Designing an Explainable agent for Generative Design Workflows with Physics based Reasoning

Generative design tools powered by large language models (LLMs) increasingly operate as opaque "black boxes," relegating human designers to passive bystanders. This opacity erodes "reflection on action" and critical thinking, leading professionals to blindly accept fundamentally flawed designs. Meaningful engagement requires a tool that generates explicit, verifiable, physics-based reasoning behind its actions. However, current LLMs and vision-language models perform poorly in rigorous engineering contexts, relying on linguistic pattern matching rather than true physical awareness, causing frequent hallucination and infeasible outputs. Both the human cognitive gap and the technical performance gap converge on the same necessity: an explainable generative model grounded in geometric and physical reasoning. To test this hypothesis, this research transforms standard LLMs and VLMs into a physics-based CAD agent. The LLM was fine-tuned on a structural dataset paired with PyBullet rigid-body simulation data, explicitly encoding real-world constraints including gravity, weight, and center of mass. The interface enables designers to concurrently draft, inspect per-brick physical reasoning, and intervene at any stage. Validated through computational diagnostics and an expert user study measuring structural stability and cognitive engagement, the prototype shows a glimpse of restoring designers' "reflection on action," enabling active negotiation, reframing, and creative agency, while demonstrably improving structural stability.