AutoDFT: A Closed-Loop Multi-Agent Framework for Autonomous DFT Calculations

AutoDFT is a closed-loop multi-agent framework that automates density functional theory (DFT) calculations by embedding LLM reasoning throughout the entire computational lifecycle, rather than just the planning phase. The system achieves 94.1% success on a 34-task benchmark and enables non-experts to obtain reliable computational chemistry results by dynamically adapting to failures and unexpected outcomes.

AutoDFT represents a meaningful advancement in automating scientific computation by addressing a critical limitation of previous LLM-based approaches: the inability to adapt mid-execution. Traditional automation tools lock in an execution plan upfront, requiring human intervention when convergence fails or unexpected physics emerges. AutoDFT's architecture—featuring a strategic planner for skeletal objectives, a step planner for just-in-time parameter generation, and a monitor-recover-reflect cycle—enables genuine closed-loop adaptation.

This development builds on growing recognition that large language models excel not only at initial reasoning but also at iterative problem-solving. The chemistry and materials science communities have long struggled with DFT's computational demands and the specialized expertise required to troubleshoot calculations. By democratizing access to reliable first-principles computations, AutoDFT lowers barriers for experimental scientists lacking deep computational backgrounds.

The benchmark results demonstrate both scope and rigor: 94.1% success across nine DFT calculation types on VASPBench, plus validation on established materials databases for electronic, magnetic, and energetic properties. This breadth suggests the framework generalizes beyond narrow, pre-planned scenarios—a critical distinction from previous implementations.

Looking forward, the success of AutoDFT signals that sophisticated scientific workflow automation requires continuous adaptation loops rather than static planning. Similar architectures could reshape laboratory automation across chemistry, materials science, and physics. The framework's reliance on GPT-5.2 raises questions about dependency on specific model versions and accessibility for researchers with computational budget constraints.

• →AutoDFT integrates LLM reasoning into every stage of DFT calculations, not just planning, enabling mid-execution adaptation to failures and unexpected results.
• →The framework achieved 94.1% task success on a 34-task benchmark spanning nine DFT calculation types, validating both breadth and quantitative reliability.
• →Closed-loop multi-agent design with strategic planning, just-in-time parameter generation, and dynamic recovery cycles eliminates need for expert manual intervention.
• →AutoDFT democratizes access to reliable computational chemistry by enabling experimentalists without deep computational expertise to obtain first-principles results.
• →Successful automation of iterative scientific workflows suggests similar architectures could transform laboratory automation across chemistry, materials, and physics research.