|We can add a new chemical reaction database to our list of chemical reaction databases. See previous database coverage here. Not a brand new one, it has been around since 2016. |
This Haverford College initiative aims to use machine learning to predict the success of chemical reactions. This concept has been described in a 2016 Nature letter (DOI): an algorithm was trained on a set of successful reactions and a set of so-called dark reactions - unsuccessful and failed -, more specifically a hydrothermal synthesis. With training done, the algorithm was then set to work and it proved better in devising new synthesis strategies than humans!
The aim of the dark reactions project (website here) is to collect failed reactions that together with successful reactions can serve as test sets for further machine learning. Failed reactions do generally not make it to publications and hence the repository. The site has a public reactions database. A csv file is available once you have registered as a user. The file contains around 4000 reactions. Curiously the published database only contains sets of reactants: no reaction conditions or products. Not sure why that is. There is also a piece of software downloadable from github here. Not sure yet what that is about either.