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

2026-04-15 · Source: Artificial Intelligence · Field: Science & Research — Materials Science & Chemistry, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Researchers have developed an exascale workflow for materials discovery utilizing atomistic graph foundation models built on HydraGNN. This system jointly trains on 16 open first-principles datasets, encompassing over 544 million structures and 85+ elements, through a multi-task architecture with per-dataset heads and a scalable ADIOS2/DDStore data pipeline. Executing on Frontier, the team conducted six large-scale DeepHyper hyperparameter optimization campaigns in FP64, identifying a PaiNN-based lead model. This model facilitates billion-scale screening, evaluating 1.1 billion atomistic structures in 50 seconds, a task that would otherwise demand years of first-principles computation. The workflow also supports data-scarce fine-tuning for diverse downstream tasks and demonstrates strong- and weak-scaling across Frontier, Aurora, and Perlmutter, enabling rapid exploration of extensive chemical design spaces.

Key takeaway

For materials scientists and computational chemists seeking to accelerate discovery, this exascale workflow offers a path to rapidly screen vast chemical design spaces. You can leverage these graph foundation models to evaluate billions of atomistic structures in seconds, significantly compressing computational time compared to traditional first-principles methods, and enabling efficient data-scarce fine-tuning for new tasks.

Key insights

Exascale graph foundation models accelerate materials discovery by screening billions of structures in seconds.

Principles

• Multi-task learning improves generalization across diverse datasets.
• Scalable data pipelines are crucial for exascale training.

Method

The method involves joint training on 16 datasets using a multi-task architecture with per-dataset heads, optimized via DeepHyper, and leveraging a scalable ADIOS2/DDStore data pipeline for exascale performance.

In practice

• Use PaiNN-based models for atomistic structure prediction.
• Employ multi-task training for broad material property prediction.

Topics

• Exascale Computing
• Graph Foundation Models
• Materials Discovery
• Multi-Task Learning
• Atomistic Data

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Exascale Multi-Task Graph Foundation Models for Imbalanced, Multi-Fidelity Atomistic Data — Exascale multi-task graph foundation models accelerate materials discovery by enabling billion-scale screening and data-efficient fine-tuning.
• Bridging three-dimensional molecular structures and artificial intelligence with a conformation description language — ConfSeq enables large language models to process 3D molecular structures by converting them into discrete token sequences.
• LEAP: A closed-loop framework for perovskite precursor additive discovery — Coupling a domain-specialized LLM with active learning and expert review accelerates perovskite additive discovery under low-data conditions.
• Accelerating science with AI and simulations — AI is at a second inflection point, integrating language and multi-modal reasoning to accelerate scientific discovery.
• ASTEROID: A Spatiotemporal Information Transformer for Forecasting Multi-Step Time Series of Molecular Dynamics — ASTEROID uses a Transformer with STI Transformation to forecast molecular dynamics, improving accuracy and reducing computational cost.
• Platonic representation of foundation machine learning interatomic potentials — Diverse MLIPs converge to a shared latent geometry when physically supervised, enabling interoperability and diagnostics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.