FORUM-AI: An Agentic AI Stack to Compress Materials Discovery Timelines
Developing lithium-ion batteries took decades. A new multi-institutional effort - FORUM-AI (Foundation Models Orchestrating Reasoning Agents to Uncover Materials Advances and Insights) - aims to shorten that cycle using AI, high-performance computing, and automated experiments.
Led by the Department of Energy's Lawrence Berkeley National Laboratory with partners at Oak Ridge National Laboratory, Argonne National Laboratory, MIT, and The Ohio State University, FORUM-AI supports the DOE Genesis Mission to accelerate discovery across energy, science, and national security. The four-year, $10M project, funded under the Scientific Discovery through Advanced Computing (SciDAC) program, will build an open-source, general-purpose AI platform for materials and the physical sciences.
What FORUM-AI Will Do
"FORUM-AI aims to be the first full-stack, agentic AI system for materials science research and discovery. It will help scientists at every step of energy materials research, from hypothesis generation and computer simulations to laboratory experiments and analysis." - Anubhav Jain
Instead of testing one idea at a time, researchers will be able to generate and evaluate hundreds of hypotheses and research plans in parallel. The system will run large-scale simulations on leadership supercomputers - including Berkeley Lab's National Energy Research Scientific Computing Center (NERSC), Oak Ridge Leadership Computing Facility (OLCF), and Argonne Leadership Computing Facility (ALCF) - and trigger robotic experiments when appropriate.
Three AI classes drive the workflow:
- Generative models to propose hypotheses, write plans, and produce structured inputs.
- Reasoning models to decompose problems, interpret results, and recommend next steps.
- Agentic models to execute actions, from launching simulations to controlling lab equipment.
Guardrails Against Bad Science
Hallucinations don't help anyone in the lab. FORUM-AI builds in safeguards so outputs align with ground truth and accepted methods.
- Verified data sources: For properties like band gaps, agents query curated databases rather than rely on model memory.
- Transparent plans: Researchers can inspect and edit the agent's research plans and reasoning traces before anything runs.
- Physics-based tools: Property predictions use benchmarked simulation codes and community-standard workflows.
Energy Use: Smaller Models, Same Results
The team will apply model distillation to train smaller models that mirror the behavior of large ones. That cuts compute needs and enables local execution.
In practice, a distilled model can run on a laptop or be attached to instruments (for example, an XRD setup) for on-the-fly guidance. This makes AI assistance accessible to more researchers while reducing energy costs.
Why the National Labs Matter
The labs have spent years building the pieces needed for AI-assisted research. The Materials Project provides a large, trusted database of materials properties. Open-source tools now automate complex simulation workflows that once required extensive manual tuning.
On the experimental side, facilities like Berkeley Lab's A-Lab support computer-controlled inorganic powder synthesis. And the DOE complex provides the compute backbone: NERSC, OLCF, and ALCF - all DOE Office of Science user facilities - enable large-scale reasoning and simulation at speed. Learn more about NERSC here.
What Success Looks Like by Year Four
The goal is an end-to-end autonomous platform that discovers new materials and insights for applications. The system will propose compositionally complex candidates that meet target properties, then orchestrate simulation and synthesis to validate and improve them.
Think property-driven loops: set a target (e.g., battery cyclability or charge/discharge rate capability), run the first round, analyze results, and iterate. FORUM-AI will use lab automation (such as A-Lab) to close the loop and climb the performance gradient with each cycle.
What's Next
Future plans include connecting FORUM-AI to experimental user facilities like light sources. The assistant could run prep experiments before a user arrives, or support them during beamtime to make each session more efficient and accurate.
The team is also exploring broader applications: catalysis, structural materials, semiconductors, and interfaces. The aim is a shared platform the community can extend across scientific domains.
Practical Takeaways for Research Teams
- Frame your problem as a set of target properties and constraints; this makes hypothesis generation and triage effective.
- Prepare clean, well-documented datasets and simulation workflows so agents can query, reproduce, and build on your work.
- Identify which steps in your pipeline benefit most from automation (structure generation, parameter sweeps, lab protocols).
- Pilot distilled models near instruments to reduce latency and cost while keeping decision logic close to the data source.
- Keep a human-in-the-loop for plan review and anomaly checks; use reasoning traces as a standard part of your lab notebook.
- If you use DOE compute, track allocations and queue times early; parallelism only helps if the runs can actually launch.
Project Snapshot
- Name: FORUM-AI (Foundation Models Orchestrating Reasoning Agents to Uncover Materials Advances and Insights)
- Lead: Lawrence Berkeley National Laboratory
- Partners: Oak Ridge National Laboratory, Argonne National Laboratory, Massachusetts Institute of Technology, The Ohio State University
- Program: DOE SciDAC
- Facilities: NERSC, OLCF, ALCF (DOE Office of Science user facilities)
- Focus: Open-source, general-purpose AI platform for materials and physical sciences
Skill Up
If you're building team capability in AI-assisted research workflows, explore targeted training options at AI Research Courses.
Your membership also unlocks:
