Technical Papers Archive

Star-LAFT: An Agent-Driven Adaptive Framework for Fault-Tolerant High-Throughput Materials Informatics


Authors: Pin Chen, Hongjin Zhong, Guixin Guo, Jiang Li, Yuewen Huang, and Yutong Lu (Sun Yat-sen University)

Abstract: High-throughput computational materials science workflows execute millions of simulations on HPC clusters to build materials genome databases for data-driven discovery. However, these workflows face critical challenges: complex task dependencies reduce scheduling efficiency, load imbalances waste resources, and parameter-sensitive simulations suffer poor fault resilience. We present Star-LAFT, an agent-driven framework integrating large language models (LLMs) for automated workflow construction, dynamic resource optimization, and multi-level fault recovery. Key innovations include: (1) LLM-based automatic inference of computational parameters from material specifications; (2) chain-of-thought-driven dynamic resource allocation; and (3) a hybrid fault-tolerance agent that pairs an LLM diagnoser with a reward-guided recovery-action selector. Benchmarks show Star-LAFT achieves a 2.12× wall-time speedup and 10.08% higher task completion with only 2.76% additional cost. At production scale (~300k materials), it generated a million-entry database in 10 weeks at 94.3% completion rate, enabling effective downstream AI model training.


Presentation: file

Back to Technical Papers Archive Listing