- 53 views
- 63 downloads
Machine Learning Enabled Prediction of Solvent-Based Reaction Energetics
-
- Author / Creator
- Mao, Yiren
-
Despite recent advancements in molecular dynamics (MD) methods, the computational costs of \emph{ab initio} molecular dynamics simulations for explicit solvation systems are still too significant. If accuracy is to be left uncompromised, new methods must be employed to reduce computational expenses. This work focuses on the development of machine learning (ML) models as proxy models for Car-Parrinello molecular dynamics (CPMD) metadynamics simulations in condensed-phase biomass reactions.
Explicit solvation CPMD metadynamics simulation data of HMF undergoing protonation in a solution of dimethyl sulphoxide (DMSO) and water is used to train various model architectures to make time-series predictions of their probability distribution functions (PDFs). For each model architecture, three models were trained to fully predict the system, one for each of the following species: reactants, water and DMSO. Each model was tested assuming an initial simulation had been performed and the proxy models were used to complete the simulation.
The long short-term memory (LSTM) autoencoder and 3D convolutional neural networks (CNN)-LSTM autoencoder architectures failed to accurately capture PDF magnitudes and locations. A binary relevance 3D CNN-LSTM autoencoder, employing different loss functions, showed marginal improvement but struggled to predict probability locations over a large horizon. Models trained on principal component analysis (PCA)-transformed and dynamic PCA (DPCA)-transformed data showed promise in training but failed in testing. Models trained on PDFs without "dead voxels" (zero probability voxels independent of time) and atomic Cartesian coordinates perform well during training but encounter challenges in testing due to teacher forcing. Teacher forcing is a training method that can potentially make the trained model over-reliant on ground truth, which is unavailable if the model is to be used as a proxy. Despite attempts to mitigate teacher forcing effects through scheduled sampling, no model architecture achieves reliable long-term predictions without ground truth data. However, the model trained on Cartesian coordinates demonstrated proficiency in making short-term predictions regarding the atomic configuration of the system.
-
- Graduation date
- Spring 2024
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.