Chenmu Zhang
Research Scientist in Materials Science, Rice University
Building autonomous agents for computational materials science
I develop LLM-based agents that autonomously write code, run simulations on supercomputers, and analyze results -- exploring how far AI can push the boundaries of computational materials research.
MatClaw uses a code-first architecture with retrieval-augmented generation to achieve ~99% API accuracy across domain libraries, and a four-layer memory system for coherent multi-day workflows.
LLM-agent research loop optimizing materials ML models
An LLM-agent research loop is set running: it repeatedly edits and trains a crystal graph neural network to predict band gaps. On the MatBench band-gap benchmark of over 100,000 crystals, the loop built the most accurate model trained without external pretraining (MAE = 0.148 eV), ahead of all 17 expert-designed band-gap models trained on the benchmark.
This shows that an AI agent can optimize an expert-designed ML model and push its accuracy further. Alongside this promise, I also examine the limitations of such loops and how to overcome them.
Previous work: electron transport from first principles
Before turning to AI, I spent years computing how electrons scatter in 2D semiconductors and nanoscale metals -- phonons, surfaces, defects, dielectric environments.
