A Neural-Operator Surrogate for Platelet Deformation Across Capillary Numbers.

Area of Science:

• Computational biology
• Biophysics
• Scientific machine learning

Background:

• Multiscale modeling of thrombosis requires accurate platelet-scale simulations, which are computationally expensive.
• Existing models struggle to balance fidelity with computational cost, hindering organ-scale predictions.

Purpose of the Study:

• To develop a highly accurate and computationally efficient surrogate model for platelet dynamics using scientific machine learning.
• To bridge the gap between platelet-scale fidelity and organ-scale computational requirements in thrombosis modeling.

Main Methods:

• A DeepONet surrogate model was trained on platelet dynamics data generated using LAMMPS simulations.
• The model input includes wall shear stress, bond stiffness, time, and initial particle coordinates.
• Adam optimization with adaptive learning-rate decay was used for training.

Main Results:

• The DeepONet surrogate achieved median displacement errors below 1% and worst-case errors below 4% across a range of elastic moduli and capillary numbers.
• Computational speedup ranged from four to five orders of magnitude compared to traditional methods.
• The model demonstrated acceptable extrapolation capabilities for stiff and compliant platelets.

Conclusions:

• Scientific machine learning, specifically DeepONet, offers a viable solution for accurate and efficient platelet dynamics simulation.
• The developed surrogate model can be coupled with continuum computational fluid dynamics (CFD) for future platelet-resolved hemodynamic simulations.
• This approach opens new avenues for predictive thrombosis modeling in patient-specific geometries.