The Department of Energy just funded something that sounds like science fiction: an AI system that will design its own experiments, run them on supercomputers, control robotic labs, and validate its discoveries. No human in the loop required.
FORUM-AI (Foundation Models Orchestrating Reasoning Agents to Uncover Materials Advances and Insights) is a $10 million, four-year project led by Lawrence Berkeley National Laboratory. The goal is building the first “full-stack, agentic AI system for materials science research and discovery.”
What “Agentic” Actually Means Here
Most AI in science today is passive. You feed it data, it finds patterns, you decide what to do next. FORUM-AI is designed to work differently.
The system will use three types of AI working together:
- Generative AI to propose new materials and write research plans
- Reasoning models to evaluate which approaches are worth pursuing
- Agentic models that execute the plan - running simulations, controlling lab equipment, analyzing results
The idea is that FORUM-AI could take a goal like “find a better battery cathode material” and autonomously explore hundreds of research pathways in parallel. It would generate hypotheses, select computational methods, run the simulations, then synthesize promising candidates in automated labs.
The Infrastructure Behind It
This isn’t just one lab playing around. The project connects:
- Supercomputers at NERSC (Berkeley), OLCF (Oak Ridge), and ALCF (Argonne)
- A-Lab at Berkeley - a fully automated facility for computer-controlled inorganic powder synthesis
- The Materials Project - a database with verified properties of hundreds of thousands of materials
The system pulls from curated, validated data rather than relying on model memory. When FORUM-AI looks up a material property, it queries an actual database with experimentally confirmed values. This is meant to reduce hallucination risks that plague LLMs in scientific applications.
Why This Matters
Lithium-ion batteries took decades to develop from concept to commercial use. The promise of agentic AI in materials science is compressing those timelines dramatically - potentially from decades to months for some discoveries.
Principal investigator Anubhav Jain leads the project. He’s a staff scientist at Berkeley Lab and associate director of the Materials Project, which has already cataloged properties for over 150,000 materials.
The collaboration includes Berkeley Lab, Oak Ridge National Laboratory, Argonne National Laboratory, MIT, and Ohio State University.
The Fine Print
Four years and $10 million is ambitious but not unlimited. The project aims to deliver a “fully integrated end-to-end platform” by completion, but how well agentic AI handles the messiness of real materials research remains to be seen.
Previous automated synthesis systems have struggled with unexpected results and edge cases that human researchers would catch immediately. Whether reasoning models can replicate that intuition is an open question.
The system will include “inspectable AI reasoning traces” - meaning researchers can review the AI’s decision-making process. This transparency is critical for scientific credibility. A discovery only matters if other labs can reproduce it.
There’s also the question of what happens when the AI fails. Materials science involves expensive equipment and sometimes hazardous materials. The handoff between AI recommendations and physical execution will need robust safety checks.
Still, if FORUM-AI works as intended, it could change how materials research happens. Instead of a researcher spending months on one hypothesis, AI could explore thousands simultaneously. The bottleneck shifts from human cognition to computational resources - and computing power keeps getting cheaper.