Northwestern researcher Edward Sargent applies AI to accelerate energy science discovery
In December 2017, Nature published a two-page comment on machine learning that, nearly a decade later, reads less like a prediction and more like a blueprint. Graduate students Phil De Luna and Jennifer Wei shared authorship with deep-learning pioneer Yoshua Bengio, quantum chemist Alán Aspuru-Guzik, and materials engineer Edward Sargent. Sargent is currently a professor in the Chemistry and Electrical and Computer Engineering Departments at Northwestern University and director of the Trienens Institute for Sustainability and Energy. Sargent and his collaborators argued that discovering better materials for solar energy, batteries, and carbon capture remained too slow and too dependent on manual trial and error. They proposed using deep reinforcement learning to accelerate the search.
“Useful here is the growing field of deep reinforcement learning,” they wrote, “in which agents explore their evolving environment to find the best solutions. Applying such algorithms to materials discovery would make searches progressively more efficient.”
The core idea has held up remarkably well. The reinforcement-learning systems described in 2017, programs that learn through repeated interaction with an evolving environment, closely resemble what are now called agentic AI systems. Instead of simply responding to prompts, these systems plan sequences of actions, delegate work to specialized tools or models, execute experiments, evaluate the results, and decide what to do next. Between 2023 and 2026, that approach moved beyond chatbots and into research laboratories. Autonomous platforms such as Coscientist demonstrated that language-model agents could plan and carry out real chemical experiments [5], while multi-agent systems became a common framework for tackling complex scientific problems. Sargent's own publications over that period reflect the same progression, with each study shifting more of the experimental workflow into an increasingly autonomous feedback loop.
Letting the model choose the next experiment
One key step, published in Nature, came through a collaboration between Sargent and Carnegie Mellon researcher Zachary Ulissi [2]. Sargent, Ulissi, and colleagues started with 244 copper-containing intermetallic compounds from the Materials Project database, generating a search space of more than 12,000 possible surfaces and nearly 229,000 adsorption sites.
Evaluating every possibility with density functional theory (DFT) would have been prohibitively expensive. Instead, the researchers used an active-learning cycle. They ran DFT calculations on a subset of sites, trained a machine-learning model on those results, used the model to identify the most promising unexplored candidates, performed new DFT calculations on those sites, and repeated the process. After roughly 4,000 DFT calculations—less than two percent of the total search space—the model could accurately predict carbon monoxide binding energies across the remaining candidates without additional quantum calculations [2]. The search ultimately identified a de-alloyed copper-aluminum catalyst that converted CO₂ to ethylene with a Faradaic efficiency above 80 percent, setting a new benchmark at the time.
Combining robotic labs with human judgment
Five years later, Sargent's group replaced virtual screening with an automated laboratory [3]. In a Joule paper led by Jiheon Kim, Suhas Mahesh, and collaborators, a robotic platform synthesized and evaluated 300 catalyst compositions spanning a 15-element chemical space for converting CO₂ into propylene. After each round of experiments, the results were fed into an interpretable machine-learning model that proposed the next candidates to test.
The human researchers never left the loop. After every iteration, they adjusted how the model weighed different chemical features, incorporating their own expertise as new data emerged. The best-performing catalyst appeared after only five rounds. Overall, the discovery process was about 165 times faster than a conventional screening campaign—roughly a 33-fold gain from laboratory automation, and another fivefold improvement from human guidance layered on top [3]. Neither the robotic system nor the researchers alone could have reached the result as quickly. Together, they identified Cu₀.₉₈In₀.₀₂, which produced propylene at 42 mmol per gram of catalyst per hour, outperforming previous reports.
Overcoming a longstanding trade-off
A companion paper in Nature Catalysis, co-authored with Jason Hattrick-Simpers, applied a similar strategy to a more difficult problem: designing ruthenium oxide catalysts for the oxygen-evolution reaction in acid-stable water electrolysis [4]. Researchers have long struggled with a familiar compromise—materials that are highly active tend to degrade quickly, while durable catalysts are often less efficient.
The team's mixed acceleration workflow combined theoretical predictions with experimental measurements to estimate not only catalytic activity but also whether each material could realistically be synthesized and how long it would remain stable. Guided by that framework, the researchers carried out 379 experiments and identified seven ruthenium-based oxides that surpassed the existing activity-stability frontier. Rather than sacrificing one property to improve the other, these catalysts proved both more active and more durable than previous leading materials.
Taken together, the four papers trace the evolution of a single idea over a decade: materials discovery becomes more effective when search is shared between machines and scientists. What began in 2017 as a proposal to apply reinforcement learning to chemistry has grown into a broader ecosystem of active learning, self-driving laboratories, and human-guided AI systems. Throughout that progression, Sargent's work has repeatedly helped define what the next stage of AI-assisted materials discovery will look like.
References
[1]: De Luna, P., Wei, J., Bengio, Y., Aspuru-Guzik, A. & Sargent, E. H. Use machine learning to find energy materials. Nature 552, 23–27 (2017). https://doi.org/10.1038/d41586-017-07820-6
[2]: Zhong, M., Tran, K., Min, Y., Wang, C., Wang, Z., Dinh, C.-T., De Luna, P., Yu, Z., Rasouli, A. S., Brodersen, P., Sun, S., Voznyy, O., Tan, C.-S., Askerka, M., Che, F., Liu, M., Seifitokaldani, A., Pang, Y., Lo, S.-C., Ip, A., Ulissi, Z. & Sargent, E. H. Accelerated discovery of CO2 electrocatalysts using active machine learning. Nature 581, 178–183 (2020). https://doi.org/10.1038/s41586-020-2242-8
[3]: Kim, J., Mahesh, S., Lee, H. S. et al. & Sargent, E. H. Accelerated discovery of CO2-to-C3-hydrocarbon electrocatalysts with human-in-the-loop. Joule 9, 102213 (2025). https://doi.org/10.1016/j.joule.2025.102213
[4]: Bai, Y., Li, K., Han, N., Kim, J., Zhang, R., Mahesh, S., Shayesteh Zeraati, A., Sutherland, B. R., Chow, K., Liang, Y., Hoogland, S., Huang, J. E., Sinton, D., Sargent, E. H. & Hattrick-Simpers, J. Stable acidic oxygen-evolving catalyst discovery through mixed accelerations. Nature Catalysis 9, 28–36 (2026). https://doi.org/10.1038/s41929-025-01463-x
[5]: Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023). https://doi.org/10.1038/s41586-023-06792-0