AI Needs Somewhere to Learn

A model can propose a new catalyst in seconds. Finding out whether it works requires making it, measuring its performance and watching what happens as it degrades. Those steps determine whether the next proposal is better informed than the last.

As AI improves, this relationship between prediction and experiment becomes more consequential. Scientific companies gain the ability to explore more possibilities. Their advantage depends increasingly on how well they turn those possibilities into knowledge they can use again.

The State of AI Report 2026 brings several parts of this shift into view: agents that search for better algorithms, systems that learn from laboratory measurements, and robots that collect their own experience. Across these examples, progress depends on an environment that can return useful feedback. In science, building that environment is becoming an important company-building opportunity. 1

The experiment becomes part of the AI system

Automated experimentation predates the current generation of models. Researchers have long connected algorithms to instruments that select and test the next candidate. The emerging opportunity is to extend the range of work these systems can coordinate, including interpreting unfamiliar results and deciding what to investigate next. 2

Software offers a relatively clear illustration. Google DeepMind’s AlphaEvolve combines models that propose code with automated evaluators that test it. Google reports that one resulting scheduling improvement recovered an average of 0.7% of its worldwide compute resources. The evaluator gives the search a direction and a way to recognise progress. 3

A laboratory has to construct its own version of that feedback. A promising material must be synthesised successfully. Instruments must measure the properties that matter. Results need enough context to distinguish a failed idea from a failed experiment. Each stage affects what the system can learn.

Lila’s September catalyst preprint makes this tangible. Its human-supervised platform evaluated 2,942 catalysts across 53 material systems, searching for alternatives to iridium and ruthenium in acidic oxygen evolution, a reaction used in hydrogen production. One palladium-based composition maintained the paper’s specified performance threshold for more than 1,000 hours in laboratory testing. The system connected synthesis and measurement with models that guided subsequent experiments. 4

These are research results, with industrial validation still ahead. Palladium also remains a precious metal. The significance is the ability to explore a difficult trade-off between activity and durability, then advance promising candidates into longer tests. A useful discovery system has to organise that progression.

Our interest in this category begins with the possibility that experimental infrastructure and AI improve together, opening scientific searches that would otherwise be too slow or costly to pursue.

The experimental loop / 01

An idea meets the physical world.

↻ Replay

What actually gets better

The phrase “learning from experiments” can describe different achievements. A system can use a new result to choose its next experiment while leaving its underlying model unchanged. It can train a model on accumulated results so that future analysis improves. More ambitiously, knowledge acquired in one scientific domain can help it solve problems in another.

These achievements have different commercial implications. Better selection can produce a valuable material. Better models can make an entire research programme more productive. Transfer across domains could support a much broader scientific platform. Evidence for the first does not establish the third.

Periodic Labs offers an example of model improvement. The company reports that training an open-weight model on its laboratory data raised success on a difficult internal X-ray diffraction analysis benchmark of 134 samples from 2.7% to 55.3%. The task is to interpret what materials an experiment actually produced. Its evaluation uses model judges calibrated against scientists, and the results remain company-reported. 5

That detail matters. Physical measurements do not always yield a simple correct or incorrect answer. Interpretation can depend on preparation conditions, competing explanations and expert judgment. The quality of this judgment becomes part of the learning system.

Periodic also reports gains on chemical systems excluded from training. That supports transfer within the measured task; it does not demonstrate a model that generalises across all of science. A specialised capability can still be commercially important. Companies should be able to build large businesses by becoming exceptionally good at a valuable class of problems.

Three different claims / 02

What does “learning” change?

01Experiment choice02Model improvement03Transfer across domains

A different next experiment

A new result changes which experiment comes next. The model can stay unchanged.

↻ Replay

The useful unit is a better decision

Once a company can run more experiments, it becomes tempting to use experiment count as the measure of progress. That number can conceal the economics.

A fast assay may be a poor guide to a product’s performance. A durability test may take weeks. An instrument can produce measurements faster than a team can interpret them. Research on self-driving laboratories already distinguishes throughput from precision, operating reliability and optimisation performance. More measurements cannot reliably compensate for poor experimental quality. 2

For a company, we would translate this into a practical question: how much time and money does it take to reach a decision that advances the product?

Consider a hypothetical coatings business. Its model can generate thousands of formulations, but the customer needs a coating that survives exposure, adheres to the right substrate and can be manufactured at an acceptable cost. The company’s advantage might lie in a screening method that predicts those requirements earlier. It might lie in selecting a small set of experiments that eliminates a whole family of unsuitable formulations.

Simulation can help by narrowing the search before expensive tests. Physical experiments then establish where the simulation is useful and where it is misleading. The commercial objective is to allocate each kind of testing to the uncertainty it resolves best.

This changes the case for automation. The largest return may come from making a previously impractical search economical. A small scientific team could pursue a product that once required the experimental capacity of a much larger organisation.

Access and ownership create different businesses

Experimental capability can reach customers in several ways. A company can use it internally to discover products and retain the resulting intellectual property. It can sell experiments to other researchers. It can supply instruments or software that make existing laboratories more productive.

Adaptyv Bio, based in Lausanne, illustrates how access is becoming programmable. Its API lets software submit protein sequences for laboratory testing and receive binding data back. That makes experimental capacity available within a computational design workflow. A customer can build around this service without first assembling the laboratory itself. 6

Each model has a different route to value. A product company needs its discoveries to survive qualification and reach customers. An experimental service needs dependable turnaround and attractive economics at useful utilisation. An equipment supplier benefits when a growing number of laboratories need its capability. A broad scientific platform additionally needs evidence that shared capabilities make successive programmes more effective.

Data rights influence which of these businesses a company can become. Performing an experiment for a customer does not automatically confer the right to train on the result or reuse it in another customer’s programme. A service that protects each customer’s information may attract demand precisely because it makes that separation clear.

Such a company can still improve its operations and build valuable know-how. Its investment case should reflect those advantages. A claim that every customer strengthens a shared scientific model requires both permission to learn from the data and evidence that the learning transfers.

Owning a laboratory is therefore only one possible starting point. Equipment can be bought by competitors, and unused capacity consumes capital. What becomes difficult to reproduce may be a validated measurement method, an accumulated understanding of failure, or the integration that turns an unexpected result into a productive next experiment.

A company can begin with one scientific advantage

For Aldea, this connects computing with biology and the physical world in a particularly practical way. The opportunity can begin inside a specialist community: a team that understands an industrial constraint, an instrument that makes a difficult property measurable, or a method that turns an inconsistent experiment into a dependable service.

Europe’s scientific founders have several routes into this market. Adaptyv demonstrates one: making local experimental expertise accessible through software to researchers elsewhere. The broader opportunity is to build global companies around capabilities whose value rises as more people and machines can design things worth testing.

The strongest starting point may be narrow. A team that improves one consequential measurement can become essential to many research programmes. A company that owns a productive discovery process can expand through related products. Broader platforms can emerge where the evidence shows that accumulated knowledge reduces the effort required to enter the next domain.

This is where we would place the ambition. Better AI can expand the number of scientific ideas worth pursuing. Businesses that make those ideas testable, and turn the results into a lasting advantage, can expand what an industry is able to build.

The next important scientific company may begin with a model available to everyone and an experiment that very few teams know how to run.

Disclosure: Aldea is an investor in Lila. 7