Machine Learning Supported Mesoscale Modeling of Gas Conduction in Nanoporous Systems

Heat transfer through nanoporous systems is a complex phenomenon that involves contributions from solid and gas phases. Existing theories for estimating thermal conductivity of gases in nanoporous materials, such as aerogels, often rely on one-dimensional heat transfer assumptions and neglect the influence of sidewalls in confined geometries. Moreover, the impact of molecular surface forces, which dominate nanoscale confinements, is often overlooked in the literature. This dissertation addresses these challenges by developing a new mesoscale boundary condition that accounts for nanoscale molecular surface forces and macroscale kinetic theory gas/solid energy exchange. Additionally, data-driven algorithms are used to develop structure-property correlations with thermal conductivity, reducing the need for extensive simulations of different pore size configurations.

Machine learning algorithms are robust predictive models. However, many algorithms exhibit "black-box" behavior, which limits their interpretability. Complex models achieve high accuracy but often fail to provide insight into the underlying physical mechanisms. This makes them less suitable for engineering design, where understanding mechanisms is crucial. This dissertation develops a macroscopic engineering model that offers a promising solution by encoding domain knowledge into mathematical structures. Integrating domain knowledge constrains model behavior, ensures physical consistency, and reduces data needs.

The macroscopic engineering model uses 96 molecular dynamics simulations that incorporate the developed mesoscale boundary conditions. This model predicts the normalized effective thermal conductivity of argon gas in nanoporous aerogel materials as a function of three key parameters: aspect ratio, Knudsen number, and sidewall temperature. The developed model is then compared with three machine learning models: random forest, gradient boosting, and Gaussian process. To ensure the explainability of the machine learning models, explainable approaches are utilized. With R² = 0.9803, the macroscopic model achieves high accuracy while maintaining complete interpretability. Feature importance analysis reveals consistent findings across all modeling approaches: aspect ratio, enabled by the mesoscale boundary conditions, dominates transport behavior (95-99% importance), while Knudsen number and sidewall temperature contribute marginally (1-2% and 2-3%, respectively). The results of this dissertation provide insights and solutions to the challenges of studying gas conduction at the nanoscale, and contribute to a deeper understanding of heat transfer in nanoporous systems.