AI Model Designs 3D Printed Bone Scaffolds in 12 Seconds

Researchers reported in PLOS One that ScaffoldForge, a diffusion-based generative AI model, designs a complete 3D printed bone scaffold geometry in about 12 seconds versus roughly 8,400 seconds for conventional topology optimization — about 700 times faster on a single NVIDIA A100 GPU. ScaffoldForge reached an estimated compressive strength of 56.6 MPa, 8.2% higher than topology optimization and 6.6% higher than the concurrent TriTopo-LGDM method, with an average gain of 3.5 MPa over TriTopo-LGDM across five training runs (p = 0.043). All results came from computer simulations on a training set of 250 scaffolds (200 software-generated, 50 real), and no scaffold was printed or lab-tested.

Researchers have developed ScaffoldForge, a generative artificial intelligence AI model that designs 3D printed bone scaffolds from a set of target structural properties. Described in a paper published in PLOS One https://journals.plos.org/plosone/ , the model produces a complete scaffold geometry in about 12 seconds. Topology optimization TO , a conventional computational design method, takes about 8,400 seconds per design. All reported results come from computer simulations. No scaffold was printed or tested in a laboratory. Large bone defects that cannot heal on their own affect more than 2 million patients each year, according to the researchers. Grafts harvested from the patient’s own bone remain the standard treatment, but supply is limited and harvesting can cause complications at the donor site. 3D printed scaffolds can be shaped to fit an individual patient. Designing one still means balancing requirements that pull in opposite directions, since a more porous structure helps blood vessels grow into the implant but also weakens it. The study proposes generating scaffold designs directly from a list of specifications in place of iterative optimization. According to the team, this could eventually let clinicians and engineers explore patient-specific designs interactively instead of waiting on lengthy optimization runs. How the Model Builds a Scaffold ScaffoldForge takes eight standardized measurements of trabecular bone architecture as its input. These cover solid volume, strut thickness and spacing, connectivity, rod or plate shape, directional alignment, and how easily fluid passes through the structure. Starting from random noise, the diffusion model refines the volume over 1,000 steps into a 3D voxel grid measuring 64 units per side that matches the requested values. During training, the model is penalized for three types of failure: - compressive strength below the level needed for load-bearing use - overhanging features that would require support structures during additive manufacturing AM - permeability too low for nutrients to reach cells The researchers describe these penalties as empirical rules, not physical equations. Strength and permeability are estimated by two pre-trained neural networks. Matched geometry and property data for real scaffolds is scarce, so the team assembled a training set of 250 scaffolds: 200 generated by software and 50 real scaffolds with experimentally measured mechanical data. Simulated Results Against Existing Design Methods The team compared ScaffoldForge with eight other approaches, including TO, evolutionary algorithms, Bayesian optimization, and earlier generative models. ScaffoldForge reached an estimated compressive strength of 56.6 MPa. That is 8.2% higher than TO and 6.6% higher than TriTopo-LGDM, a concurrent diffusion method that starts from TO designs. Across five training runs, the average gain over TriTopo-LGDM was 3.5 MPa p = 0.043 . TO produced scaffolds with slightly higher permeability. The authors attribute this to ScaffoldForge favoring stiffness while keeping permeability above the level associated with blood vessel growth. On a single NVIDIA https://www.nvidia.com/ A100 GPU, ScaffoldForge generated a design roughly 700 times faster than TO, while TriTopo-LGDM needed 340 seconds. Without the penalty terms, generated structures showed disconnected clusters and floating material that could not be fabricated. A comparison of 500 generated designs with the training set indicated that the model produces new geometries and does not copy training samples. The researchers also demonstrated a pipeline that converts patient CT scans into scaffolds. It identifies the defect, generates a matching scaffold, and trims it to the defect boundary. Across 20 patient cases from a public spine CT dataset, the fit error stayed within the 1 mm tolerance used for implants. Limits of a Small, Mostly Synthetic Dataset Accuracy declined on bone shapes the model had not seen during