AutoSiMP: Autonomous Topology Optimization from Natural Language via LLM-Driven Problem Configuration and Adaptive Solver Control
In an innovative leap towards automating structural optimization, researchers have introduced AutoSiMP, a groundbreaking pipeline that converts natural language problem descriptions into validated binary topologies without the need for manual configuration. This advancement, detailed in the paper (arXiv:2603.27000v1), showcases the application of large language models (LLMs) in engineering domains traditionally reliant on expert knowledge.
Overview of AutoSiMP
AutoSiMP encompasses five integral modules that work in concert to achieve its objectives:
- LLM-Based Configurator: This module parses plain-English prompts and translates them into a structured specification that includes geometry, supports, loads, passive regions, and mesh parameters.
- Boundary-Condition Generator: It generates solver-ready degrees of freedom (DOF) arrays, force vectors, and passive-element masks, preparing the problem for computational analysis.
- Three-Field SIMP Solver: A robust solver featuring Heaviside projection and pluggable continuation control, allowing for sophisticated optimization processes.
- Eight-Check Structural Evaluator: This component assesses the structural topology through various metrics, including connectivity, compliance, grayness, volume fraction, convergence, and three additional informational quality metrics.
- Closed-Loop Retry Mechanism: Ensures that if a problem does not meet the required quality checks, it can be reconfigured and solved iteratively.
Evaluation and Performance
The evaluation of AutoSiMP was conducted across three critical axes:
- Configuration Accuracy: The configurator demonstrated a remarkable ability to produce valid specifications across ten diverse problems, achieving a median compliance penalty of only +0.3% compared to expert ground truth.
- Controller Comparison: In a study involving 17 benchmarks and six different controllers, the LLM controller performed admirably with a 76.5% pass rate, although it resulted in the lowest median compliance. In contrast, a deterministic schedule achieved a 100% pass rate with a slightly higher compliance of +1.5%.
- End-to-End Reliability: Utilizing the schedule controller, all problems configured by the LLM passed every quality check on the first attempt, indicating a high level of reliability without the need for retries.
Conclusion and Future Work
AutoSiMP stands out as the first system to successfully complete the entire process from natural-language problem description to validated structural topology. The implications of this work are vast, potentially transforming how engineers approach topology optimization by reducing reliance on manual input and expertise.
Upon journal acceptance, the complete codebase, all specifications, and an interactive web demo will be made available to the public, paving the way for further advancements in automated engineering solutions.