Automat: AI Agent Designs Compositional Descriptors for Materials Science
Researchers have introduced a novel AI framework named Automat, which employs a large language model (GPT-5.5 through OpenAI Codex) to independently create composition-only descriptors aimed at forecasting material properties. The system suggests, executes, and enhances descriptor strategies grounded in chemistry, focusing on band gaps in inorganic substances and Curie temperatures in ferromagnetic materials. This study, available on arXiv (2605.14671), broadens the scope of autoresearch from merely selecting models to the design of input descriptors.
Key facts
- Automat is an autoresearch framework for designing compositional descriptors.
- Uses GPT-5.5 via OpenAI Codex as the coding agent.
- Descriptors are derived solely from chemical formulas.
- Evaluated using a random forest workflow.
- Applied to predict experimental band gaps and Curie temperatures.
- Published on arXiv with ID 2605.14671.
- The agent iteratively proposes, implements, and tests descriptor strategies.
- Restricted to information derivable from chemical formulas.
Entities
Institutions
- OpenAI
- arXiv