A small model with a visual feedback loop
At the core of the system is Bioinspired3D, a 3-billion-parameter language model fine-tuned on a curated dataset of more than 4,500 examples spanning helical, cellular, and tubular biological design motifs. Despite its compact size, the model achieves a near fourfold performance improvement over its non-finetuned base and outperforms substantially larger state-of-the-art language models on a new benchmark designed by the team to evaluate bioinspired 3D generation.
To handle complex prompts, the researchers embed Bioinspired3D in a graph-based agentic framework that pairs the language model with a vision–language critic alongside dedicated repair and refinement agents. The vision-language model evaluates each rendered geometry against the original prompt, flags physical inconsistencies such as floating components or overlapping struts, and iteratively guides the system toward designs that better match the intended structure.
"What was striking was how much capability already lives inside small open-source models," said Luu, lead author and PhD candidate in MIT's Department of Materials Science and Engineering. "By grounding generation in executable code and adding a visual feedback loop, we can get reliable 3D design behavior from a model orders of magnitude smaller than what's typically used for this kind of task."
From a biological name to a printable part
The system integrates with BioinspiredLLM, the team's earlier conversational language model for biological and bioinspired materials. Together, the two enable an end-to-end pipeline: a user provides only the name of a biological material (for example, "toucan beak" or "crab exoskeleton") and the coupled system reasons about its known microstructure, drafts a design prompt, and produces a corresponding 3D structure. In one demonstration, a prompt derived from "horse hoof wall" was correctly interpreted as a tubular architecture with gradient porosity, even though graded geometries were not explicitly represented in the training data.
Several of the AI-generated designs were exported as STL files and 3D-printed in ABS, PLA, and PETG, demonstrating direct compatibility with standard additive manufacturing workflows.
Lowering the barrier to generative scientific design
The authors argue that the broader significance of the work lies in lowering barriers to AI-driven scientific discovery. By showing that compact, modular pipelines can match or exceed much larger models on domain-specific 3D design tasks, Bioinspired123D points toward a future in which research groups without access to large GPU clusters can still meaningfully participate in generative materials design. The same agentic framework, the authors note, can be extended in future work to incorporate domain-specific critics for finite element analysis, manufacturability checks, and multiscale or hierarchical design.
The work was supported in part by Google, the MIT Generative AI Initiative, the U.S. Department of Agriculture, the National Institutes of Health, and the National Science Foundation Graduate Research Fellowship Program.
All code, datasets, benchmarks, and model weights are released openly. Code is available at github.com/lamm-mit/Bioinspired123D ; model weights are hosted on Hugging Face.
Reference: Rachel K. Luu, Markus J. Buehler. Bioinspired123D: Generative 3D Modeling System for Bioinspired Structures[J]. AI for Science. DOI: 10.1088/3050-287X/ae61d1
