Automated Solver Selection for Nuclear Engineering Simulations

An important objective of the NEAMS program is to enable widespread use of the software tools among the industry, academia, and regulatory communities. For solving the problems occurring at various stages of NEAMS simulations, typically there are several possible choices for numerical subroutines. Furthermore, the best method for a numerical problem may also evolve over the course of the simulation. The choice of the method can significantly enhance the portability of the NEAMS tools across a wide range of NEAMS user base and computing platforms. General Statement RNET and its subcontractors will develop an addin feature called “SolverSelector” for performance optimization of NEAMS simulations in terms of CPU time, accuracy, resilience, and energy efficiency. The plugin will analyze the sub problem characteristics at runtime and select the optimal solver with minimal overhead based on previously trained machine learning odels. Phase I Work: The Phase I has investigated robust machine learning models and feature sets using the data generated from NEAMS applications and other standard datasets (Florida Sparse Matrix collection). High prediction accuracy has been demonstrated for optimal linear solver selection in terms of execution time. Phase II Work: The Phase II work will investigate enabling solver selection throughout the spectrum of NEAMS tools, investigate other performance objectives and their prioritization, and integrate the techniques into a pluggable software component. In addition, software indirections will be developed to adapt the techniques to other numerical softwares. Commercial Applications and Other Benefits: The NEAMS users will be benefitted in the form of faster simulations, portability across computing
platforms, and increased reliability by avoidance of spurious failures due to nonconvergence issues. The project is beneficial is also beneficial to NEAMS developers by avoiding lot of experimentation required in choosing default solvers. The end product is applicable to range of numerical softwares across the government institutions and private companies. The targeted customers include power companies, DoE agencies, and NASA divisions, DoD and its Prime Contractors, CFD software providers, oil and gas companies, and semiconductor design companies. Key Words: Automatic Solver Selection, Machine Learning, Feature Analysis