materials discovery pipelines.
• →Precision-performance tradeoffs (BF16/FP32/FP64) enable flexible deployment strategies balancing accuracy requirements with computational efficiency.
• →The approach addresses the data scarcity problem in computational chemistry by leveraging transfer learning from large-scale pre-training.