Amorphous silica (a-SiO₂) plays a vital role in various fields, including catalysis and gas adsorption. Owing to its broad pore size distribution—spanning both micro- and mesoporous regimes—it provides greater accessibility to reactive sites compared to crystalline analogs such as zeolites. In addition, undercoordinated surface defects exhibit high reactivity and enhanced gas adsorption, making them suitable anchoring sites for transition metals.

Despite these advantages, the intrinsic structural disorder of a-SiO₂ presents major challenges for conventional atomistic analysis. The high variability of local atomic environments (LAEs) necessitates a systematic characterization and accurate prediction of their properties to support the rational design of amorphous materials.

To address this, we propose a multiscale machine learning (ML) framework that combines classical Molecular Dynamics (MD) and Density Functional Theory (DFT) to predict adsorption and catalytic properties across statistically representative a-SiO₂ surfaces. Classical MD is used to efficiently generate diverse amorphous surface models, while high-throughput DFT calculations enable accurate evaluation of adsorption energies, transition-state energies, and reaction rates at chemically relevant sites. Unsupervised and probabilistic ML models are then applied to link structure and function across scales.