SIGA: Self-Evolving Coding-Agent Adapters for Scientific Simulation

Researchers introduce SIGA, an AI adapter system that enables general coding agents to operate specialized scientific simulators without extensive domain training. The system achieves a 36x speedup compared to human experts on GEOS multiphysics simulator configuration, demonstrating that lightweight grounding layers can make general AI tools practical for scientific software.

SIGA addresses a fundamental friction point in scientific computing: domain scientists spend substantial time learning simulator-specific languages and interfaces before conducting actual research. The research demonstrates that coding agents possess foundational capabilities—file navigation, code editing, command execution—but lack the simulator's operational constraints and vocabulary. By injecting this domain-specific contract through retrieval mechanisms, procedural memory, and validation rules, the system bridges this gap efficiently.

This work sits at the intersection of AI capability enhancement and scientific software accessibility. Rather than training new models or creating domain-specific agents from scratch, SIGA uses lightweight adapters that evolve through self-improvement mechanisms. The validation that TreeSim scores improved from 0.720 to 0.789 on held-out test cases shows that the approach generalizes beyond initial training conditions. The transfer experiments across OpenFOAM and LAMMPS reveal important principles: validation becomes critical when structural completeness bottlenecks performance, while memory and retrieval matter most when domain correctness is the constraint.

For the scientific computing ecosystem, this represents a potential democratization path where non-expert users can leverage complex simulators through AI intermediaries. The 36x speedup compared to three-hour human expert configuration translates directly to research productivity gains. However, the practical impact depends on adoption barriers—whether simulation communities embrace AI-driven configuration approaches and how robust these systems perform on edge cases in real research workflows. The self-evolution capability suggests these adapters become more valuable over time as they process more trajectories, creating network effects within scientific communities.

• →SIGA achieves 36x speedup in scientific simulator configuration compared to human experts, reducing setup time from hours to minutes
• →Lightweight adapter layers combining retrieval, memory, and validation can turn general coding agents into practical scientific software operators
• →Self-evolution mechanisms allow adapters to improve from prior trajectories, matching or outperforming hand-designed configurations
• →Different interface bottlenecks require different adapter strategies—validation for structural completeness, memory/retrieval for domain correctness
• →Transfer across GEOS, OpenFOAM, and LAMMPS demonstrates generalizable principles for grounding agents to specialized scientific tools