training. On vertebral anatomy from an external dataset, the error in matching target measurements was 65% higher than on the main test set. Expanding the training set to 2,050 scaffolds barely changed accuracy on familiar geometries but cut the error on unfamiliar anatomy by about 27%. The team takes this as evidence that dataset size, not model design, is the main constraint on performance with new patients. The software-generated training scaffolds matched real bone within half a standard deviation on all seven shape measurements, but only two of seven met a stricter threshold. Settling that comparison would require around 100 real micro-CT scans, double the number available. A neural network classifier could still distinguish generated scaffolds from real bone, though less easily than it could the synthetic training data. That difference was no longer statistically significant with the enlarged training set. The authors state that closing the gap will require more real bone scans in training, not more synthetic data. Further limitations include the grid resolution, which captures features down to roughly 0.5 mm, coarser than the 50 µm scale of natural trabecular bone. Regulatory assessment was also outside the study’s scope. The researchers note that the speed advantage applies to the design stage alone, not to fabrication, testing, or approval. The code, data generation pipeline, and trained models are available on GitHub https://github.com/vectorlab5/scaffold-diffusion-bone . Building Physical Limits Into AI-Generated Designs Checking whether a candidate scaffold will hold up and print typically requires a simulation, and iterative methods such as TO repeat that step many times per design. ScaffoldForge applies those checks during training, so the finished model produces geometry already steered toward them, with one simulation to validate each result. Researchers at the Massachusetts Institute of Technology https://www.mit.edu/ MIT , with collaborators from Google https://www.google.com/ , Stability AI https://stability.ai/ , and Northeastern University https://www.northeastern.edu/ , developed MechStyle https://3dprintingindustry.com/news/mit-unveils-mechstyle-a-generative-ai-system-for-structurally-sound-3d-design-248329/ , a generative AI system that restyles 3D models while tracking structural stability. With earlier stylization approaches, around 26% of modified models stayed structurally sound. MechStyle runs finite element analysis when edits threaten weak regions and reached up to 100% structural viability across 30 test models. Like ScaffoldForge, it aims to keep AI-generated geometry mechanically sound at the design stage. It differs in editing existing printable models for style and cannot repair flawed geometries, while ScaffoldForge generates new scaffold geometry from numeric targets. At Politecnico di Torino https://www.polito.it/en and Maastricht University https://www.maastrichtuniversity.nl/ , researchers built a Voronoi-based scaffold design tool https://3dprintingindustry.com/news/researchers-develop-voronoi-based-scaffold-design-tool-for-extrusion-3d-printing-of-lung-tissue-models-251045/ for lung tissue models. It generates printing toolpaths from parameters such as seed number, pattern diameter, and fiber thickness, and predicts properties including porosity and pore area. Printed scaffolds came within about 2.2% of predicted porosity for fused deposition modeling and 1.7% for melt electrowriting. The tool shares ScaffoldForge’s aim of linking design inputs to expected properties before printing, but it targets a different tissue and has already checked its predictions against printed parts, a comparison ScaffoldForge has yet to make. The authors identify fabrication and mechanical testing of selected designs against their simulated predictions as the immediate next step. Titled “ ScaffoldForge: Bioinspired generative design of 3D-printed bone scaffolds via conditional diffusion for large segmental bone defect reconstruction https://doi.org/10.1371/journal.pone.0361041 ,” the study was conducted by Yaohua Zhang, Hongyue Zhang, and Huaqing Yang. To stay up to date with the latest 3D printing news, don’t forget to subscribe to the 3D Printing Industry newsletter or follow us on LinkedIn . Explore the full Future of 3D Printing and Executive Survey series from 3D Printing Industry, featuring perspectives from CEOs, engineers, and industry leaders on the industrialization of additive manufacturing , 3D printing industry trends 2026 , qualification, supply chains, and additive manufacturing industry analysis . Featured image shows Overview of the ScaffoldForge AI framework for generating 3D printed bone scaffolds. Image via Zhang et al.