MIT’s PhysiOpt system blends AI with physics to produce structurally sound 3D printed objects
Finite element analysis is closing the gap between AI-generated designs and real-world fabrication
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a system called PhysiOpt, that integrates generative artificial intelligence (genAI) with physics simulations, to enable users to generate 3D printable designs for personal items — such as cups, keyholders, and bookends — that hold up under real-world use.
The system takes a known limitation of genAI design tools to task: while platforms like Microsoft’s TRELLIS can produce three-dimensional models from text prompts or images, the resulting blueprints often fail structurally when fabricated. A chair design, for instance, might have disconnected components or insufficient support to bear weight.
PhysiOpt seeks to tackle this by running a physics simulation known as finite element analysis. It stress tests a 3D model and generates a heat map indicating structurally weak areas, before then making incremental adjustments to reinforce them without altering the object’s overall appearance or intended function.
Users enter a text description of what they want to create and specify how much force or weight the object should handle, as well as the fabrication material — such as plastic or wood — and how it will be supported. The system then delivers a refined 3D model in roughly 30 seconds.
“PhysiOpt combines GenAI and physically-based shape optimization, helping virtually anyone generate the designs they want for unique accessories and decorations,” stated Xiao Sean Zhan, an MIT electrical engineering and computer science PhD student and CSAIL researcher, and a co-lead author on the paper.
“It’s an automatic system that allows you to make the shape physically manufacturable, given some constraints. PhysiOpt can iterate on its creations as often as you’d like, without any extra training.”
The system relies on a pre-trained model rather than task-specific training, allowing it to draw on prior knowledge of shapes and aesthetics. This is a property the researchers referred to as “shape priors”.
“Existing systems often need lots of additional training to have a semantic understanding of what you want to see,” stated co-lead author Clément Jambon, also an MIT EECS PhD student and CSAIL researcher. “But we use a model with that feel for what you want to create already baked in, so PhysiOpt is training free.”
In comparative testing against DiffIPC, a method that similarly simulates and optimizes 3D shapes, PhysiOpt was nearly 10 times faster per iteration while producing more realistic outputs.
The researchers’ work was presented in December at the Association for Computing Machinery’s SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia, and was supported, in part, by the MIT-IBM Watson AI Laboratory and the Wistron Corp.
