Why did Max shift toward AI for science and climate-related problems?

Max notes two drivers: first, the opportunity to apply AI tools to scientific problems (e.g., protein folding and force fields) and second, the desire to have a positive societal impact as climate change becomes more urgent. This motivates starting CusAI to accelerate material discovery and sustainable technologies. (Timestamp: 475)

What is equivariance and why is it useful in neural networks?

Equivariance means building models that respect symmetry properties so that a transformed input (like a rotated object) leads to a correspondingly transformed output. This reduces data requirements and enforces consistent behavior, which is valuable for physical problems and graph-structured data. The discussion also notes trade-offs with data augmentation. (Timestamp: 1743)

How does CusAI's platform handle generation, testing, and feedback?

The platform uses a generative component to propose candidates, then a digital twin/multifidelity pipeline to test cheap options first and escalate to more expensive experiments. It also includes autonomous agents that surface literature and orchestrate computations and experiments, with human experts providing governance and final material judgments. (Timestamp: 1116)

Will the lab ever be fully automated and dark, without human in the loop?

Max argues for a future with increasing automation, but not a fully autonomous dark lab in the near term. He envisions powerful tools that empower domain experts and keep humans in the loop to guide decisions, especially since problems are highly vertical and context-specific. (Timestamp: 1420)

When is the book Generative AI and Stoastic Thermodynamics expected to be published?

Max says the book is expected to be published around April, with a keynote at ICLR/ICLear as a potential milestone; publication timelines depend on the publisher. (Timestamp: 2023)

🔬Max Welling: Materials Underlie Everything

Latent Space Podcast

Science & Technology3 min read35 min video

Feb 25, 2026|1,913 views|47|2

Save to Pod

Key Moments

TL;DR

Physics-first AI for science: materials, PPU, and a human-in-the-loop lab

Key Insights

Physics is the unifying thread across Max Welling's work, guiding both theory (symmetries, gauge invariance) and practical ML approaches (equivariance, diffusion models).

AI for science aims to empower scientists and chemists with powerful tools, not replace them, leveraging partnerships and deep domain insight.

CUSAI and the self-driving lab concept aim to treat materials discovery as a computable process where nature’s experiments complement digital computation.

The platform blends generative models, digital twins/multiscale modeling, and autonomous agents, with data quality and human input as critical competitive advantages.

Automation is gradual and instrumental: start with workflows, automate components (DFT parameterization, testing), but keep domain experts in the loop for decisions.

A bold future dream ties generative AI to stochastic thermodynamics; the forthcoming book argues for a deep mathematical equivalence that could cross-fertilize ML and physics.

PHYSICS AS THE THREAD: FROM QUANTUM GRAVITY TO AI

Max Welling describes a through-line in his career: physics, not as a distant backdrop but as the core thread that informs how he thinks about problems in machine learning. His early work in two-dimensional quantum gravity highlighted frontier questions, yet the real undercurrent was symmetry—gauge symmetries, rotational invariance, and the role of mathematical structure. This fascination with symmetry evolved into the application of these ideas to ML, particularly in graph neural networks and equivariance, where models respect the underlying physics of the problem. More recently, he connects diffusion models and stochastic thermodynamics, showing that the mathematics of generative AI echoes that of non-equilibrium physics. This fusion—physics guiding algorithm design and vice versa—frames his view that AI should be used to illuminate deep physical questions while yielding practical, impactful tools.

A MISSION AT THE INTERFACE: IMPACT, CLIMATE, AND MATERIALS

As he looks beyond traditional academic milestones, Welling emphasizes impact as a driving force. He describes a shift toward climate-oriented technology, motivated by the urgency of reducing emissions and enabling rapid energy transition. This leads to the formation of CUSAI (CosAI in the transcript), a startup designed to accelerate materials discovery through AI-enabled experimentation. He frames the problem as one where underlying materials govern every layer of technology—LLMs, GPUs, and wafers sit atop a much deeper materials problem. Batteries, fuel cells, solar cells, and novel perovskite structures become the practical targets that could change the world, aligning scientific curiosity with societally meaningful outcomes.

THE PLATFORM ARCHITECTURE: PPU, SELF-DRIVING LABS, AND MULTISCALE TWIN

At the heart of his platform vision is the physics processing unit (PPU): nature’s own computer that complements data centers. The design envisages a loop where a generative component proposes materials candidates, a digital twin with multiscale, multifidelity models screens and ranks options, and experiments (in lab or in silico) provide feedback. Over time, autonomous agents can mine literature and orchestrate computations and experiments. Yet the emphasis remains on tools that empower scientists, not replace them. The moat, he notes, lies in data: high-quality, domain-relevant data is what makes the platform valuable and differentiable, enabling faster, more reliable discovery.

HUMAN IN THE LOOP: TOOLS, WORKFLOWS, AND INDUSTRY PARTNERSHIPS

A central theme is human-in-the-loop design. The platform starts with manual workflows, then progressively automates components such as parameter selection for DFT calculations and streamlined decision-making. Over time, agents—driven by AI, LLMs, or specialized chemists—can perform complex orchestration, but domain experts stay in charge of what counts as a “good” material. He highlights the importance of partnerships with industry to ensure that breakthroughs translate into real-world materials and processes, from water filtration to carbon capture, reinforcing that AI-for-science must be co-developed with practitioners in the field.

GENERATIVE AI AND STOCHASTIC THERMODYNAMICS: A UNIFIED THEORETICAL VISION

One of the most provocative threads is the link between generative AI and stochastic thermodynamics. Welling notes that the mathematics underpinning diffusion models, reinforcement learning, and sampling aligns with the non-equilibrium theory of systems driven away from equilibrium. He has even written a book tying free-energy concepts in ML to stochastic thermodynamics, aiming to unify these domains and reveal how insights from physics can accelerate AI methods—and how AI can, in turn, accelerate science. The broader goal is to foster cross-disciplinary fertilization that yields more powerful algorithms and deeper scientific understanding.

Mentioned in This Episode

●Books

●People Referenced

Common Questions

A PPU is described as nature doing computations that complement digital computation. The idea is to run simulations or experiments in the physical world and let those processes contribute to finding new materials, essentially pairing a data-center computation with nature's computations. It highlights a pathway to accelerate material discovery by integrating physical experimentation with AI-driven workflows. (Timestamp: 0)