Deep learning (DL) can potentially substitute traditionally used expensive quantum mechanical (QM) methodologies in chemistry to predict various chemical properties by offering to generate fast and accurate mathematical models better suited to everyday computers. However, training DL models like
neural networks require hundreds of thousands to millions of data points of a particular chemical property to attain good generalization. Such a requirement is currently hindering the development of DL models for chemistry because the generation of large training data using accurate QM methodologies has an infeasible computational cost.

The SUDS Scholar will tackle this problem and utilize a novel quantum mechanics-based approach to efficiently yet accurately generate large QM training data (hundreds of thousands to millions of data points) for a chosen chemical property. Once the QM training data becomes available, they will utilize it to generate new DL models for the chosen property to demonstrate the acceleration in DL for chemistry provided by applying the novel quantum mechanics-based approach.

Fast and accurate computational modeling of chemical properties and structures has immense potential to accelerate discovery in various fields, including drug design and catalysis. In this context, predicting transition state (TS) structures of chemical reactions that cannot be obtained experimentally offers a powerful way to gather mechanistic details and generate energy profiles. Unfortunately, most traditional computational chemistry methods for predicting TS structures are still very costly to expedite high-throughput applications. Deep learning (DL) can potentially substitute traditionally used expensive quantum mechanical (QM) methodologies in chemistry to predict these structures by offering to generate fast and accurate mathematical models better suited to everyday computers. In this project, the student will further explore this promising avenue and utilize a chemical reaction data set being currently generated in-house. They will utilize it to develop a novel graph neural network (GNN) based model for predicting the highly desired transition state structures at an unprecedented speed and demonstrating the acceleration provided via DL. The research aims at providing a novel way to cut down the computational cost and manual intervention associated with TS structure prediction. Such developments hold immense potential to advance computational chemistry and accelerate high-throughput applications like drug discovery and catalysis. The student will assist in solution design, development of research code (with PyTorch, Keras, Tensorflow, and Python), deployment and running on HPCs. They will help with the generation of reference chemical property datasets. They will then be engaged mainly in designing graph convolutional neural network architecture for structure predictions. The proposed research will be coordinated by the Supervisor and a postdoctoral fellow with experience in running interdisciplinary collaborations. A key responsibility will be to provide one-on-one support to the student, which includes guidance on interdisciplinary method development, data analysis, result interpretation, and effective research communication in the form of published articles and presentations.