Investigation of machine learning-based schemes for the development of coarse-grained force fields for organic molecules

Summary

This master thesis presents the development of an innovative evaluation and ranking protocol that integrates physical insights to assess the performance of models and employs statistical tests for validation. The protocol was applied to address a multi-objective optimization problem related to self-adapting the weights of a Graph Convolutional Neural Network model. Specifically, the Graph Convolutional Neural Network model was designed to simulate a force field for predicting molecular system configurations in a coarse-grained setting. The SchNet architecture, a deep learning framework tailored for atomistic systems and capable of modelling quantum interactions in molecules, served as the foundation for the proposed approach. The training dataset consisted of multiple frames obtained from atomistic simulations of benzene molecules. This research mainly focused on achieving a self-balancing of weights within the dual components of the Graph Convolutional Neural Network model’s loss function. This self-balancing aimed to optimize the trade-off between different objectives in the multi-objective optimization problem, enhancing the model’s performance. To identify the most effective self-balancing method among various alternatives, the developed protocol was employed to evaluate the performance of each approach. The evaluation results revealed the most optimal self-balancing approach for this specific multi-objective optimization problem within the context of simulating a force field for predicting coarce-grained molecular system configurations. The presented methodology offers a valuable contribution to the field of deep learning applied to atomistic systems and multi-objective optimization, paving the way for further advancements in molecular dynamics simulations and related research.

Keywords:

Molecular dynamics simulation ; Graph Convolutional Neural Network ; SchNet ; Multi-component loss function ; Multi-objective optimization