Neural Operator Modeling of Platelet Geometry and Stress in Shear Flow

Thrombosis involves processes spanning large-scale fluid flow to sub-cellular events such as platelet activation. Traditional CFD approaches often treat blood as a continuum, which can limit their ability to capture these microscale phenomena. In this paper, we introduce a neural operator-based surrogate model to bridge this gap. Our approach employs DeepONet, trained on high-fidelity particle dynamics simulations performed in LAMMPS under a single shear flow condition. The model predicts both platelet membrane deformation and accumulated stress over time, achieving a mode error of ~0.3% under larger spatial filtering radii. At finer scales, the error increases, suggesting a correlation between the DeepONet architecture's capacity and the spatial resolution it can accurately learn. These findings highlight the importance of refining the trunk network to capture localized discontinuities in stress data. Potential strategies include using deeper trunk nets or alternative architectures optimized for graph-structured meshes, further improving accuracy for high-frequency features. Overall, the results demonstrate the promise of neural operator-based surrogates for multi-scale platelet modeling. By reducing computational overhead while preserving accuracy, our framework can serve as a critical component in future simulations of thrombosis and other micro-macro fluid-structure problems.