Machine learning can guess how well a reaction will perform, but it rarely explains why. Most models lean on descriptors that are hand-picked and hard to trace back to what the molecules are actually doing, so chemists get a yield figure and no mechanism to act on. A team at Hokkaido University's ICReDD (Institute for Chemical Reaction Design and Discovery) wanted both at once: a reliable yield prediction and a readable picture of the chemistry driving it. ICReDD runs its automated synthesis on a Chemspeed platform, which put reproducible experimental data at the center of the work.
- Instead of arbitrary descriptors, the team built an energy descriptor from the calculated energies of the reaction's possible intermediates, a feature grounded in real chemistry rather than convenience.
- Those intermediates were mapped through automated reaction path exploration (SC-AFIR), so the search ran comprehensively instead of resting on chemist intuition.
- The output is a simple, interpretable regression model that predicts yield and flags the intermediates responsible, delivering mechanistic insight alongside the number.
- Automated experimentation on the Chemspeed system generated the yield data behind the model, keeping measurements consistent and reproducible enough to trust what was built on top.
- Computation and physical experiment ran as one connected loop rather than three disconnected steps: predict, test, explain.
A yield number tells you what happened. Pair Chemspeed automation with reaction path exploration, and the model also tells you which intermediate to thank.
Authors and institutions involved: Takahiro Doba, Yu Harabuchi, Yuuya Nagata, Satoshi Maeda.
Institute for Chemical Reaction Design and Discovery (WPI-ICReDD), Hokkaido University; ERATO Maeda Artificial Intelligence for Chemical Reaction Design and Discovery Project, Hokkaido University; Department of Chemistry, Faculty of Science, Hokkaido University; Institute for Chemical Research, Kyoto University.
Published in
Journal of the American Chemical Society (JACS), 2026, 148 (25), 26150-26160.
https://doi.org/10.1021/jacs.6c05203
This text is an original summary prepared by Chemspeed for editorial purposes. All rights to the original publication remain with its authors and the publisher. No figures, tables, or passages from the original article are reproduced here. For full details, please consult the original publication via the link above. All trademarks and product names are the property of their respective owners.
