ÐÂÔÂÖ±²¥appÏÂÔØ

Biography

Ed Pyzer-Knapp FRSC has always been passionate about working at the interface of multiple disciplines, and has a particular interest in the link between advanced AI methods and chemistry. Ed is a co-founder and chief scientific officer at Xyme – an AI startup building science-first AI to enable the discovery and development of new biocatalysts for industry. Previously, he was head of research innovation at IBM UK and Ireland, providing technical leadership on the convergence of HPC, AI and quantum computing to accelerate scientific discovery.

Ed is interested in the use of powerful emerging technologies to help to answer some of the biggest scientific challenges of our time. He obtained his PhD from the University of Cambridge working on computational methods to predict crystal structures of porous materials, and then moved to Harvard University where he built some of the first deep learning approaches applied to large scale ntum chemistry as part of the Harvard Clean Energy Project. He returned to the UK in 2015 to help start the IBM Research Lab there. In the following nine years, he led interdisciplinary research programmes with both universities and major industrial companies as part of IBM’s contribution to the Hartree Centre. In 2018, he was granted an honorary professorship at the University of Liverpool and in 2019 became editor-in-chief of the Wiley journal Applied AI Letters. In 2025 he was made a Bye-Fellow of Downing College, Cambridge. He has authored more than 90 papers and conference proceedings, filed multiple patents, and written a textbook on the use of AI for physical science, published by Wiley in 2021. 

Ed was elected to the UK Young Academy in 2023, and was appointed to the Executive Group as co-chair of the UK Young Academy in the same year – a position to which he was recently re-elected. In 2024, Ed was appointed a Fellow of the ÐÂÔÂÖ±²¥appÏÂÔØ.

The most interesting questions are almost always at the boundaries between fields, and the only way to find those questions is to spend time with people who think differently.

Edward Pyzer-Knapp

Q&A

Can you tell us more about your work?

My research develops AI methods for accelerating the discovery of new molecules and materials. This involves both making our discovery pipelines both more efficient, by building powerful approximators, and more effective, by building smarter optimisation engines. 

My guiding principle is that AI works much better for science when it is built from an understanding of the underlying scientific principles, rather than relying solely on large volumes of data. By embedding the laws of physics and chemistry into machine-learning models, as well as understanding the uncertainty of the models themselves, I aim to unlock new uncharted areas of chemical space for application to some of the world’s most impactful problems.

In 2024 I brought all these threads together to co-found Xyme; an AI company building science-first AI for the discovery of new (bio)catalysts. By working from core chemical principles, we build the catalytic chemistry to unlock new feedstocks, processes and systems; with the aim of replacing the heat and pressure of today’s chemical industry with the ambient conditions that nature has shown us is possible.

What have been the biggest challenges that you have faced over the course of your time in science, and what have you learned from those experiences? 

The biggest challenge has been pursuing fundamental research outside traditional academic structures. Most of the people whose work I've admired have university appointments. I never have. That was a deliberate choice, but it meant I had to build credibility, audiences and collaborators from a different starting point each time. Industrial research environments offer unusual freedoms, particularly around scale and engineering, but they also live closer to the chopping block. Fundamental work is often the first thing trimmed when budgets tighten, regardless of how strong the science is.

A second, related challenge has been working at the seam between communities that do not naturally talk to each other. Chemists, machine-learning (ML) researchers, and software engineers each have their own vocabularies, incentives and ideas of what counts as a good result. For a long time, work in the middle was treated with suspicion by all sides. The same paper that excited the chemistry audience would be dismissed as derivative by the ML one, and the other way round.

What I have taken from this is that the work is only half the job. Building a community of people who can read it, evaluate it and use it is the other half, and it does not happen by accident. I now spend almost as much time investing in those relationships, and giving back to the community as I do in the science itself, and I think our progress is actually better for it.

Thinking back to earlier in your career, are there any words of wisdom that you wish someone had told you? 

The first is that the output of your work matters more than the address it comes from. Early in my career I worried, sometimes a great deal, about whether being outside academia would limit what I could do. It has not. The constraints are real, but they are different constraints, not bigger ones. Some of the most interesting science of the last decade has come out of industrial labs, foundations, government centres and small companies. If the work is good, it will travel.

The second is that it’s OK to be unfashionable sometimes. I realise that this sounds odd coming from someone working in AI, but when I started it was a struggle to convince people that this line of research held merit. It is important to keep doing the work in those quiet years, because the discomfort itself is not a signal that the idea is wrong. If anything, it tends to be a signal that the idea is far enough from the consensus to be interesting.

