Machine Learning Accelerated Lattice Boltzmann Simulation for Multiscale Platelet Modeling

<b>Background</b>: Cardiovascular disease (CVD) is responsible for upwards of 17.9 million deaths globally per year. CVD-related death commonly occurs on account of platelet-mediated thrombosis. Modeling platelet dynamics is necessary to develop a mechanistic understanding of platelet-mediated thrombosis at multiple scales. Multiscale modeling (MSM) links microscale, mesoscale, and macroscale simulations to speed up computation while maintaining desired accuracy. Previous simulations model platelet aggregation using a MSM model that uses dissipative particle dynamics (DPD) for the shear flow and coarse-grained molecular dynamics (CGMD) for the platelet dynamics. However, the resolution of the fluid particles in the DPD simulation occupies unnecessary computational resources. In contrast, the lattice-Boltzmann method (LBM) may be up to 20-50x faster than DPD while still resolving the relevant fluid phenomena. In this work, we present a machine learning (ML) accelerated LBM for platelet dynamics.<br/><br/><b>Methods</b>: We evaluate two methods for LBM: the conventional LB algorithm and the ML-LBM model. The conventional LBM solves the discrete Boltzmann equation for velocity distribution functions to predict flow dynamics. An in-silico model of the system was developed for LAMMPs using CGMD for platelets and LBM for fluid dynamics. We initially predicted future dynamics with this framework. The ML-LBM model trains based on existing data to accelerate computation. First, input data is preprocessed by normalization into a series of tensors indicating velocity distributions. The proposed architecture consists of an encoding layer, utilizing a ConvLSTM architecture to effectively resolve system dynamics in both space and time, and a decoding layer, utilizing the ResNet CNN architecture to map these inputs back to the original space. The training process minimizes the Mean Square Error (MSE) loss to iteratively improve predictive accuracy. The model outputs a series of tensors representing the final velocity distribution based on the LB dynamics. Our workflow exports data from LAMMPs, predicts the flow dynamics with the ML model, and then sends these predicted velocity distributions back into LAMMPs. Accuracy was computed by comparison of flow velocity and platelet stresses from our model to CGMD-DPD simulations. Simulations were performed with a total of 4 aggregated platelets and 3 flowing platelets within a box of 16x16x39.1 micrometers, for a total of 1,148,441 particles in the simulation.<br/><br/><b>Results</b>: The conventional LB algorithm was 9x faster than DPD while maintaining 96.4% and 96.2% accuracy for flow velocity and platelet stress respectively. Evaluation of the ML-LBMyielded a 26x speedup with a 96.1% and 95.8% accuracy for flow velocity and platelet stress respectively. Both algorithms were able to capture major flow dynamics and platelet stresses while evaluating significantly faster than DPD. We attribute the lower accuracy of the machine learning LB model to the inability to effectively learn the dynamics of the platelet-fluid boundary, where lowest accuracy was observed.<br/><br/><b>Discussion and Future Work</b>: We assessed the relative performance of LB methods in recreating the dynamics of DPD. Both the conventional LB algorithm and the novel ML-LBM captured the system dynamics of blood flow and platelet stresses. Due to the inherent limitations of LBM, both models used in this study were less accurate near the platelet-fluid interfaces, since LBM cells are much larger than CGMD particles with the added resolution. This added resolution, albeit rather costly, has not caused errors in the platelet dynamics. Future work should implement and assess the efficiency of the ML-LBM method for studying larger and more complex platelet clusters for platelet dynamics modeling.<br/><br/><b>Acknowledgements</b>: The project is supported by the Garcia High School Program and the Morin Charitable Trust.