Physically informed machine learning for functional materials
File(s)
Author(s)
Keeratikarn, Keerati
Type
Thesis
Abstract
This thesis investigates the incorporation of physically informed machine learning techniques—particularly Gaussian Process (GP) regression and Atomic Cluster Expansion (ACE)—to effectively study and model the complex physical behaviours of functional materials. Understanding the electronic, thermal, and optoelectronic properties and applications of these materials requires accurate atomistic modelling. The atomistic modelling continues to pose computational difficulties due to the sophisticated anharmonic effects in crystalline inorganic systems and the complex electronic coupling dependencies in organic semiconductors.
The first section presents a derivative Gaussian Process framework that effectively evaluates harmonic and cubic anharmonic lattice force constants directly from data obtained through density functional theory (DFT). This GP-based approach significantly improves computational efficiency, surpassing conventional finite-displacement techniques in predicting phonon dispersion, lifetimes, and thermal conductivity. Our model assesses multiple descriptor strategies, encompassing Cartesian and phonon-based representations, emphasising their influence on accuracy and computational efficiency.
The second section enhances the Atomic Cluster Expansion methodology to precisely predict charge transfer integrals essential for charge transport in organic semiconductor materials. ACE models (ACEpotential.jl), employing systematically developed symmetry-adapted descriptors, effectively capture the electronic coupling's dependence on molecular orientation and separation in various molecular dimers, including naphthalene, thiophene, ethylene, and their fluorinated derivatives. Our results underscore the significance of descriptor selection and illustrate enhanced model generalisation across chemical systems.
Our research collectively establishes a unified methodology that integrates rigorous physical insights with sophisticated machine learning techniques. This integration greatly enhances the computational resources for predicting material properties, facilitating efficient exploration and precise optimisation of innovative functional materials. The thesis concludes by delineating future research avenues, encompassing computational improvements, sophisticated descriptor formulation, and applications to non-equilibrium and real-time dynamic processes.
The first section presents a derivative Gaussian Process framework that effectively evaluates harmonic and cubic anharmonic lattice force constants directly from data obtained through density functional theory (DFT). This GP-based approach significantly improves computational efficiency, surpassing conventional finite-displacement techniques in predicting phonon dispersion, lifetimes, and thermal conductivity. Our model assesses multiple descriptor strategies, encompassing Cartesian and phonon-based representations, emphasising their influence on accuracy and computational efficiency.
The second section enhances the Atomic Cluster Expansion methodology to precisely predict charge transfer integrals essential for charge transport in organic semiconductor materials. ACE models (ACEpotential.jl), employing systematically developed symmetry-adapted descriptors, effectively capture the electronic coupling's dependence on molecular orientation and separation in various molecular dimers, including naphthalene, thiophene, ethylene, and their fluorinated derivatives. Our results underscore the significance of descriptor selection and illustrate enhanced model generalisation across chemical systems.
Our research collectively establishes a unified methodology that integrates rigorous physical insights with sophisticated machine learning techniques. This integration greatly enhances the computational resources for predicting material properties, facilitating efficient exploration and precise optimisation of innovative functional materials. The thesis concludes by delineating future research avenues, encompassing computational improvements, sophisticated descriptor formulation, and applications to non-equilibrium and real-time dynamic processes.
Version
Open Access
Date Issued
2025-11-21
Date Awarded
01/12/2025
Advisor
Frost, Jarvist
Sponsor
Thailand. Krasuang Witthayasat lae Theknoloyi
Publisher Department
Department of Physics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)