Machine learned domain decomposition scheme applied to parallel multi-scale muscle simulation

Abstract

Since multi-scale models of muscles rely on the integration of physical and biochemical properties across multiple length and time scales, they are highly processor and memory intensive. Consequently, their practical implementation and usage in real-world applications is limited by high computational requirements. There are various reported solutions to the problem of parallel computation of various multi-scale models, but due to their inherent complexity, load balancing remains a challenging task. In this article, we present a novel load balancing method for multi-scale simulations based on finite element (FE) method. The method employs a computationally simple single-scale model and machine learning in order to predict computational weights of the integration points within a complex multi-scale model. Employing the obtained weights, it is possible to improve the domain decomposition prior to the complex multi-scale simulation run and consequently reduce computation time. The method is applied to a two-scale muscle model, where the FE on macroscale is coupled with Huxley’s model of cross-bridge kinetics on the microscale. Our massive parallel solution is based on the static domain decomposition policy and operates in a heterogeneous (central processing units + graphics processing units) environment. The approach has been verified on a real-world example of the human tongue, showing high utilization of all processors and ensuring high scalability, owing to the proposed load balancing scheme. The performance analysis shows that the inclusion of the prediction of the computational weights reduces execution time by about 40% compared to the run which uses a trivial load balancer which assumes identical computational weights of all micro-models. The proposed domain decomposition approach possesses a high capability to be applied in a variety of multi-scale models based on the FE method.