I’ve been collaborating with professor Craig Butts and his PhD student Will Gerard on trying to predict scalar coupling constants from molecular structures using machine learning. So far, the chemistry machine learning community has mostly focused on the prediction of molecular or atomic properties, so there’s no precedence on how to predict atom-pair properties like scalar coupling constants. While we’ve had some initial success modelling one bond couplings between carbon and hydrogen by extending the methods used for atomic properties, I felt like we should be able to improve on this.
Last year I went to the International Workshop on Machine Learning for Materials Science conference in Helsinki. As mentioned in a previous blog post, Dr. Christopher Sutton gave a great talk there on his experience with competition-based research on Kaggle and I figured that a similar competition could yield a wide range of distinct and well performing algorithms for predicting scalar coupling constants.
Because of this, my last few months have been busy with preparing a data set suitable for the competition format as well as sorting out the funding. Luckily, I was able to secure the funding via my grant (CHAMPS).
Kaggle waives their usual fees for a few competitions a year if the competition is hosted by an academic institution for research purposes and luckily they decided that they would do so as well for my proposed competition. At this point it should be noted that all communication with the Kaggle team as been very painless and that the team has been extremely helpful. They even decided to add a large pot of their own money to the prize pool!
The competition launched less than a week ago and there’s already been hundreds of submissions. The community has proved very helpful in gathering the related theory and information, answering questions and providing helpful code snippets or full solutions. The leaderboard continues to show improvements daily and I’m very excited to see what results from this.
Check out the competition by pressing the image link below!