Literature Seminar

Usage of data driven approaches such as machine learning (ML) is raising due to the success they achieve in predicting outcomes in chemical problems.  However, there has been constant criticism from the scientific community about what do the ML based predictive models learn from the input data. A small step into understanding the learning process of ML methods was made by the Zimmermann group in University of Michigan. In the recently published paper [1], the authors examine various input representations for a same set of data to understand how the ML models learn. The chemical problem considered in the study involves prediction of activation barriers of atmospheric reactions using the various input representations. Upon decoding how ML models learn, the authors compare the prediction results from ML based models to an already existing simple chemical relationship i.e., Evan-Polanyi relationship. Finally, the authors address how chemical understanding is necessary for better predictions using ML techniques.

Zoom Link