SHA-PF: New LLM framework cuts simulation-driven design evaluations

Expensive simulation-driven design is critical in engineering fields like antenna design, where each high-fidelity simulation can take hours or days. While LLMs have been used to automatically translate natural-language requirements into optimization formulations, they typically ignore whether the resulting formulation enables an efficient search process. The new SHA-PF framework addresses this gap by introducing a search hardness-aware objective. It scores candidate formulations based on how effectively they prioritize rare samples—those that offer the greatest potential for progress—thereby guiding the solver toward better designs with fewer evaluations.

SHA-PF operates by generating, repairing, and evolving formulations through LLM-based operations, all guided by the search hardness metric. The framework was tested on a real-world multi-objective benchmark and five expensive antenna design benchmarks. Results showed that formulations discovered by SHA-PF required substantially fewer simulation evaluations to reach design requirements compared to existing methods, including other LLM-based approaches. This work bridges the gap between language-driven problem formulation and practical optimization efficiency, offering a promising path for accelerating engineering design cycles.