Exascale simulation of cardiac electrophysiology to support arrhythmia research – Exacard

Heart disease is one of the two most important causes of death. About half of these deaths are caused by malfunctioning of the electrical activation system of the heart. In each cardiac muscle cell, millions of molecular-scale ion channels, pumps, and exchangers generate electric signals that serve to synchronize the contraction of the heart muscle. Pathologies that affect this intricate system can result in life-threatening arrhythmia: situations where the cardiac activation is disorganized and the contraction becomes ineffective.

Due to the complexity of the cardiac activation system, research in the area of cardiac electrophysiology has traditionally had a strong numerical component. Typical models of cardiac electrophysiology consist of a system of two reaction-diffusion equations that model the quasi-static electrical balance between the extra- and intra-cellular media, and ionic currents originating at the cell membrane. They are coupled to stiff and highly nonlinear systems of ordinary differential equations that model the dynamics of the cell membrane. Currently, the more realistic whole-heart or heart-body models have 10 million to 100 million mesh elements and 10 to 30 variables per element, and require high-performance computing (HPC) systems to be practically usable.

Even in these models each element still represents hundreds of cardiac muscle cells. This resolution suffices to simulate a structurally normal tissue, but there is a growing consensus that structural abnormalities play a crucial role in nearly all cardiac arrhythmia. Initiation of arrhythmia relies on structural details at the cellular and sub-cellular scale, and tissues that are likely to contain such "arrhythmogenic" details may generate characteristic features in body-scale measurable signals such as the electrocardiogram. For a realistic representation of these events the element size must be reduced from the customary 100 or 200um to less than 10um.

This scale leap cannot be achieved by merely deploying existing codes on the next generation of HPC infrastructure. The increased problem size and spatial resolution have many consequences for the available choices and numerical properties of the solvers. Moreover, the heterogeneity in the new generation of supercomputers is increasing, integrating a variety of multicore CPUs and accelerators such as graphics processing units. This heterogeneity requires a flexible load management within the code. In addition, output of results is already a bottleneck encountered in simulations, and will certainly become more difficult as the scale of the simulations grows.

Our objective is to look for solutions to reach the desired 10um spatial resolution. To this aim we plan to combine advances numerical strategies and a new task-based parallelisation scheme. We are two research teams specialists in these fields, that will tightly collaborate to reach this goal. We will base our research on a highly mature HPC code for cardiac electrophysiology and a modern runtime system. The project will require to develop new numerical methods with low communication requirements, as well as innovative approaches in runtime systems. Finally, we want to maintain the practical usability of the code, meaning that it can be used for applied heart modeling studies aimed at improving understanding and diagnosis of cardiac diseases. Therefore we have brought together a consortium with an overlapping expertise in parallel computing, runtime systems, large-scale scientific computing, numerical methods, and applied cardiac simulation studies. A new code will ultimately be available that allows to reach a 1000-fold increase in problem size.

We expect this project to have an impact on 1) cardiac simulation, which will become much more realistic especially when structurally abnormal cardiac tissue is simulated, 2) numerical methods for reaction-diffusion problems, and 3) the development and practical usability of runtime systems.