A supervised learning-assisted multi-scale study for thermal and mechanical behavior of porous Silica.

This paper presents a comprehensive investigation of mesoporous Silica utilizing a multi-scale modeling approach under periodic boundary conditions integrated with machine learning algorithms. The study begins with Molecular Dynamics (MD) simulations to extract Silica's elastic properties and thermal conductivity at the nano-scale, employing the Tersoff potential. Subsequently, the derived material characteristics are applied to a series of generated porous Representative Volume Elements (RVEs) at the microscale. This phase involves the exploration of porosity and void shape effects on Silica's thermal and mechanical properties, considering inhomogeneities' distributions along the X-axis and random dispersion of pore cells within a three-dimensional space. Furthermore, the influence of pore shape is examined by defining open and closed-cell models, encompassing spherical and ellipsoidal voids with aspect ratios of 2 and 4. To predict the properties of porous Silica, a shallow Artificial Neural Network (ANN) is deployed, utilizing geometric parameters of the RVEs and porosity. Subsequently, it is revealed that Silica's thermal and mechanical behavior is linked to pore geometry, distribution, and porosity model. Finally, to classify the behavior of porous Silica into three categories, quasi-isotropic, orthotropic, and transversely-isotropic, three methodologies of decision tree approach, K-Nearest Neighbors (KNN) algorithm, and Support Vector Machines (SVMs) are employed. Among these, SVMs employing a quadratic kernel function demonstrate robust performance in categorizing the thermal and mechanical behavior of porous Silica.