Skip to content
Science

2026 Summary

This page provides a high-level summary of the meetings held in 2026.

Key Highlights

  • Stormer Weather Model Scaling: Conducted an empirical study on the Stormer weather forecasting model, discovering that large-scale transformer training becomes communication-dominated. Gradient accumulation was identified as the most critical optimization for reducing GPU communication overhead.
  • Protein Engineering Bottlenecks: Analyzed resource usage in generative AI for protein design, identifying significant GPU/CPU idle time caused by the sequential feedback loop between generative and predictive models—a common challenge in hybrid science-plus-ML workloads.
  • Stakeholder Review & Taxonomy: Advanced the "Stakeholder Review" paper to identify gaps in current scientific AI benchmarks and initiated the development of a benchmark taxonomy categorized by scientific domain, data type, and evaluation metric.
  • Time Series Dataset Catalog: Successfully completed a comprehensive catalog of 804 time series datasets to support future benchmark development.

Major Decisions

  • Publication Standards: Agreed to utilize the ACM format for the stakeholder review paper to ensure professional management and dissemination.
  • Strategic Focus: Decided to decline a special journal edition offer from Frontiers on Earth Science to maintain focus on core expertise.
  • DOE Alignment: Determined that "Area 18" (performance prediction) of the DOE AI and HPC solicitation is the primary alignment point for MLCommons benchmark efforts.