Researchers at Beijing Rehabilitation Hospital, Capital Medical University, have developed ScaffoldForge, a generative artificial intelligence model that produces a bone scaffold design in about 12 seconds on a single graphics processing unit.

Topology optimization, the conventional method, took 8,400 seconds (140 minutes) per design, making the new approach roughly 700 times faster. TriTopo-LGDM, the strongest baseline tested, took 340 seconds. All results were computational, and no scaffolds were printed or mechanically tested.

ScaffoldForge is a 3D conditional diffusion model. It takes eight trabecular bone descriptors, including bone volume fraction, trabecular thickness and spacing, connectivity and permeability, and returns a scaffold architecture on a 64-voxel cubic grid. During training, a physics-constrained loss penalized low predicted compressive strength, overhangs that would complicate printing and low permeability. The authors described these penalties as empirical soft constraints rather than governing physical equations.

Strength estimates and realism

ScaffoldForge designs reached an estimated compressive strength of 56.6 megapascals (MPa), 8.2% above TO and 6.6% above TriTopo-LGDM. Across five seeds, the 3.5 MPa gain over TriTopo-LGDM was statistically significant (p = 0.043). TO retained slightly higher permeability.

Generated scaffolds matched real bone within 0.5 standard deviations on all seven shape descriptors tested, but only two met a stricter 0.3 standard deviation margin, a gap the authors tied to the small pool of 50 real volumes. 

Expanding the training set to 2,050 scaffolds reduced descriptor error on unseen anatomy from 0.079 to 0.058. The study also outlined a pipeline that segmented defects from patient CT scans and trimmed generated scaffolds to fit them. The authors left physical, biological, manufacturing and regulatory validation to future work.

The work, by Yaohua Zhang, Hongyue Zhang and Huaqing Yang, appeared in the journal PLOS ONE.