Rosa Del Mar

Daily Brief

Issue 56 2026-02-25

Closed Loop Materials Discovery Compute Plus Experiment

Issue 56 Edition 2026-02-25 8 min read
General
Sources: 1 • Confidence: Medium • Updated: 2026-03-02 20:00

Key takeaways

  • Max Welling proposes treating physical experiments as a "physics processing unit" (nature-as-compute) that should be integrated with data-center computation in a materials-discovery workflow.
  • Max Welling states CuspAI was started about 20 months prior to the interview to develop technology for carbon dioxide removal motivated by climate-change concerns.
  • Max Welling describes equivariance as hard-coding symmetry constraints (such as rotations, translations, permutations) into neural network weights to improve generalization across transformed inputs with less data.
  • Max Welling describes AI-for-science (and science-for-AI) as an emerging discipline focused on the interface between physics/science and machine learning.
  • Max Welling states his forthcoming technical book argues diffusion models and generative AI share identical mathematics with stochastic non-equilibrium thermodynamics and aims to enable cross-fertilization of methods between the fields.

Sections

Closed Loop Materials Discovery Compute Plus Experiment

  • Max Welling proposes treating physical experiments as a "physics processing unit" (nature-as-compute) that should be integrated with data-center computation in a materials-discovery workflow.
  • Max Welling describes CuspAI's platform as using a generative model to propose candidates and a multi-scale, multi-fidelity digital twin to filter candidates from cheap to expensive evaluations before experimental validation.
  • Max Welling states CuspAI's automation approach starts with humans manually assembling modular tool workflows and then progressively automates components and sequencing using agents or optimizers.
  • Max Welling states CuspAI has added agents that search chemical literature, propose experimental suggestions, and orchestrate computations and experiments.
  • Max Welling states an expectation that domain experts will remain in the loop for a long time, with near-term goals focused on empowerment and acceleration rather than fully autonomous labs.
  • Max Welling states an expectation that materials R&D can increasingly be approached as an automated search over the space of molecules/materials rather than a slow hypothesis-experiment loop.

Cuspai Company Scope And Commercialization Gates

  • Max Welling states CuspAI was started about 20 months prior to the interview to develop technology for carbon dioxide removal motivated by climate-change concerns.
  • Max Welling states CuspAI has grown to about 40 people and raised about 130 million in investment.
  • Max Welling states CuspAI commits to new material-development directions only when it has a strong industrial partner to collaborate through deployment with domain expertise.
  • Max Welling states he helped start CuspAI to work on climate change through technology while engaging with deep scientific problems in materials science.
  • Max Welling states CuspAI is pursuing a mixed strategy consisting of a confidential long-term lighthouse material proof point alongside shorter paid projects such as training force fields for clients.
  • Max Welling states CuspAI is working with Kemira on a water filtration project aimed at removing PFAS.

Physics Inspired Ml Inductive Bias And Training Tradeoffs

  • Max Welling describes equivariance as hard-coding symmetry constraints (such as rotations, translations, permutations) into neural network weights to improve generalization across transformed inputs with less data.
  • Max Welling states data augmentation can sometimes outperform hard-coded equivariance because constraints can make optimization harder and impede finding good minima.
  • Max Welling states inductive bias trades off against data, and if the imposed bias is not exactly correct it can cap achievable performance.
  • Max Welling states physics is the central thread linking his work and emphasizes the importance of symmetry concepts from theoretical physics to machine learning.
  • Max Welling states his symmetry-to-ML work with Taco Cohen progressed from rotational symmetries to gauge symmetries on spheres and related domains.

Ai For Science Field Dynamics Capitalization And Bubble Risk

  • Max Welling describes AI-for-science (and science-for-AI) as an emerging discipline focused on the interface between physics/science and machine learning.
  • Max Welling attributes AI-for-science momentum to successful applications (including protein folding and machine-learning interatomic potentials) and to a belief that the field has reached a capability inflection point.
  • Max Welling states there is a Jeff Bezos-backed AI-for-science startup that raised a 6.2 billion seed round.
  • Max Welling states AI-for-science investment and activity are rapidly accelerating and may be forming a bubble.

Diffusion Models Linked To Non Equilibrium Thermodynamics

  • Max Welling states his forthcoming technical book argues diffusion models and generative AI share identical mathematics with stochastic non-equilibrium thermodynamics and aims to enable cross-fertilization of methods between the fields.
  • Max Welling claims diffusion models, reinforcement learning, Schrödinger bridges, and MCMC share mathematical structure with stochastic thermodynamics for non-equilibrium systems.
  • Max Welling states he taught a course at the African Institute for Mathematical Sciences (near Cape Town) and turned it into a book that has been sent to a publisher.
  • Max Welling states his upcoming book focuses on relationships among free energy, diffusion models (generative AI), and stochastic thermodynamics.

Watchlist

  • Max Welling describes AI-for-science (and science-for-AI) as an emerging discipline focused on the interface between physics/science and machine learning.

Unknowns

  • What are the measured iteration-time and throughput gains (end-to-end) from integrating experiments as a closed-loop compute component versus a conventional simulation-then-validation pipeline?
  • What quantitative screening efficiency metrics exist for the multi-fidelity digital twin (e.g., hit rate at each fidelity tier, calibration error, compute cost per surviving candidate)?
  • How mature and reliable are the agents used for literature search and experiment/computation orchestration (error rates, reproducibility outcomes, required human oversight)?
  • What is the identity and verified financing structure of the asserted Jeff Bezos-backed AI-for-science startup with a 6.2 billion seed round?
  • What are the concrete commercialization milestones and performance targets for CuspAI’s CO2 removal focus (capture performance, durability, operating conditions, and any pilot plans)?

Investor overlay

Read-throughs

  • Closed loop integration of experiments as compute suggests demand for lab automation plus orchestration software, with value shifting toward iteration time and throughput rather than isolated model accuracy.
  • Multi fidelity digital twin screening implies a market for tools that manage tiered evaluation, calibration, and compute allocation, where efficiency metrics become the procurement language.
  • AI for science interest and possible bubble dynamics imply near term capital availability and partner driven pilots, benefiting vendors that enable tooling first and human in the loop workflows.

What would confirm

  • Published or customer reported end to end iteration time and throughput gains for closed loop experiment plus compute versus conventional simulation then validation pipelines.
  • Quantitative multi fidelity screening metrics such as hit rates by tier, calibration error, and compute cost per surviving candidate reported consistently across programs.
  • Demonstrated reliability of agents for literature search and experiment orchestration, including reproducibility stats, error rates, and clear boundaries for required human oversight.

What would kill

  • No measurable cycle time reduction from closed loop workflows after deployment, with bottlenecks remaining in physical experimentation rather than decision making.
  • Digital twin tiers fail to correlate with higher fidelity outcomes, showing poor calibration or low hit rate improvements that negate screening efficiency gains.
  • Agent driven orchestration produces frequent errors or irreproducible runs, forcing full manual reversion and preventing scaled adoption beyond pilots.

Sources