Chenmu Zhang

Postdoctoral Researcher in Materials Science, Rice University

I build autonomous AI agents for computational materials science.

My work connects scientific computing, agent evaluation, and materials discovery.

Current research

MatClaw

An autonomous agent that writes code, runs simulations, and analyzes materials.

Given a task in plain language, MatClaw writes and runs Python against domain libraries and submits jobs to supercomputers. It works in a persistent Python session: loaded models and running jobs stay available between steps, and its tools are the code it writes. It is the agent behind the benchmark and Cu–In–Se study below.

MatClaw architecture: an autonomous LLM agent writes Python to orchestrate materials science tools and supercomputer jobs. View full-size figure ↗
Paper GitHub

CompMat-Bench

Testing AI agents on the steps of real materials research.

A benchmark of 94 tasks from 15 published studies, with expensive simulations run in advance and answers graded by fixed rules, without an LLM judge. With full guidance, agents pass 66–90% of single research steps. Long workflows with reduced guidance are harder; most failures are scientific errors.

CompMat-Bench: published studies become research steps, expensive simulations are precomputed, and fixed rules evaluate agent answers. View full-size figure ↗
Project page

Materials discovery in Cu–In–Se

Human-defined scientific tasks, carried out by an autonomous agent.

Applied to the system of the solar absorber CuInSe2, the agent connects MatterGen, machine-learning potentials, VASP, and Phonopy in one workflow. From a single request, it completed the screening shown here with no further human input.

The authors reviewed the scientific validity of the outputs and extended checks where the automated runs were incomplete.

Screening funnel: 1,000 generated structures, 578 after deduplication, 268 after machine-learning screening, 45 validated with DFT, 19 at or below the Materials Project hull, and 6 on the combined hull. View full-size figure ↗
From 1,000 generated structures to 45 DFT-validated candidates.
Project page Data

Autonomous research for band-gap prediction

An LLM research loop improves an expert-designed crystal graph network.

The agent repeatedly edits and trains a model on the MatBench band-gap benchmark of over 100,000 crystals. It achieved 0.148 eV MAE, ahead of all 17 expert-designed band-gap models trained on the benchmark.

This is the most accurate model trained without external pretraining in that comparison. Foundation models using external pretraining remain ahead. Alongside this result, I study the limitations of autonomous research loops and how to overcome them.

MatBench band-gap leaderboard: the agent's model achieves 0.148 eV MAE, ahead of the expert-designed models trained without external pretraining, behind foundation models using external pretraining. View full-size figure ↗
Band-gap prediction error. Lower is better; models with external pretraining are identified separately.
Paper GitHub

Previous research

Electron transport from first principles

Before turning to AI, I studied how electrons scatter in 2D semiconductors and nanoscale metals: phonons, surfaces, defects, and dielectric environments.

Electron-phonon coupling in 2D semiconductors

Electron–phonon

Quadrupole scattering and high-mobility 2D semiconductors.

 Electron-surface scattering in metal films

Electron–boundary

Surface scattering in nanoscale metals and graphene–Cu.

 Dielectric screening in heterostructures

Electron–environment

Dielectric screening in van der Waals heterostructures.

 Defect scattering in transition metal dichalcogenides

Electron–defect

Defect-limited mobility in transition metal dichalcogenides.