Reproduce “AI-Guided Human-In-the-Loop Inverse Design of High Performance Engineering Structures” on OpenResearch

About this paper

Dat Quoc Ha, Md Ferdous Alam, Markus J. Buehler, Faez Ahmed, Josephine V. Carstensen

Researchers at MIT developed HiTopAI, an AI-guided human-in-the-loop framework for topology optimization that predicts user-preferred modification regions. This system accelerates the design process, achieving a 39% higher buckling load with only a 4% increase in design time, while enhancing designs for complex performance metrics and manufacturability.

Key insights

• AI can effectively predict human preferences for design modifications within an interactive topology optimization workflow, significantly reducing the manual trial-and-error bottleneck.
• The trained ML model demonstrates robust generalizability, accurately identifying critical features in unseen and non-standard design aspect ratios (up to 1.5:1) despite being trained exclusively on square domains.
• The AI co-pilot exhibits emergent behavior, identifying multiple critical features (e.g., multiple near-maximum-length members) even when explicitly trained on annotations for only a single feature, suggesting it learns deeper structural patterns.

Summary by alphaXiv.