Physics-embedded graph neural operator for interaction-controlled colloidal aggregation.

Area of Science:

• Environmental science
• Water chemistry
• Colloid science

Background:

• Colloidal aggregation in water is complex, influenced by particle and water chemistry.
• Predicting aggregation across different regimes is difficult due to nonlinear dependencies.

Purpose of the Study:

• Develop a graph neural operator surrogate model to predict colloidal aggregation dynamics.
• Embed transport physics directly into the network for improved accuracy.

Main Methods:

• Represent particle size classes as graph nodes with encoded Brownian collision kernels.
• Utilize attention mechanisms conditioned on ionic strength and zeta potential to model collision efficiency.
• Incorporate extended DLVO theory for electrostatic interactions.

Main Results:

• Achieved R² > 0.99 for predicting aggregation kinetics across various conditions, including temporal extrapolation.
• Outperformed baseline and physics-regularized neural network models.
• Validated against experimental data from bacteriophage, polystyrene, and cerium oxide systems, reproducing aggregation regime transitions.

Conclusions:

• The graph neural operator model provides a physically grounded framework for exploring colloidal aggregation.
• Enables rapid analysis relevant to water quality prediction and treatment optimization.
• Demonstrates effective learning of collision efficiency variations based on electrochemical conditions.