Session

DescriptionSimulations of physical phenomena consume a large fraction of computer time on existing supercomputing resources. Today, the challenge of scaling multiscale simulations is primarily addressed by brute-force search-and-sample techniques and are computationally expensive. Emerging Exascale architectures pose challenges for simulations such as efficient and scalable execution of complex workflows, concurrent execution of heterogeneous tasks and the robustness of algorithms on millions of processing cores, data and I/O parallelism, and fault tolerance. Therefore, incremental approaches that scale simulations will not be successful in achieving the throughput and utilization on such machines. Machine learning (ML) techniques can be integrated with system and application changes and give many orders of magnitude higher effective performance. We term this convergence of high-performance computing (HPC) and ML methodologies/ practice as MLforHPC. Nowhere is the impact of MLforHPC methods likely to be greater than multiscale simulations in biological and material sciences, with early evidence suggesting several orders of magnitude improvement over traditional methods. Fueled by advances in statistical algorithms and runtime systems, ensemble-based methods have overcome some of the limitations of traditional monolithic simulations. Furthermore, integrating ML approaches with such ensemble methods holds even greater promise in overcoming performance barriers and enabling simulations of complex multiscale phenomena.