CuspAI: Turning Materials Discovery into a Search Workflow

The difficult interface is physical

Discovering a material requires computation and physical experiments to work together. The laboratory needs apparatus and careful setup; connecting it to a digital workflow is cumbersome. A useful metaphor from this Latent Space conversation with Max Welling is the physics processing unit: nature performs part of the computation through experiments, alongside the work performed in a data center.

CuspAI’s proposed materials search engine would let a scientist specify desired properties, launch computation and experiments, inspect candidates, and refine the request. That simple interface would coordinate substantial work in remote laboratories and data centers. Full automation remains an aspiration; the practical objective is to give chemists and materials scientists a more powerful research tool.

From fundamental physics to practical impact

Welling’s research priorities have evolved from intellectual curiosity toward a combination of scientific depth and positive impact. Fundamental physics supplied questions worth pursuing for their own sake. Later, concern about climate change and frustration with political responses motivated a technological approach through CuspAI. Materials science offered both difficult scientific problems and consequential applications.

Physics supplies tools as well as questions. Welling credits Taco Cohen as the main driver of their work bringing symmetries into machine learning, progressing from rotations to gauge symmetries on spheres. He recommends his former PhD student Maurice Weiler’s book on symmetries in AI. More recently, he has explored stochastic thermodynamics—thermodynamics for systems out of equilibrium—and its mathematical connections to diffusion models and other learning methods.

His interest in fundamental questions persists. Welling admires his PhD adviser Gerard ’t Hooft’s pursuit of an alternative account of quantum mechanics, particularly the effort to make the theory understandable. He remains interested in understanding quantum mechanics without invoking multiverses or collapses.

Welling sees AI for science gaining momentum because useful methods can now model important scientific problems effectively. He points to protein folding and machine-learning force fields, also called machine-learning interatomic potentials, as successful applications involving symmetries. Health, drug development, energy materials, and carbon capture add practical urgency. He also sees the rapid growth in investment as capable of producing a bubble.

Participation requires better entry points for engineers without specialist science training: curricula joining the fields, accessible books, and introductory tutorials before specialist workshops. An hour of shared foundations before a workshop is one concrete proposal for making the field easier to enter.

Searching for materials that do not yet exist

Even software-centered AI depends on materials. Language models run on graphics processors, whose fabrication depends on materials and lithography. As miniaturization becomes harder, the argument goes, new materials become increasingly important to hardware progress. Batteries, fuel cells, and solar panels make the same dependence visible in the energy transition.

Adding perovskite layers above silicon in solar panels illustrates the potential. The solar example invokes a theoretical capacity to capture up to 50% of light against a tentative baseline of about 22%. This is theoretical potential, not a demonstrated efficiency gain: neither the device configuration nor measurement conditions are specified.

The proposed shift in research is to search possible molecular structures, including ones never made. Instead of relying on a slow cycle of reading papers, forming a hypothesis, and running an experiment, a scientist could specify a goal and use computation and experimentation to explore candidates, then revise the query. Searching all possible molecules describes the ambition’s scope, rather than demonstrated exhaustive coverage.

A plastic that breaks down after a few weeks and becomes fertilizer makes that ambition tangible as an aspirational design target. Carbon removal provides a founding motivation: the account emphasizes the difficulty of removing atmospheric carbon dioxide at enormous scale alongside reducing emissions.

At the time of the conversation, CuspAI was described as roughly 20 months old, with Chad Edwards named as co-founder and a team of about 40 people. Its computational platform covered several material classes and needed continued expansion. Large language models were being incorporated to automate more work, while high-throughput experimentation was being connected to the platform to provide faster feedback.

Generate broadly, spend selectively, test physically

The architecture combines a trained candidate generator with a digital twin, a computational representation used to evaluate candidates. Its multiscale, multifidelity design evaluates materials at different scales and levels of detail. Multiscale modeling has gained importance since the original design, and self-driving laboratories have become a timely addition.

Candidates climb a ladder of increasingly expensive evaluations: 1. Generate candidates with the trained generative component. 2. Reject obvious failures cheaply before committing more resources. 3. Apply more expensive evaluations to the remaining candidates. 4. Experiment on a small set and collect feedback.

Agentic components support two distinct responsibilities: - Literature search: search chemical literature to suggest chemicals or experiments. - Workflow orchestration: coordinate the computations and experiments that need to run. These components have different levels of maturity and substantial room for improvement. The company locates its defensibility in access to data and the difficulty of building the platform.

