🔬There Is No AlphaFold for Materials — AI for Materials Discovery with Heather Kulik

Summary

• Materials AI has no AlphaFold-like shortcut because materials involve many more building blocks, highly variable bonding, and sparse experimental ground truth. Kulik contrasts AlphaFold’s success with globular proteins, primarily using 20 natural amino acids, with materials whose current potentials are “certainly not correct across all of chemical space” and can fail more catastrophically without a clear experimental check.
• Kulik described a clear AI-enabled discovery: a polymer network made about four times tougher through a design that surprised experimentalists and worked in the lab. AI searched thousands to tens of thousands of candidates whose individual experiments could take months to years, uncovering a “fully quantum mechanical phenomenon” in which electron rearrangement stabilizes a molecular component as it breaks.
• Active learning is especially valuable when materials must satisfy many simultaneous constraints. Kulik’s direct-air-capture campaign optimizes seven objectives—including cost, humidity stability, CO2 selectivity, and mechanical and thermal stability—with even imperfect models offering “at least a hundred- to a thousandfold speedup for every dimension.”
• Claims that neural potentials have already displaced physics-based simulation remain ahead of demonstrated performance. One unnamed model that made a major splash was only about five times faster than Kulik’s fastest GPU DFT calculation and “doesn’t work all the time.” Her transformative threshold would be a reliable replacement for DFT at roughly two orders of magnitude greater speed.
• General-purpose LLMs can augment chemistry knowledge, but they still require an expert error detector. ChatGPT is “super good at Wikipedia-level chemistry knowledge,” yet repeatedly fails Kulik’s simple request for a ligand containing exactly 22 atoms and binding through two nitrogen atoms. The operating rule is to “learn chemistry well enough to know when these models are right or wrong.”
• Experimental data, validation, and manufacturing process are major bottlenecks alongside model scale. Literature-derived labels conflict depending on whether they come from a graph or an author’s interpretation, autonomous labs struggle with experiments humans find easy, and materials performance at device scale depends on processing—an area where Kulik says, “We’re at ground zero. We’re nowhere.”
• Compute-rich companies change how academics should choose problems. Kulik contrasts academic resources with Microsoft and Meta’s “basically infinite resources,” while pointing to neglected chemistry, better evidence, creative problem selection, shared cloud labs, and machine-readable experimental reporting as opportunities.

Deep dive

1. AI found a polymer design chemists did not anticipate

• Kulik described a clear proof point from screening thousands to tens of thousands of materials whose individual laboratory experiments could require months to years. The selected polymer-network design was unexpected to experimental collaborators, yet testing confirmed that it made the material about four times tougher.

• The design uses molecular components that break apart in a way that makes the overall structure tougher. AI found a “fully quantum mechanical phenomenon” in which electrons rearrange to stabilize the molecule precisely as it begins breaking—an effect resembling catalytic or enzymatic chemistry but not previously demonstrated in these polymers.

• A host compared the mechanism to a controlled fuse preserving the Bay Bridge during an earthquake. Kulik’s correction matters: sacrificial breaking itself had appeared in Science a couple of years earlier; her group’s contribution was the specific, electronically enabled design that produced it.

2. Active learning earns its keep on seven-way trade-offs

• Kulik entered data-driven chemistry through impatience with studying “one molecule at a time” in the mid-2000s. Around 2015–2016, she stopped calling the work cheminformatics and started calling it machine learning; student Jean-Paul Janet turned an initial design discussion into neural-network work for a class assignment.

• The polymer search was active learning “in principle,” although the team stopped after one generation because it had exhausted the space. The larger promise is beginning optimization before models are accurate, particularly when the target is a multidimensional “needle in a haystack.”

• Her current metal-organic-framework campaign for direct CO2 capture balances seven objectives: cost, stability in humid or aqueous conditions, CO2 selectivity, mechanical stability under force, thermal stability, and other constraints. Kulik says even a not-very-accurate model can provide “at least a hundred- to a thousandfold speedup for every dimension” being optimized.

• MOFs are modular “Tinker Toys or Legos” used in gas storage, sensing, separations, polymer composites, CO2 capture, catalysis, and even drug delivery. Their precise chemical groups can form a molecular “glove” around a targeted guest, but stability remains a central limitation.

3. Chemistry expertise remains the error detector

• Transition metals have unpaired electrons that make them more reactive and useful in catalysis, including processes such as Haber–Bosch ammonia synthesis. Their behavior is traditionally modeled through approximations to the Schrödinger equation, but one quantum-mechanical prediction can require hours, days, or weeks.

• Machine learning can accelerate those calculations and help select the appropriate approximation, a choice too complex for simple heuristics. Kulik’s group feeds information from the quantum-mechanical wave function into neural networks to learn which method a particular material requires.

• Shawn Wang’s provocation—why learn chemistry if ChatGPT has PhD-level understanding?—elicited Kulik’s 22-atom ligand test. LLMs repeatedly miss the exact atom count even after being told the ligand must bind through two nitrogen atoms, something an expert chemist “could do in a second.” They are useful tutors and knowledge augmenters, but not starting points for blind reliance.

4. Materials has neither AlphaFold’s bounded space nor its ground truth

• The less-served ML opportunities involve scarce or diverse data: reaction prediction, transition-metal bonding, warm dense matter, and excited states produced by shining light on materials. By contrast, “really boring chemistry”—organic molecules and protein binding—already has familiar datasets, benchmarks, and leaderboards.

• Materials Project and Open Catalyst Project offer leaderboards built largely on relatively low-fidelity DFT calculations, not experimental truth. Unlike CASP, they do not supply comparable experimental ground truth, so “all the smartest ML engineers” are learning on data that may not reflect laboratory behavior.

• Kulik says each new “foundation potential” can look excellent until her group applies it to a real problem and molecules simply fall apart. One unnamed model that made a huge splash that summer was only about five times faster than her fastest GPU DFT calculation and did not work consistently; reliable replacement of DFT at two orders of magnitude greater speed would genuinely change the science.

• Shawn Wang characterized AlphaFold as solving ground-state structures; Kulik emphasized that it handles globular proteins, primarily using 20 natural amino acids, whereas materials range from comparatively easy aluminum to iron oxide, metal-organic bonding, and high-entropy alloys. At larger length and time scales, missing or interpretive experiments leave “no real way to know if you’re right or wrong.” Kulik also noted that AlphaFold has failures, while materials models could fail more catastrophically.

5. The bottleneck shifts from computation to evidence and process

• Autonomous laboratories do some experiments that humans find difficult, while struggling with others humans find easy; researchers are also exploring how to introduce the serendipity or noise a human might experience. Beyond the material itself, people working to get materials to device scale say “it’s not just the material, it’s the process,” yet Kulik says machine learning has barely addressed processing effects.

• Literature extraction creates a quieter data-quality problem. Kulik’s group can derive a MOF’s breakdown temperature from its experimental graph or from the authors’ written interpretation, but “those two things do not line up”; modern LLM extraction remains sensitive to false positives, making verification an overhead.

• Kulik describes an open question: whether a model trained on a field’s first 30 years can predict discoveries from the next 20. Uncertainty quantification might identify the most interesting materials to add, while shared cloud labs and standardized, machine-learning-ready reporting could make experimental results reusable from publication day.

• Kulik contrasts academic resources with Microsoft and Meta’s “basically infinite resources” and now avoids projects that could be solved mainly by throwing compute at them. Her group makes transition-metal and MOF design tools available through molSimplify and MOFSimplify—websites, Conda, and GitHub—with machine-learning predictions and novel-structure generation—and seeks feedback from users, including companies already using them.