Delta learning and its applications in computational chemistry

Abstract

This thesis explores the innovative framework of delta Learning and its diverse applications in computational chemistry, focusing on enhancing predictive accuracy and efficiency in chemical modeling. The first chapter provides a comprehensive literature review of delta Learning, outlining its foundational principles and methodologies. By synthesizing existing research, this chapter identifies the strengths and limitations of traditional machine learning approaches in capturing complex chemical phenomena, thereby establishing the necessity for a Delta Learning paradigm that adapts dynamically to new data.

In the second chapter, a GNN model followed by a delta Learning architecture is applied to predict the heats of formation for small molecules, reaching MAE 1.8 kcal/mol, with 𝑅2=0.9972. The process is a pre-training and fine- tuning process: a GNN is applied to extract molecular structural information from molecular graphs, and predict heats of formation at DFT level; then a fully connected neural network is exploited to calibrate DFT level heats of formation to experimental level. Utilizing a dataset of only 405 molecular structures and their corresponding thermodynamic properties, delta learning is demostrated to effectively improve accuracy of computed properties.

The third chapter extends the application of delta learning to the prediction of open circuit voltages of lithium-ion batteries, a vital area in energy storage research. By integrating experimental data with molecular dynamics simulations, this chapter illustrates how delta learning can enhance the fidelity of voltage predictions across different states of charge. The calibration is able to predict open circuit voltage at MAE of 0.13 V, while 𝑅2 reaching 0.933, which demonstrates the potential of delta learning in refining computational models for battery materials. There also presents an automated workflow of delta learning, which can be easily applied to other chemical systems for non- experts in model training. Overall, delta learning is a robust framework that bridges the gap between empirical data and predictive modeling, offering a versatile approach to enhancing the accuracy and efficiency of chemical predictions. The findings presented in this thesis contribute to the growing body of knowledge in computational chemistry and provide a foundation for future research that leverages delta learning in diverse chemical applications.