Revitalizing Simulations of Colloidal and Interfacial Systems: From Atomistic Trajectories to Data-Centric Insights

Abstract

This dissertation advances methodological frameworks for investigating complex systems characterized by competing interactions. Structured around three interrelated themes, the research develops methods to tackles challenges in water-in-oil emulsions, chemical inhibition in asphaltene aggregation, and adsorption energy prediction. While each theme addresses unique challenges, they all contribute to quantitative understanding of interactions across scales, integrating structural, thermodynamic, and energetic perspectives to establish governing principles for system behavior. The first theme elucidates the interplay between asphaltene aggregation and water-in-oil droplet coalescence. Molecular Dynamics (MD) simulations conducted in pentane solvents reveal the mutual influences of asphaltene and water droplets, highlighting a nonmonotonic trend in polyaromatic stacking with increasing droplet sizes. An innovative in-house tool is introduced to analyze droplet coalescence modes, demonstrating that droplet growth predominantly occurs around a nucleation site, which is the largest droplet. In the second theme, a novel approach for calculating partial molar volumes (PMVs) directly from MD simulation trajectories is presented. This approach has been validated against experimental data, yielding an average error of 4.41%. When applied to systems containing model asphaltenes, organic solvents, and chemical inhibitors, the PMV analysis elucidates molecular-level inhibition mechanisms. Specifically, it shows that inhibitors enhance solubility by altering nanoaggregate size and number, with their polar and nonpolar segments interacting with different regions of the asphaltene molecules and the solvent. The third and final theme highlights the critical need for efficient and accurate adsorption energy predictions, a fundamental aspect of catalysis, materials design, and a key factor in mitigating asphaltene deposition. Traditional methods of calculating adsorption energy such as Density Functional Theory (DFT) are computationally demanding, particularly for large, complex adsorbates like aromatics. To overcome these limitations, this work applies pretrained graph neural networks (GNNs) with a directional message passing architecture. The model is trained to capture geometric relationships between small adsorbates and metallic or metal oxide surfaces, linking them to energy and atomic forces. Here, it is fine-tuned to adapt these learned interactions for larger molecules, including aromatics. Two curated datasets, a diverse adsorbate-substrate collection and a specialized aromatic subset, are employed to balance generalizability with specificity. Results indicate that the sheer volume of fine-tuning data has a greater impact on model adaptation than using smaller but more domain-specific datasets. Moreover, the results show that selectively fine-tuning the early layers, which focus on geometric features, achieves performance comparable to full model retraining. This highlights the crucial role of geometric feature adaptation. Collectively, these themes contribute to the refinement of data-driven methodologies for complex interfacial and amorphous systems, providing actionable insights into structural, thermodynamic, and energetic properties that were previously inaccessible. Nano-scale simulations with extensive temporal and spatial sampling are essential, as competing interactions make it impossible to predict dominant factors without comprehensive analysis. By reutilizing existing simulation data, the innovations presented here enhance sustainability and efficiency, with broad implications for physical chemistry, materials science, and industrial applications.