Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a new system, called PhysiOpt, that combines generative artificial intelligence (genAI) with physics simulations to create 3D-printable designs for everyday objects. The system addresses a key limitation of current genAI tools: their inability to consistently produce designs that are structurally sound and capable of sustaining real-world use.
While genAI models like Microsoft’s TRELLIS can generate creative 3D models from text or image prompts, these designs often lack a fundamental understanding of physics, resulting in unstable or impractical objects. A chair generated by such a model, for example, might have disconnected parts or be unable to support weight. PhysiOpt aims to bridge this gap by rapidly testing and refining designs to ensure their viability before fabrication.
“PhysiOpt combines GenAI and physically-based shape optimization, helping virtually anyone generate the designs they want for unique accessories and decorations,” explained Xiao Sean Zhan, an MIT electrical engineering and computer science PhD student and CSAIL researcher, and co-lead author of a paper detailing the work. “It’s an automatic system that allows you to produce the shape physically manufacturable, given some constraints.”
The system allows users to input a desired object and its intended use, or upload an image, and then generates a realistic 3D model within approximately 30 minutes. Researchers demonstrated PhysiOpt’s capabilities by creating a flamingo-shaped drinking glass, complete with a handle and base resembling the bird’s leg. During the design process, PhysiOpt automatically made subtle adjustments to ensure structural integrity.
PhysiOpt operates by iteratively optimizing a design through a process called “finite element analysis,” a physics simulation that stress-tests the 3D model. This analysis generates a “heat map” highlighting areas of weakness or insufficient support. For instance, when designing a birdhouse, the system might identify and reinforce weak support beams. The system can then refine the design based on this feedback.
Researchers showcased PhysiOpt’s versatility by fabricating more complex designs, including a steampunk-style keyholder with intricate robotic hooks and a “giraffe table” with a flat surface for placing items. The system’s ability to interpret and generate designs based on abstract concepts like “steampunk” relies on a pre-trained model that has already been exposed to a vast library of shapes and objects. This avoids the need for extensive retraining.
“Existing systems often need lots of additional training to have a semantic understanding of what you want to see,” said Clément Jambon, co-lead author and 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 comparisons with DiffIPC, another shape simulation and optimization method, PhysiOpt demonstrated significantly faster performance. CSAIL researchers found their system was nearly ten times faster per iteration while producing more realistic designs when tasked with generating items like chairs.
The researchers, including Kenney Ng, Evan Thompson, and Mina Konaković Luković, presented their work in December at the Association for Computing Machinery’s SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia. Their work was supported by the MIT-IBM Watson AI Laboratory and Wistron Corp. Zhan and Jambon indicated future development will focus on removing occasional artifacts that appear in the 3D models and modeling more complex constraints for various fabrication techniques, such as minimizing overhanging components for 3D printing.