There are no big data battery papers, Sam Cooper says. There are only medium data ones.
The data that would justify the phrase sits inside manufacturers and stays there, because it is a genuine industrial secret, and that limits the kind of model anyone outside a cell plant can build. A few corners of industry are beginning to share, which he calls an exciting breakthrough, but the constraint shapes what the field is able to learn.
His own position is unusually plain for an academic. The cluster at Imperial is not the bottleneck: on his gut feeling the battery world sits at the data-limited end of machine learning rather than the compute-limited one. That might not be true inside CATL, where cells have been tested for a very long time. From where he sits it is, and he uses the conversation to ask manufacturers directly for data.
Batteries make the problem worse by being several problems at once, running from the arrangement of a few atoms through microstructure, cell design and pack design up to the behaviour of a vehicle fleet and the charging network. At the atomic end, first-principles calculations are expensive, so people build cheap surrogate models that stand in for the physics. At the pack end there is what feels to him like a hundred papers a week relating what you asked a cell to do to its remaining charge, its efficiency or its remaining useful life. Radically different questions, all requiring the same skills: coding in Python, sending work to a cluster, and cleaning data until it tells the story that needs telling.
They are also quick to acquire. Cooper co-created Imperial's Mathematics for Machine Learning course on Coursera, which has taken over 600,000 students through it since launching around seven years ago. Electrochemistry takes years before you can write down equations and trust them; machine learning lets you train a model that does one small thing in a week.
Learning the microstructure instead of simulating it
The middle of that stack is where Cooper works, and it is the part worst served by physics.
Making an electrode involves mixing material in a bucket, spreading it onto a foil, letting the solvents dry off and crushing it down to make it denser. Simulating those steps from first principles is, in his description, insanely expensive and slow, and currently just not very accurate. The state of the art amounts to dropping imaginary spheres into a box, rumbling them around and pressing them down in the hope the result resembles a real electrode. It does not.
Polaron, the spin-out where he is chief scientist, skips that entirely. Feed a model images of microstructures produced under a range of known manufacturing conditions and ask it to learn the relationships. As he puts it, he does not much care how it learns them. What comes out is an algorithm that can search the space quickly and find the best microstructure for a given application, on the assumption that a large enough dataset implicitly contains the physics.
The payoff is engineering rather than chemistry. Take a factory with, for the sake of argument, a thousand interacting parameters to optimise. Tune them into the right settings and you might get a cell with 20% more accessible energy at a particular discharge rate. The chemistry has not changed. No machine has been replaced. Every knob has simply been squeezed as tight as it will go.
He points at CATL and BYD. Rather than inventing a radical chemistry and reinventing every process around it, they finessed the engineering until conventional chemistries performed remarkably well.
The counterweight to all that precision is a story he heard from Eric Darcy, NASA's head of batteries, who builds deliberate faults into cells to see how they fail. NASA spent a long time trying to make very good batteries for spaceflight and, with the resources of NASA behind the effort, could not. Good cells require enormous scale. So the solution is to buy a very large number of cells from whoever the biggest supplier is at the time, grade them, and fly the best 1%. You cannot make a small number of excellent batteries however much money you throw at it.
Most battery papers could be thrown away
In an academic lab, a cell is made by a PhD student or postdoc inside a glove box, because the gas environment has to be controlled or the material surfaces rot. Each component is cut out by hand and dropped on top of the last, in the hope that the pieces align and that the same quantity of electrolyte went on as last time, then crimped together in the hope the result is repeatable. Often it is not, and there is rarely time for more repeats, which leaves the researcher in an awkward position over what to report.
He accepts that the conclusion is cynical and states it anyway: you could probably throw away 99% of the battery papers ever written and society would not have lost any information.
What changes this is robots. A robotic lab is coming online in Magda Titirici's group at Imperial, essentially an automated cell assembly line inside a glove box, doing the cutting, stacking and crimping a student would otherwise do by hand. The deeper change he wants is in how methods are recorded. Simulations and data analysis are already code; experimental methods are still prose, a sentence saying these things were mixed using that instrument. If the method were code submitted to a robot, running it here or in another country would give the same result, because the process would be controlled tightly enough to make that true.
What a language model is actually good for
On large language models Cooper is practical. They may not be genius oracles yet. What they are is tireless workers.
