Reaction to 'AI helps chemists develop tougher plastics' by the MIT News

MIT News just published an article detailing how researchers were able to create stronger polymers through a machine learning oriented approach. The breakthroughs accomplished in this article testify to the capabilities of machine learning in advancements across various fields such as chemistry.

The article highlights the difficulties of identifying mechanophores by hand; an arduous task that 'requires either time-consuming experiments or computationally intense simulations of molecular interactions'. Enter machine learning; by employing a model that utilizes a neural network to identify potential ferrocenes as mechanophores, researchers at MIT and Duke were able to speed up the characterization of these molecules. Using obtained data from computational simulations of around 400 ferrocenes and their compound structures, the scientists trained the model to predict the force required to activate the mechanophore. From the model, the scientists discovered a new cause of increased tear resistance internal to polymers, in that large bulky molecules counterintuitively made molecular dissemination more imminent.

As the scientists noted themselves, this trait was 'not something a chemist would have predicted beforehand, and could not have been detected without AI'. This further exemplifies the usefulness of machine learning in unlocking hidden insights in complex processes, notwithstanding the biological and chemical phenomenons focused on in this article. As a whole, this article represents one of many optimistic representations of machine learning, and the potential good it can do for us in the future.

However, I am a bit curious about the methodology of the scientists when designing the machine learning model. Firstly, what was the rationale behind their implementation of a neural network? Was the model a simple network with only a few layers, and if so, what were the hidden layers in action behind the face of the model that were incorporated into the model? Would the employment of a machine learning model with a differing base, for example with deep learning and a higher number of layers, have increased performance on force predictions? I wonder what the thought process of the researchers was as they designed the model from the ground up.

Another impeding question I have on their analysis that I believe any respectable data scientist or machine learning engineer would ask is related to the quality of their data. Renowned MIT postdoc Ilia Kevlishvili states that due to diminishing concerns of synthesizability, the lab was able to 'pick a really large space to explore with a lot of chemical diversity, that also would be synthetically realizable'. It also seems that the data taken for training the model was taken exclusively from ferrocenes. I wonder if the 'chemical diversity' that the lab was able to amplify allowed them to account for possible bias in the data analysis, such as possible overfitting? What measures did the lab take to ensure for evaluating the performance of the model and how did they partition/evaluate the data (batch training, on-line learning, etc.)? These are all questions I am curious about.

Overall, the researchers' implementation of machine learning to derive new chemical insights invisible to the naked eye demonstrates a creative way of exploiting the potential of this discipline. The success and breakthroughs expressed in this article are inspirational towards those (including me) who are actively involved in harnessing this field to derive innovative solutions for research or community service.