TritonDFT: Automating DFT with a Multi-Agent Framework

A research team led by Zhengding Hu and 10 other collaborators from institutions including UCSD has introduced TritonDFT, a multi-agent AI framework designed to fully automate Density Functional Theory (DFT) calculations. DFT is a fundamental computational method in materials science used to predict material properties, but its practical application involves coordinating complex, multi-step workflows that require significant domain expertise. Existing tools and LLM-based solutions only automate isolated parts of this process. TritonDFT addresses this gap by providing a unified system that manages the entire workflow—from initial setup to final analysis—while adapting to diverse scientific tasks and optimizing the critical trade-off between calculation accuracy and computational cost.

The framework's architecture employs specialized AI agents that leverage an expert-curated, extensible workflow design, a novel Pareto-aware parameter inference engine, and multi-source knowledge augmentation from scientific literature and databases. This allows it to make intelligent decisions about simulation parameters and computational resources. Alongside the framework, the team released DFTBench, a comprehensive benchmark for evaluating AI agents across multiple dimensions including scientific expertise, trade-off optimization, High-Performance Computing (HPC) knowledge, and cost efficiency. With its open-source code and user interface, TritonDFT represents a significant step toward making advanced computational materials science more accessible and efficient, potentially accelerating the discovery of new materials for batteries, semiconductors, and other critical technologies.