On the computational acceleration of aggregating nanoparticle transport models using deep operator networks: Correlation with experimental data

Accurate modeling of nanoparticle (NP) transport in porous media requires taking into account several processes and mechanisms, including aggregation, reversible and irreversible attachment. These models are computationally intensive, especially when the participating processes evolve on different time scales. Consequently, their use in large scale simulations and fitting applications is rather limiting. In this work, a Deep Operator Networks (DeepONets) model is proposed that can learn the complete space-time evolution of NP concentrations, incorporating not only traditional transport mechanisms such as advection and dispersion, but also more complex processes like aggregation. The DeepONet is trained on a dataset obtained from numerical simulations produced by a nanoparticle transport model reported in the literature, achieving speedups of up to five orders of magnitude compared with conventional simulations. A detailed analysis of this dataset revealed that aggregation significantly influences plume dynamics, with larger aggregates leading to higher plume velocities and greater plume spreading. The trained DeepONets model was subsequently integrated into a Tandem Neural Network Architecture (TNNA), enabling the direct inference of transport coefficients. These inferred parameters were then used as high quality initial estimates that significantly accelerated the fitting of experimental breakthrough curves by 33%. These results establish the TNNA and DeepONets as valuable tools for simulating and predicting the fate of aggregating NP transport in porous systems.