📝 2025 Annual Report

Draft annual report covering scientific accomplishments, publications, presentations, and goals at ALCF.

• Goals for Next Year (2026)
• Goals from Last Year (2024)
• Contributions to ALCF
• Publications
• Presentations
• Posts
• Organizational Efforts
• Mentoring
• Scientific / Technical Accomplishments
• References

Goals for Next Year (2026)

• Build out generic training services for science teams
• Continue to push on resilient / fault-tolerant training techniques

Goals from Last Year (2024)

• Continue to contribute to division(/lab)-wide efforts
• Continue to work with application teams to efficiently scale on ALCF systems
• [WIP] Publish retrospective on initial pre-training of AuroraGPT

Contributions to ALCF

• AERIS: Argonne Earth Systems Model for Reliable and Skillful Predictions

  • ACM Gordon Bell Prize Finalist (co-author)
  • Contributed to model development, performance analysis, and scaling studies

• MProt-DPO: Breaking the ExaFLOPS Barrier for Multimodal Protein Design with DPO

  • Finalist for the 2024 ACM Gordon Bell Prize (first-author)

• AuroraGPT

  • Co-lead Models and Training team with Venkat Vishwanath
  • Ongoing writeup of pre-training efforts
  • Successfully pre-trained:
    • AuroraGPT-7B on 2T tokens
    • AuroraGPT-2B on 4T tokens (ongoing)

• Catalyst for:

  • Arvind Ramanthan’s INCITE Project (FoundEpidem)
  • Zheng Zhang’s ALCC Project
  • Rao Kotamarthi’s ALCC Project

• Member of Software Committee

• Intro to HPC Undergraduate Bootcamp:

  • Project lead for Intro to {AI, HPC} for Science

Publications

1. AERIS: Argonne Earth Systems Model for Reliable and Skillful Predictions (Hatanpää et al. (2025))¹
2. Aurora: Architecting Argonne’s First Exascale Supercomputer for Accelerated Scientific Discovery (Allen et al. (2025))
3. HiPerRAG: High-Performance Retrieval Augmented Generation for Scientific Insights (Gokdemir et al. (2025))
4. Automated Tuning for HMC Mass Ratios (Torsiello et al. (2025))
5. MOFA: Discovering Materials for Carbon Capture with a GenAI and Simulation-Based Workflow (Yan et al. (2025))
6. MProt-DPO: Breaking the ExaFLOPS Barrier for Multimodal Protein Design with DPO (Dharuman et al. (2024))²

Presentations

• Scientific AI at Scale: AI for Science @ Open SkAI 2025
• Scientific AI at Scale: Distributed Training @ Open SkAI 2025
• Large Scale Training on Diverse Accelerators @ Scalable Deep Learning, SIAM AN2025
• LLMs on Aurora: 🌌 AuroraGPT @ 2025 ALCF INCITE GPU Hackathon
• LLMs on Aurora: 🍋 ezpz @ 2025 ALCF INCITE GPU Hackathon
• AuroraGPT: Foundation Models for Science @ Foundation Models for the Electric Grid
• Parallel Training Methods @ AI-for-Science on Supercomputers
• AuroraGPT @ 2024 ALCF Hands-On HPC Workshop
• Machine Learning and Foundation Models at Scale @ 2024 ALCF Hands-On HPC Workshop

Posts

• 📊 pbs-tui : TUI for PBS Job Scheduler Monitoring
• 🍹 BlendCorpus + TorchTitan @ ALCF
• 🏗️ Building PyTorch 2.8 from Source on Aurora
• 🚧 Frameworks Issue with numpy > 2
• 🔥 Building PyTorch 2.6 from Source on Aurora
• 🪛 Torchtune on Aurora
• 🚑 Torchtune Patch on Aurora
• 💾 Converting Checkpoints

Organizational Efforts

• Organizer for:
  • SC25 Workshop: High Performance Python for Science at Scale (HPPSS)
  • SC25 Tutorial: Accelerating and Scaling Python for HPC
  • SC24 Workshop: High Performance Python for Science at Scale (HPPSS)
• Served as reviewer for:
  • HiPC 2025
  • SPIGM @ NeurIPS
  • ML4PS Workshop @ NeurIPS’24
  • AI4Science Workshop @ NeurIPS’24
  • GenBio Workshop @ NeurIPS’24
  • AI4Science Workshop @ ICML’24

Mentoring

• Khalid Hossain: Supported Khalid’s successful transition from postdoc to staff
• Joseph Frimpong: Postdoc in Center for Nanoscale Materials
• Hung Nguyen: Graduate student @ UIUC

Scientific / Technical Accomplishments

References