🔬There Is No AlphaFold for Materials — AI for Materials Discovery with Heather Kulik
TL;DR
AI found a polymer design human chemists wouldn’t have proposed — and it made the material about 4x tougher in the lab. Kulik describes screening tens of thousands of candidates to discover a quantum-mechanical bond-breaking mechanism in a polymer network that experimentalists were surprised by and then successfully validated.
Materials still has no AlphaFold-style breakthrough because the space is messier, the bonding is more varied, and the ground truth is weaker. Unlike proteins with 20 natural amino acids and strong experimental benchmarks like CASP, materials span far more building blocks, heterogeneous bonding regimes, and often rely on low-fidelity DFT rather than large experimental datasets.
Active learning matters most when you’re optimizing across many tradeoffs, not just one property. In Kulik’s current metal-organic framework work for direct air capture, the team is juggling seven objectives at once — including cost, CO2 selectivity, humidity stability, mechanical stability, and thermal stability — where even rough models can give 100x to 1000x speedups per dimension.
LLMs are useful chemistry copilots, but they still fail on expert-level design tasks that a real chemist does instantly. Kulik’s favorite stress test is asking an LLM to design a ligand with exactly 22 atoms that binds a metal through two nitrogen atoms; despite improvements, she says she still can’t get a correct one-shot answer.
A lot of ML in materials is being trained on the wrong target: convenient data instead of experimentally meaningful data. Kulik points to repositories like Materials Project and Open Catalyst Project as useful, but says the field’s best ML engineers are often learning from low-fidelity DFT data that may not reflect experiment, making model replacement claims premature.
The real bottleneck is the interface between bits and atoms — especially process, reporting, and experimental infrastructure. Kulik argues that beyond prediction models, the field needs cloud-lab style shared facilities, machine-learning-ready reporting standards, and ways to model not just structure and properties but also the processing steps that determine whether a material actually works in a device.
The Breakdown
The 4x-tougher plastic that AI surfaced first
Kulik opens with a concrete win: her group used AI to screen tens of thousands of material candidates, where each lab experiment would have taken months to years, and found a polymer-network design that made the material roughly four times tougher. The surprise wasn’t just speed — it was chemistry the experimentalists said they never would have proposed themselves, and then it worked in the lab.
A controlled-break “fuse,” but driven by quantum mechanics
She explains the mechanism with a vivid picture: a small part of the material breaks so the whole structure survives, like a fuse preserving integrity. What her team found was not just the known idea of sacrificial bond breaking, but a new way to achieve it — electrons reorganize at the breaking point and stabilize the molecule in a way that couldn’t have been guessed from simple intuition.
From “cheminformatics” to ML, thanks to impatience and a class project
Kulik says she was drawn early to pattern-finding in data because she was impatient with doing chemistry “one molecule at a time.” Around 2015–2016, she realized it was smarter to call it machine learning than cheminformatics, and credits student John Paul Jana — now at AstraZeneca in Sweden — with turning an ideas conversation and even a homework assignment into neural-network work on inverse design.
Why active learning shines in messy, seven-objective searches
For Kulik, one of ML’s most promising roles in chemistry is handling multidimensional optimization. In current MOF work for direct air capture of CO2, the team is optimizing seven objectives at once — cost, humid stability, selectivity, mechanical durability, thermal stability, and more — and she argues active learning lets you start searching for the “needle in a haystack” before models are fully accurate.
MOFs as Legos, and the old quantum methods ML is trying to outrun
She describes metal-organic frameworks as “tinker toys” or Legos: modular building blocks that can be combined in near-infinite ways to place chemical groups precisely for host-guest interactions. Before ML, much of the field relied on quantum mechanical modeling of systems like transition-metal catalysts, where a single high-fidelity calculation could take hours, days, or weeks, and her group still uses ML partly to accelerate those calculations and even choose which quantum approximation is best.
The 22-atom ligand test: where LLM chemistry still falls over
When asked whether people can skip learning chemistry because ChatGPT “has PhD-level understanding,” Kulik gives a sharp reality check. She says LLMs are great for “Wikipedia-level chemistry knowledge,” but her standing challenge is simple: ask one to design a ligand with exactly 22 atoms and two nitrogen donor atoms for binding a metal — something an expert chemist can do immediately — and it still usually can’t do it.
Why there’s still no AlphaFold for materials
Kulik says the analogy breaks down because materials are much less standardized than proteins: not 20 building blocks, but huge chemical diversity and wildly variable bonding. She also warns that the community’s benchmark culture is weaker; repositories like Materials Project and Open Catalyst Project help, but they’re often based on lower-fidelity DFT rather than experimental ground truth, so flashy “foundation” potentials can look great until you try them on a real lab problem and “molecules fall apart.”
Mining papers, catching contradictions, and dreaming of cloud labs
Her group has spent years extracting data from papers — first with classic NLP and graph digitization, now with LLMs — and found that even basic facts like a MOF’s decomposition temperature can disagree depending on whether you read the graph or the author’s interpretation. She’d love a future with shared cloud-lab infrastructure, machine-learning-ready reporting standards, and public experimental datasets, because right now too much scientific value gets trapped in PDFs and too much of materials discovery still depends on brute-force compute that only companies like Microsoft and Meta can easily afford.