His example comes from a proposal recently submitted to the European Union, aimed at pulling the drudgery out of science. Designing an experimental campaign, given two years, two million pounds and a question to answer, means reconciling budgets, available chemicals, funder interests and the capabilities of every machine involved, each of which may come with several hundred pages of documentation. That synthesis currently depends on one experienced person recalling what worked before. A model with a long enough context window can instead be told to read the manual and extract what matters for a campaign testing cells of a given type. Asking a PhD student to do it is a great deal of work for a small result, and, in his word, unkind.
Reproducibility here he considers unsolved. Open models let you see inside the machine but lag the closed ones behind an API, and the closed ones change without warning. His suggested minimum is what you would do with a commercial physics solver: state the exact release. That still leaves the fact that these models are stochastic by design and do not give the same answer twice.
Materials discovery by brute machine search he calls stubborn. Cobalt is common in batteries and uncommon in the crust, concentrated in places including the Democratic Republic of Congo and extracted in conditions he describes as generally very unethical, so there are good reasons to want a replacement. But finding the crystal that takes up the most lithium at the best voltage settles almost nothing. It also has to hold a stable interface with the electrolyte, remain stable internally for years and be economically viable, and promising structures rarely clear all those thresholds at once. Then there is the decades-long path from computer to car.
Leaving the university made industry easier
Polaron came out of Cooper's group after roughly seven years of work on microstructure. Two of his PhD students, Isaac Squires and Steve Kench, are now chief executive and chief technology officer, have raised money and won prizes, and work from an office in Shoreditch. He goes in one day a week.
The tool is deliberately not battery specific. Alloy makers care about the relationship between crystal morphology and mechanical performance, and so do concrete, additive manufacture and the polymer coatings on pharmaceutical tablets, where the coating's properties determine whether the medication works. Cooper suspects batteries are among the hardest of those cases rather than the obvious one. You cannot buy the tool with a click yet; the plan is to work with a few companies in a few industries until it meets their needs.
What surprised him most was how much easier industry became to talk to. A project with the university means contracting and IP arrangements in which the college may want a stake, and publication is one of a university's key performance metrics. A company provides a service and is not trying to put the result in the public domain. Inbound interest has been far higher than expected, from people who heard about the work on the grapevine. He also expected startups to be adversarial and academia collegiate, and found the reverse: two other Imperial spin-outs in batteries gave advice generously enough to keep his team out of a lot of holes, while academia can be a bit barbed. His advice to frustrated academics is to start a company.
After eight years of saying everything he knows, he has also had to learn to keep some of it quiet.
The gap nobody has filled
Asked where this goes, Cooper's attention lands on grid storage rather than cells for cars.
He points to the third Tesla master plan, published as an account of how to electrify the world, and to how shocking the gap is between the grid storage it requires and the technologies actually available. Serious people are still discussing storing hydrogen in caverns. Redox flow batteries have been in people's minds for decades without quite arriving. He credits Don Sadoway at MIT with an excellent pitch for liquid metal batteries, which have also turned out harder than expected, while thinking the pitch contains the right idea: if you want dirt cheap batteries, you have to make them with dirt. Nothing containing cobalt will be storing gigawatt hours.
This is the one place he does not think engineering finesse will rescue the situation, and he says so having just watched it do exactly that elsewhere. Nobody expected ten years ago that lithium iron phosphate packs would be energy dense enough for a perfectly good city car. They are. The higher energy materials are only needed, in his framing, if you want to drive silly cars.
Aviation he will not call. It is not clear that long-haul flight with 400 people aboard is ever an electrochemical problem rather than a synthetic fuel one.
What he does insist on is reading the claims properly. He is sometimes asked to review another company's technology, and a recurring trick is metrics reported separately: a plane that flies this fast, for one minute; this far, at ten miles an hour; this long, if you never use it. Announcements of that kind deserve caution, because you cannot see what was compromised to reach them. Batteries are a bubbly space in the way machine learning is, with companies exploding into prominence and fading once the technology turns out to be excellent in one metric only.
His objection to that is not aesthetic. Every dollar invested in a dud is a dollar that could have gone into something meaningful.
This piece draws on the full conversation, which is available with a complete transcript on the episode page.