← All research

Agent-driven discovery in Cu–In–Se

Humans define the scientific tasks, and an autonomous LLM agent carries out the calculations. We apply this workflow to Cu-In-Se, the system of the solar absorber CuInSe₂, and search for low-energy structures on both the Cu-rich and the Cu-poor side. The search reveals an electronic asymmetry between the two sides.

Zhenkun Yuan* and Chenmu Zhang*

Department of Materials Science and NanoEngineering, Rice University
* These authors contributed equally.

Paper forthcoming Data MatClaw

01 / Workflow

Humans define the tasks. The agent connects the methods.

The agent joins structure generation with MatterGen and icet, machine-learning potentials, DFT with VASP, phonon calculations with Phonopy and a band-gap model. It writes the Python that connects them and runs the jobs on a supercomputer.

From a single request, the agent completed the first task in 183 steps with no further human input. The authors reviewed the scientific validity of the outputs and extended the checks where the automated runs were incomplete.

The human provides three task descriptions and the agent runs the structure generation, screening, DFT and analysis tools. A timeline shows 184 agent steps with decisions and self-corrections. — (a) The task-oriented workflow. The human provides task descriptions, and the agent orchestrates the structure-generation, simulation and analysis tools. (b) Agent behavior after the first request. Gray bars show when each activity was performed, blue circles mark decisions made by the agent, and red diamonds mark self-corrections after environment or workflow failures.

02 / Screening

From 1,000 generated structures to 45 DFT-validated candidates.

The first task screens generated structures through deduplication, a machine-learning potential, and DFT. Of the 45 DFT-validated candidates, 19 lie at or below the Materials Project hull. Six remain on the hull built from the Materials Project entries and all 45 candidates.

Screening funnel: 1,000 generated structures, 578 after deduplication, 268 after screening with a machine-learning potential, 45 validated with DFT, 19 below the Materials Project hull, 6 on the combined hull. — Hierarchical screening of the generated structures in the first task. Of 1,000 generated structures, 45 were validated with DFT. 19 lie at or below the Materials Project (MP) hull, and 6 remain on the hull built from the MP entries and all 45 candidates.

03 / Electronic structure

The search reveals an electronic asymmetry in Cu–In–Se.

Within 25 meV per atom of the reference hull, the Cu-rich side holds 202 structures, all classified as metallic, and 198 of them have no match in the Materials Project, OQMD, Alexandria or GNoME. The Cu-poor side holds 40, including 34 semiconductor candidates.

Four composition directions from CuInSe2, the energies of enumerated structures relative to the reference hull, their predicted band gaps, and a comparison of two machine-learning potentials. — (a) Four composition directions from CuInSe₂: to CuSe (D1), InSe (D2), In₂Se₃ (D3) and Cu₂Se (D4). (b) Energy of the enumerated structures relative to the reference hull. Shading marks the 25 meV per atom window. (c) Predicted band gaps of the low-energy structures along D1 and D3. Blue marks metals and orange marks semiconductors. (d) Energies from two machine-learning potentials, MACE and MatterSim, for the same low-energy structures.

Research paper

Off-stoichiometric energetic and electronic asymmetry in ternary Cu-In-Se: task-oriented materials discovery by an autonomous LLM agent

Zhenkun Yuan* and Chenmu Zhang*
Department of Materials Science and NanoEngineering, Rice University
* These authors contributed equally.