🔬Max Welling: Materials Underlie Everything

From quantum gravity to equivariant networks

Welling applies mathematical tools from particle physics and general relativity—specifically gauge symmetries and group theory—to machine learning, driving foundational work on variational autoencoders, graph neural networks, and equivariant architectures with collaborators like Taco Cohen.

Diffusion models mirror non-equilibrium thermodynamics

His upcoming book details the deep mathematical equivalence between generative AI techniques (diffusion models, MCMC, Schrödinger bridges) and stochastic thermodynamics, the physics of systems far from equilibrium.

Gerard 't Hooft's challenge to quantum mechanics

Inspired by his PhD advisor Nobel laureate Gerard 't Hooft—who now argues quantum mechanics is incomplete—Welling pursues "understandable" physical theories that avoid multiverse interpretations while maintaining mathematical rigor.

Investment volumes reaching billions

The field has exploded from hundreds of millions to multi-billion-dollar funding rounds, including a recent $6.2 billion investment in an AI science startup backed by Jeff Bezos, signaling recognition that this is a "virgin field" for innovation.

Escaping advertising algorithms for climate impact

Researchers are redirecting AI tools from advertising and multimedia toward high-impact scientific domains where symmetry-aware architectures have already succeeded, such as AlphaFold for protein folding and machine learning force fields for materials.

A new interdisciplinary discipline

AI for science represents an emerging academic and industrial field at the interface of machine learning and natural sciences, requiring new curricula and educational resources to train engineers in scientific domains.

The material foundation of digital technology

Even software like LLMs ultimately depend on materials innovation—for example, GPUs rely on novel materials deposited on wafers using EUV lithography, and further computing advances require solving fundamental materials bottlenecks.

Molecular space as a search engine

Unlike traditional hypothesis-driven science, new AI tools can search the entire space of possible molecules (not just known compounds) to discover materials for batteries, perovskite solar panels reaching 50% theoretical efficiency, and self-destructing plastics.

Carbon capture as an unsolved imperative

Staying within 2°C of warming requires not just emissions reduction but a century of carbon dioxide removal at massive scale, a problem that demands new material discovery for capture technologies that do not yet exist.

CuspAI's $130M automation bet

Founded 20 months ago with co-founder Chad Edwards, CuspAI has grown to 40 people and aims to fully automate materials discovery by integrating computational modeling with high-throughput experimentation to screen novel carbon capture materials.

Nature as a computational backend

Welling conceptualizes laboratory experiments as a "Physics Processing Unit" (PPU)—nature performing computations that are physically impossible to simulate digitally—which must seamlessly integrate with data center processing to discover new materials.

Closing the loop between simulation and reality

The platform automates the iterative cycle where AI suggests molecules, robotic experiments validate properties, and results feed back into the model, creating a search engine where researchers type desired properties and receive synthesized candidates.