🔬Searching the Space of All Possible Materials — Prof. Max Welling, CuspAI

Episode Summary

Executive Summary: Max Welling frames AI for science as a new interface between computation and nature, where physics-inspired methods help search the space of materials and molecules. He explains why he shifted from pure theoretical curiosity toward climate and real-world impact, how symmetry, diffusion models, and stochastic thermodynamics connect physics and ML, and how Cuspai is building a human-in-the-loop platform to accelerate materials discovery and carbon-removal technologies.

Main Topics: Physics as the unifying thread across careers (Priority: 5/5): Welling describes physics as the common foundation behind his work in quantum gravity, symmetry-aware machine learning, diffusion models, and materials discovery. Why AI for science is surging (Priority: 5/5): He argues the field is exploding because AI has already delivered successes in protein folding and machine-learning force fields, while climate, health, and energy transition needs create urgent demand. Cuspai’s mission and climate motivation (Priority: 5/5): He explains that Cuspai was founded to address climate change by accelerating materials for carbon capture, water filtration, batteries, and other sustainability-critical applications. Materials discovery as computation plus experiments (Priority: 5/5): Welling introduces the idea of experiments as a 'physics processing unit'—nature doing computation—and argues future platforms must blend digital computation with automated laboratory feedback. Human-in-the-loop automation, not full replacement (Priority: 4/5): He emphasizes that domain experts remain essential; the goal is to progressively automate steps like DFT setup and workflow orchestration while keeping chemists in control of decisions. Equivariance and symmetry in ML (Priority: 4/5): He explains equivariance as building symmetry into neural networks so models generalize across rotations, permutations, and other transformations with less data. Generative AI and stochastic thermodynamics (Priority: 4/5): He highlights his upcoming book connecting diffusion models, generative AI, and non-equilibrium thermodynamics, arguing the mathematics is essentially shared and can cross-fertilize both fields.

Key Arguments: Physics provides a durable intellectual thread because symmetry, gauge structure, and non-equilibrium dynamics are deeply useful for machine learning and scientific discovery. AI for science is attractive not just because it is intellectually rich, but because it can have direct societal impact in climate, energy, health, and materials. Materials are the hidden foundation of modern computing and the energy transition; improving them is more fundamental than software-layer innovation alone. The space of possible molecules and materials can now be searched like an information retrieval problem, with AI proposing candidates and experiments validating them. Fully autonomous 'dark labs' are not the near-term vision; expert chemists and industrial partners are necessary for successful real-world deployment. Incorporating symmetry into models can reduce data needs and improve generalization, though data augmentation sometimes competes with or even outperforms hard-coded equivariance depending on optimization. Generative AI and stochastic thermodynamics share the same mathematical structure, suggesting a two-way transfer of theory and methods between physics and ML.

Data Points: Years until retirement (approx.): 10 years - Welling says he is 'closer to retirement' and wants more impact-oriented work. Cuspai age: about 20 months - He says the startup began roughly 20 months before the interview. Company size: about 40 people - Cuspai has grown to around forty employees. Total investment raised: $130 million - He says Cuspai has collected $130 million in investment so far. Largest startup funding reference: $6.2 billion - He cites a Jeff Bezos-backed AI-for-science startup’s C round as evidence of field momentum. Carbon removal timeline: another half century to a century - He argues staying within 2°C requires not only net-zero emissions by 2050 but sustained CO2 removal for decades afterward. Current solar capture efficiency: about 22% - Used as a benchmark when discussing silicon solar panels improved by perovskite layers. Theoretical solar capture efficiency: up to 50% - He mentions perovskite-on-silicon stacks could theoretically reach this level. Climate threshold: 2 degrees - He frames Cuspai’s mission around avoiding warming beyond two degrees Celsius. Potential severe warming scenario: 4 degrees - He warns four degrees would be 'very bad' and underscores urgency.

Pivotal Quotes: "I want to think of it as what I would call a sort of a physics processing unit, like a PPU... nature doing computations for you." — Max Welling: He describes his conceptual model of experiments and nature as computation. "We need to get through the energy transition fast if we don't want to kind of mess up this world." — Max Welling: He explains why materials discovery is tied to climate action and startup motivation. "In the end, the vision is it will be a search engine where somebody, a chemist, will type things and will get list candidates, but the chemist will still decide." — Max Welling: He clarifies that Cuspai’s platform is meant to empower experts, not replace them.

Implications: AI for science is becoming a practical industrial stack, not just a research trend. Expect more human-in-the-loop platforms that compress materials R&D, especially for climate and energy applications, while physics-based priors remain a key competitive advantage.