Predicting Bond Energies with AI

University of Notre Dame researchers developed a deep learning-based system that can accurately determine bond energies.

“Neural networks can be used to make quantitative models of chemical concepts that are not possible with just quantum mechanics,” explains John Parkhill, Assistant Professor of Chemistry & Biochemistry at the University of Notre Dame in Indiana and co-author of the paper. “[Before], I got asked all the time: ‘How much stronger is this bond?’ […] Now I can answer that question, because my machine learnt chemical bonding concepts.”

Using Tesla K80 GPUs and the cuDNN-accelerated TensorFlow deep learning framework, Parkhill and his team trained their neural network on a database of over 130,000 molecules. The trained Bonds-in-Molecules Neural Network (BIM-NN) is able to make predictions of relative bond strengths as well as a trained synthetic chemist.

The network learns the total energies of the popular GDB9 database to a competitive MAE of 0.94 kcal/mol on molecules outside of its training set, is naturally linearly scaling, and applicable to molecules consisting of thousands of bonds. More importantly, it gives chemical insight into the relative strengths of bonds as a function of their molecular environment, despite only being trained on total energy information. Source: American Chemical Society

“[Our network] predicts [data] quantitatively and reproducibly. It saves chemists from the impossible tedium of predicting an energy a billion times,” said Parkhill in regards to their software replacing trained chemists.

Read more >