One piece of advice I was given, and I want to take this opportunity to pass on, is what I call Kirk’s Law (named after my mentor, Kirk Jordan at IBM). He always taught me that people only really think on the following timescales: coffee time, lunch time, overnight, too long; and only really appreciate progress when you move the execution time between those scales. A calculation that used to run overnight and now finishes by lunch registers immediately. The same calculation going from thirty minutes to twenty does not, even if the underlying improvement is the same.

What future directions or opportunities do you see for your work? 

The first is the maturing of multi-fidelity AI. Most scientific data exists at very different levels of accuracy and cost: a quick computational estimate, a careful quantum chemistry calculation, a real experiment in a real lab. The useful move is to fold these into a single reasoning system that knows how to combine cheap rough data with expensive accurate data, and that decides for itself which kind of evidence to gather next. 

The second is closed-loop discovery, the natural extension of multi-fidelity AI into the automated lab. Generative models propose candidates; robotic platforms run the experiments; and the AI in the middle makes the real-time decisions about what to do next. For years this was a slide in talks more than a working system. That is starting to change, and the version that matters is the one that closes the loop fully, with the model updating itself as the experiments come in.

The third, particularly relevant to enzymes, is the growing overlap across length scales. Until recently, the quantum mechanics of a single bond-forming step, the structure of the protein around it, and the macroscopic properties of the enzyme as an industrial catalyst were studied in different communities with different methods. They are now starting to be addressed in a single framework, which makes it possible to co-design across scales rather than handing a result from one community to the next.

The fourth, and the most fundamental, is a shift in how we build AI for science. The dominant recipe in AI today is to train large models on large data. In my view, that recipe is not always the right shape. We already understand a great deal about how the world works that is not easily expressed as data, and we work on problems where the data is sparse, expensive, or absent. The interesting frontier is models that bake in this prior understanding as a foundation, leaving data to do the work it is best at. If we get that right, it unlocks the ability to attack problems where the dataset to ‘scale up’ does not exist, which describes many of the problems worth solving.

What do you wish more people understood about your field or the chemical sciences in general? 

Two things. The first is that the recipe behind the current generation of AI, which is essentially to scale up language models on internet data, is not the recipe for AI in science. The internet does not contain the data we need. Most of what matters in chemistry has never been written down, and the parts that have been are skewed toward what worked, not toward what failed. Pouring more compute into a model trained on that skew does not solve the underlying problem. If anything, it concentrates the bias. The AI that will matter for science needs to be built from the principles of science, not borrowed from somewhere else and pointed at it.

The second is the role of chemistry in the energy transition. When most people think about the path to net zero, they think about generation, batteries, and grids. Chemistry sits behind all of it, and behind every other industrial sector that has to be decarbonised: fertiliser, cement, plastics, pharmaceuticals, transport fuels. Roughly a third of global emissions come from chemical and materials production. There is no version of net zero that does not involve a wholesale reinvention of the chemistry, the feedstocks and the processes that underpin these sectors, and that reinvention will need new tools, new theory, and new kinds of catalysts. Chemistry is the single biggest point of leverage in that process. To put it another way, to achieve global change, would you rather convince six billion people to change their behaviour, or to change the chemistry that underpins everything?

How important would you say collaboration is for producing high quality science? How has collaboration influenced your work? 

Almost everything I have done that I am proud of has involved someone whose training is different from mine.

The deepest collaborations have been with ‘real world’ partners. Working with companies like GSK, Dow and Unilever forced my methods to confront real problems with real data, real timelines and real consequences. That confrontation made the science better. A method that works on a curated benchmark and fails on a messy industrial dataset is not, in any useful sense, a working method. The pressure of having to deliver something useful sharpened the work in ways that purely academic feedback rarely does.

The other side has been collaborations with people whose questions I could not have asked on my own. I have worked with structural biologists, quantum chemists, geologists, plant scientists and microbiologists, and each of those collaborations has taught me something I would not have learned otherwise. Some of my favourite results, including work on the microbiome and on circadian regulation in crops, came out of conversations that started with someone saying "your methods might be useful for this problem you've not heard of."

The broader lesson is that the most interesting questions are almost always at the boundaries between fields, and the only way to find those questions is to spend time with people who think differently. I try to be useful to people whose science I do not fully understand, and to ask them, in return, to push back on mine.