The physics-processing-unit metaphor points toward tighter integration of experiments and computation. Experiments have served as an end-stage check; the vision is to make their feedback part of a more closely connected system. That requires coordinating digital execution with difficult experimental setup and bulky physical infrastructure.

Learn the workflow before automating it

Chemists assemble the workflow first. Development begins with modular tools for a particular application, which experts initially select and run manually. Automation follows once the workflow is understood.

A porous material that collapses when shaken illustrates why this matters. The workflow has exposed a missing requirement, so the team adds a stability-testing tool. Selecting and sequencing existing tools is only part of the problem: experts must recognize when the toolset fails to test something consequential.

Once the tools exist, an automated controller could select and sequence them. Proposed approaches include a Bayesian optimizer or a language model, potentially trained for chemistry. These are possible orchestration mechanisms, rather than a single prescribed implementation.

A concrete target is the repeated specialist consultation surrounding a DFT calculation. Its underlying theory is left as a prerequisite, but the responsibilities to automate are clear: - Choose an appropriate calculation for the scientific problem. - Set parameters and runtime suited to that task. - Assess the outcome, rather than treating completion as evidence of a useful result.

Automating those decisions would reduce routine consultation while chemists retain judgment over candidate materials. The intended search interface returns candidates for expert evaluation. A laboratory that independently decides what is interesting is rejected as a near-term objective, and the expertise required to make and deploy materials is not expected to be fully automated within five years.

A breakthrough can precede full autonomy

An industrial partner is a condition for entering a new direction. CuspAI’s stated policy is to invest only once it has a suitable partner, because bringing a material into use requires domain expertise. Its preferred engagement is a deep partnership, but its commercial strategy accommodates two forms: - Long-term development: work toward consequential materials, including an undisclosed moonshot intended to demonstrate a material reaching real-world use with AI playing an essential role. - Bounded projects: paid work such as training a force field and analyzing a particular problem, with further collaboration left open.

The commercial premise is that tools can deliver value as they are built, before a major discovery arrives. Water filtration is one illustrative customer problem. Early breakthroughs are expected to depend on close collaboration between chemists who understand the application and AI researchers who develop and train models.

Automation is application-specific. Moving to a new materials problem can require changes across the system: - Experimental setup: different apparatus and procedures. - Characterization equipment: different machines for examining the materials. - Models: retraining or fine-tuning for the new material class. Lessons transfer, but humans must help define the new problem and redirect the system. The hoped-for sequence is a breakthrough material before complete automation: scientific progress need not wait for an autonomous research process.

Symmetry reduces data needs—but constrains optimization

Equivariance embeds symmetry into a neural network, here through constraints on its weights. In the bottle example, incorporating rotational symmetry lets learning in one orientation generalize across orientations, reducing data requirements. Other symmetry groups include translations and permutations, with permutations relevant to graph neural networks.

There are two ways to introduce this structure: - Architectural constraints encode symmetry directly. They avoid needing training examples for every possible rotation; exact coverage through augmentation would require infinitely many rotations. - Data augmentation supplies transformed examples. It lacks that exact guarantee, but can sometimes perform better with abundant data and easier optimization.

Hard-coded symmetry does not guarantee better performance. The suggested explanation is that constraining weights can complicate optimization and make good minima harder to find. The result depends on the application, available data, and interaction between constraints and training.

The broader tradeoff is between data and inductive bias, the structure a model brings to learning. An incorrect bias can impose a performance ceiling. Known symmetries remain useful structure, but architectures should also support scaling unless the dataset is tiny. The expectation is that lessons from language-model scaling will eventually apply to scientific modeling too.

Physics can improve learning algorithms too

The conversation closes with Generative AI and Stochastic Thermodynamics, a forthcoming technical book developed from Welling’s course at the African Institute for Mathematical Sciences near Cape Town. He says it took two years to complete and had been sent to a publisher. The book maps mathematical correspondences between generative AI and physical systems out of equilibrium.

Welling sees shared mathematics across stochastic thermodynamics, diffusion models, reinforcement learning, and other sampling methods. The book discussion connects this work to earlier variational free-energy work by Geoff Hinton and Radford Neal, and to Karl Friston’s free energy principle and active inference. Its technical treatment places models from machine learning alongside corresponding concepts from physics.

The opportunity runs in both directions: physics can guide better learning algorithms, and learned models can help scientists. Results such as fluctuation theorems are proposed as resources for algorithm design, much as symmetries have been. At the time of the conversation, publication in April, in time for a planned ICLR keynote, remained a hope dependent on the publisher.