Sep 19, 2024 · 38m · catalyst
Can AI revolutionize materials discovery?
Doğuş Çubuk (Doj Chubuk) · 20m spoken Shayle Kann · 10m spoken Sponsor Voiceover (Bloom Energy & Engie) · 2m spoken Sponsor Voiceover (EnergyHub) · 1m spoken
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Catalyst, host Shayle Kann speaks with Google DeepMind research scientist Doğuş Çubuk to evaluate whether artificial intelligence can overcome data scarcity and physical constraints to revolutionize materials discovery for climate technologies.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Shayle holds 29.9% of the talking time here. How this is scored →
Shayle as informed peer 4.8 Guest teaching 5.5 Guest disagreement 0.1 Shayle pushing back 1.3
speaking balance: gold is Shayle, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 17:25 Sober assessment of simulation breakthroughs
Çubuk delivers a candid and slightly contrarian reality check when asked what AI has proven so far, stating 'not a whole lot' and widening the critique to decades of simulation work.
Hardest push from Shayle ▶ 15:47 Probing room-temperature superconductor discoveryKann pushes past general enthusiasm to delineate strict boundaries, challenging whether computational ML is merely optimizing existing high-temperature recipes rather than reaching room-temperature breakthroughs.
Biggest teaching moment ▶ 22:14 The severe experimental data bottleneckÇubuk educates Kann on the stark quantitative reality of materials data, revealing that out of 200,000 known crystal structures, only 1,000 to 2,000 have well-characterized physical properties.
Shayle holds their own ▶ 27:42 Explaining system integration barriers in battery materialsKann demonstrates strong domain expertise in climate tech venture by explaining why novel battery materials fail without complex multi-component interface compatibility, validating why startups focus on standalone MOFs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Shayle as informed peer | Guest teaching | Guest disagreement | Shayle pushing back | Why |
|---|---|---|---|---|---|---|
| Historical Serendipity in Materials Discovery | 4 | 5 | 0 | 1 | Kann asks for a historical primer on pre-AI materials discovery, prompting Çubuk to share detailed historical examples of accidental breakthroughs like ductile tungsten, Li-ion batteries at Exxon, and Bardeen's transistor trials. | |
| The Paradox of Optimization Versus Radical Innovation | 3 | 6 | 0 | 0 | Çubuk outlines the core paradox in both materials science and machine learning: high domain expertise leads to incremental optimization rather than out-of-distribution radical innovation. | |
| Computational Simulation and Superconductor Optimization | 6 | 5 | 0 | 2 | Kann synthesizes Çubuk's explanation of cuprate superconductor optimization to infer that computational ML is well suited for optimizing existing high-temperature recipes but unlikely to discover room-temperature superconductors. | |
| Assessing Real-World Computational Materials Breakthroughs | 4 | 6 | 1 | 2 | Kann presses for concrete commercial achievements from AI materials discovery, and Çubuk gives a candid reality check that simulations have yielded very few commercial products over several decades. | |
| Data Bottlenecks and the Continuing Need for Lab Work | 6 | 6 | 0 | 2 | Çubuk explains the massive training data gap between internet-scale LLMs and crystal databases. Kann astutely counters with a chicken-and-egg argument: replacing lab work may ironically require a massive surge in upfront experimental lab data generation. | |
| Why Climate Tech Startups Focus on MOFs for Carbon Capture | 7 | 4 | 0 | 1 | Kann highlights the pattern of climate tech startups focusing on MOFs for carbon capture. Both discuss system integration barriers, with Kann detailing why standalone materials bypass complex battery interface engineering. | |
| Evaluating Applications: DFT Capabilities and Limitations | 4 | 6 | 0 | 1 | Kann asks Çubuk to define density functional theory (DFT) and evaluate where it works best. Çubuk details how DFT handles bulk stability well for batteries but struggles with electronic band gaps and messy catalytic surfaces. | |
| The Search for an 'AlphaFold Moment' in Materials Science | 4 | 6 | 0 | 1 | Kann queries whether materials science will have an AlphaFold moment. Çubuk explains why that is difficult due to experimental noise exceeding computational error benchmarks, then reviews DeepMind's GNoME roadmap. |
Statements from this episode (12)
Kann: AI for gigaton-scale emission reductions is harder than simple efficiency
“Of course, the world abounds with ways to use AI to do things more efficiently, but to really move the needle on like gigatons of emissions, I think is trickier.”
Çubuk: Deep expertise optimizing existing materials hinders discovering radically different ones
“The better you know a system,
The more you can continue optimizing the system, but it doesn't necessarily mean that knowledge will help you discover something different.
And I think this is probably why a lot of important discoveries are serendipitous, because…”
Çubuk: Optimization of decades-old materials impedes commercialization of new discoveries
“If you think about plastics, we're still mostly using things that we discovered 70 years ago, 80 years ago, and we've gotten so good at, you know, manufacturing them, optimizing them, so now
For someone to come up and say, oh, I discovered a completely differe…”
Çubuk: AI models excel at textbook science but cannot generate paradigm shifts
“AI hasn't really done this yet. Even today's best AI models seem to be really good at kind of doing the textbook stuff, you know, like high school, college, but then when you think about being more creative and, you know, trying to shift the paradigm, it's bee…”
Çubuk: Physicists still cannot explain why high-temperature superconductivity occurs
“And we still don't know as physicists where high temperature superconductivity comes from. It's like a crazy thing. You know, it's been around for 50 years 40 years. We don't know why it happens.”
Çubuk: Duracell batteries use a cathode material discovered via computational simulations
“There's one example that often gets talked about. I think one of the cathode materials maybe from Sader group and materials project, I think is in Duracell batteries.”
Çubuk: AI simulations will never completely eliminate physical lab work
“So I think, I can't imagine a future where we completely eliminate lab work. Because, first of all, we don't know if quantum mechanical simulations will ever become good enough to correctly predict experiments, you know, all the time.”
Kann: Battery material discovery requires co-optimizing complex material interactions
“In battery world, nothing exists in a vacuum. So you can't create a novel material and then be done with it. You have to figure out not only the material, but then its interaction with all the other materials, which are also in flux. And that's part of what ma…”
Çubuk: Physical materials experiments are as noisy as computational simulation errors
“In material science, I think one of the issues is the experimental data is actually quite noisy. So, you know, this is something that you might hear often that simulations and DFT isn't very accurate, and that's true, but maybe one thing that People don't noti…”
Çubuk: Materials science lacks a CASP-like benchmark database for machine learning
“And I think now, maybe because now machine learning really needs this high precision, large data set, there are these bigger efforts trying to create a CAASPP-like database, but it's not there yet.”
Çubuk: Deep learning scaling laws apply to quantum mechanics and materials
“The more training data you put into LLMs, the better results you get, and how much better your results are actually predictable. It's kind of like a power law. This comes back from, you know, a paper from Baidu Research from back in 2016, I think, and it seems…”
Çubuk: DeepMind’s next materials goal is predicting finite temperature stability
“One of our next goals is to predict not just zero Kelvin stability, but finite temperature stability, and this is much harder because at finite temperature, there's entropy effects, and we're interested in, you know, finding not just materials that are stable,…”