According to MIT CSAIL, generative AI has reshaped digital content creation, but its leap into the physical world has lagged. While AI can generate personalized 3D models, those designs often fail once fabricated because they ignore real-world physics.

Researchers at the MIT Computer Science and Artificial Intelligence Laboratory argue that the core issue is mechanical integrity. AI-driven design tools typically focus on aesthetics, not whether an object will survive everyday use. MIT EECS PhD student and CSAIL engineer Faraz Faruqi has been working on this problem, developing systems that balance visual stylization with functional performance and tactile intent.

In collaboration with researchers from Google, Stability AI, and Northeastern University, Faruqi helped create MechStyle, an AI-powered system that generates objects that are both visually expressive and structurally sound. Users upload or select a 3D model – such as a vase or wall hook – then guide its transformation using text or image prompts. A generative model modifies the geometry, while MechStyle continuously simulates the physical impact of those changes to protect vulnerable regions. Once satisfied, users can fabricate the object via 3D printing.

Earlier 3D stylization methods were fragile. CSAIL researchers found that only 26% of stylized models remained structurally viable. “We want to use AI to create models that you can actually fabricate and use in the real world,” said Faruqi. “So MechStyle actually simulates how GenAI-based changes will impact a structure.”

The system integrates finite element analysis (FEA) with adaptive scheduling, running physics simulations only when geometry changes threaten structural stability. This approach allowed MechStyle to achieve up to 100% structural viability across 30 tested models with styles inspired by bricks, stones, and cacti. Users can explore ideas quickly in a freestyle mode, then switch to MechStyle mode for durability-aware refinement.

MechStyle can already generate functional decor, accessories, and assistive devices, though it can’t yet repair models that are structurally unsound from the start. Future versions may generate entirely new 3D models from scratch, expanding access for non-experts.

The work was supported by the MIT–Google Program for Computing Innovation and presented at the ACM Symposium on Computational Fabrication in November.