What the research is about
Inside factories, raw materials mix together, heat rises, and new substances are produced. Behind these reactions is a key player called a catalyst. Catalysts help chemical reactions proceed faster and more efficiently, making them essential in the production of many materials used in everyday life. However, finding the best catalyst for a reaction is not easy.
Traditionally, catalyst development has relied on trial and error. Researchers propose possible candidates and repeatedly test them in experiments. Theoretical calculations have also been used, but they often require enormous computing time. Predicting multiple complex reactions and proposing new catalysts can require supercomputer-level computing.
Artificial intelligence (AI) has recently been introduced into this field. However, most existing AI systems only predict which option might perform better within known data. They are not designed to create entirely new catalysts.
To address this challenge, a research team led by Associate Professor Masahito Ohue at Institute of Science Tokyo (Science Tokyo) has developed a new AI framework called CatDRX. This AI learns the fundamental principles of chemistry from large datasets and can reason about chemical reactions even when only limited data are available.
Why this matters
A key feature of CatDRX is that it can design catalysts while taking reaction conditions into account. The system reads information about reactants, products, and reagents, interprets the chemical reaction, and proposes suitable catalyst structures. It then evaluates the generated candidates using chemical knowledge. The research team also verified the AI's proposals using theoretical calculations to confirm that its reasoning was consistent with chemical theory.
Even for catalysts that previous AI systems could not propose, CatDRX can begin with the goal-"We want this reaction to succeed"-and work backward to identify catalysts that could achieve it. This reverse-design approach represents a major step forward in catalyst discovery.
One of the biggest challenges in the study was determining whether CatDRX could truly generate new catalysts, rather than simply recombining existing knowledge. To overcome this, the team trained the AI on diverse chemical reaction data from around the world before applying it to specific tasks. Through this pretraining strategy, CatDRX learned broad chemical patterns and can now propose plausible catalysts even for reactions with limited available data.
CatDRX can independently identify trends such as which types of catalysts tend to work well with certain combinations of raw materials. It can also distinguish what shapes and properties are suitable for catalysts in different reactions. This suggests that the AI is beginning to understand the compatibility between chemical reactions and catalyst structures, effectively allowing it to reason about chemistry.
What's next
This technology could significantly accelerate the discovery of new catalysts in the chemical and pharmaceutical industries. More efficient reactions can reduce waste and energy consumption, contributing to more environmentally friendly manufacturing.
Catalyst research is moving away from a process driven mainly by experience and intuition. Instead, researchers can now start with a clear goal and use data-driven tools to narrow down promising candidates more efficiently.
Comment from the researcher
For a long time, the design of new materials has depended heavily on the experience and intuition of experts, and that will continue to be important. Our goal is to combine that expert knowledge with AI to push research even further.
AI is evolving from a tool that simply predicts answers into one that can offer new perspectives. By combining AI with human imagination, we believe it will become possible to discover catalysts that no one has ever seen before.
(Masahito Ohue, Associate Professor, Department of Computer Science, School of Computing, Institute of Science Tokyo)
