DiffractGPT, an open artificial intelligence (AI) model for materials design, has been downloaded roughly 240,000 times on Hugging Face “more than DeepSeek”, told its creator to the room during our recent 3DHEALS virtual event Biomaterials Frontier for Medical 3DPrinting. That creator is Kamal Choudhary, a materials scientist at Johns Hopkins University and a research associate at the National Institute of Standards and Technology (NIST). In fifteen minutes, he laid out what AI can already do in materials discovery: JARVIS, the open materials database he built at NIST; ALIGNN, a graph neural network that predicts material properties; and AtomGPT, a “ChatGPT for materials”. That said, almost none of these can directly benefit bioinks, implantable biomaterials, or medical 3D printing community. This article explains what AI-driven materials design can do today, why biology is the bottleneck, and where the commercial opportunity sits for bioprinting and medical device companies.
What AI can already do
Start with the model behind that download count. DiffractGPT reverses the usual workflow: give it an X-ray diffraction (XRD) pattern and it proposes the atomic structure that would produce it, including for compounds that do not yet exist in any database. Choudhary said at our recent event that it has been downloaded around 240,000 times on Hugging Face, “more than DeepSeek”, he noted. This is a remarkable reach for a tool built for crystallographers. And AtomGPT, his “ChatGPT for materials,” now connects to general chatbots, including OpenAI and Claude, through the Model Context Protocol (MCP), so a language model answers a materials question by querying a physics-grounded tool instead of inventing a plausible-looking wrong one. In a field where a confidently wrong number can send a lab down a six-month dead end, that plumbing matters more than it sounds.
Behind the generative models sits the data, and the scale is easy to underestimate. Choudhary built JARVIS (stands for: Joint Automated Repository for Various Integrated Simulations), an open repository of materials data while at NIST; he says it has around 200,000 users and roughly 100,000 materials. On top of that sits a family of models. ALIGNN, the Atomistic Line Graph Neural Network, was the one to pay attention to. Earlier graph models described a crystal as atoms joined by bonds. ALIGNN added the angles between those bonds, and that single change improved accuracy sharply, by his account up to 44 percent, while staying fast. Its reported error for formation energy is about 0.022 electron-volts per atom, close enough to first-principles calculations to work as a screening tool. His group used it to computationally pick superconductors and a metal-organic framework (MOF) for carbon capture, then synthesized and validated them. Prediction, then a real material. If you want the map rather than the demo, his group’s 2022 review in npj Computational Materials, now cited more than 1,200 times, is the standard reference.
Why biology is still outside
During the event, the moderator Craig Rosenblum asked if the same AI tool can be useful to characterize a material’s “inorganic properties like biocompatibility, the degradation rate, the bioactivity”.
“Great question,” answered Choudhary.
The lack of domain-specific data in biology is where the gap is.
The data problem is lopsided. There is a great deal of it for formation energy, bulk modulus, and band structure, the physical properties of inorganic, crystalline materials. There is very little for the properties that decide whether something belongs in a body, the behavior of a soft hydrogel or an absorbable polymer as it dissolves.
AI has gone deep in certain verticals but left the biological slice mostly untouched.
The loop problem is worse and more structural. Most of these models learn from simulation. To trust them, you check them against experiments, and the experiments are the bottleneck. Characterization tools like X-ray diffraction are slow. They generalize poorly to messy, defect-laden, biological systems. So, the theory-to-experiment loop that would let a model propose a bioink and then learn from how it performed does not close easily. Choudhary was candid that his cleanest results were for near-perfect systems, not for the heterogeneous reality of biomaterials. That said, Choudhary’s own showcase, the AI for Materials Design Laboratory at Hopkins, is a genuine closed-loop facility: robots move samples past automated X-ray, indentation, and laser-impact stations while AI decides what to test next. It is also built for materials in extreme environments, like defense and aerospace, not for biological systems. The machine that would do the same for biomaterials, feeding biological outcomes back into the model, mostly does not exist yet. (See our previous event focusing on AI and 3D printing, where NUS researchers have successfully created a ML-feedback loop to generate bioprinted gum tissue.)
Where the money is
That absence is the opportunity, and Choudhary said.
The winners in AI-driven materials, he argued, will not be whoever trains the biggest general model. They will be whoever owns a proprietary dataset in a specific domain, or builds the narrow, expert model for it. His half-joking example, a “collagen GPT,” is exactly the shape of the bet, knowing one of the speakers is an expert in human collagen. Sell the model no one else can train or sell the data no one else has.
For the 3DHEALS audience, that could be the strategy.
Bioink and biomaterial companies already sit on the scarce asset: batch after batch of biological performance data that never leaves their benches. Instrument the printers. Label the outcomes. Build the loop between what you printed and how the tissue responded. The company that does this for its niche, whether degradable polymers, bioinks, or regenerative materials, will own a moat that a foundation model cannot cross, because the model has never seen the data.
The AI, increasingly, is the easy part. The wet-lab evidence is the hard (and expensive) part, and therefore the valuable one.
References
• Biomaterials Frontier (3DHEALS event): https://3dheals.com/biomaterials-frontier/
• Choudhary et al., “Recent advances and applications of deep learning methods in materials science,” npj Computational Materials (2022): https://www.nature.com/articles/s41524-022-00734-6
• AI for Materials Design Laboratory (AIMD-L), JHU HEMI / CAIMEE: https://hemi.jhu.edu/caimee/center-facilities/aimd-l/
•AtomGPT: https://atomgpt.org • Choudhary demos (YouTube): https://www.youtube.com/@dr_k_choudhary
•Artificial Intelligence Updates For 3D Printing and Bioprinting (on-demand course)
Organizations
• Hopkins Extreme Materials Institute (HEMI): https://hemi.jhu.edu/
• Center for Artificial Intelligence and Materials Engineering (CAIMEE), Johns Hopkins: https://hemi.jhu.edu/caimee/
• National Institute of Standards and Technology (NIST): https://www.nist.gov/
• NIST JARVIS (Joint Automated Repository for Various Integrated Simulations): https://jarvis.nist.gov/
Comments