About This Presentation
Presented by:
Dirk Van Essendelft – NETL
Summary. NETL has developed a novel stabilized, explicit solver that is accelerated with Machine Learning (ML) and is releasing it as part of the Carbonaceous Chemistry for Computational Modeling (C3M) software package. The solver uses several stabilization techniques and a new methodology we are calling “logically convergent, variable load” to ensure that the solver functions at maximum efficiency on GPU hardware. Each cell is solved in parallel and can evolve at its own time step and is removed from the integrator as it reaches the end time. The stabilization techniques adjust the time step for each cell to extend the explicit capability to stiffer problems. Further, the thermodynamic calculations were replaced by ML representations for a nearly 10% net time savings. We are exploring use of the Volta architecture for further acceleration on tensor cores. The solver was integrated in NETL’s MFiX and was demonstrated to be as much as 280 times faster depending on the circumstances and problem size relative to the current state of the art LSODA solver on CPU. C3M users will be able to access the TensorFlow protobuf files for the solvers which contain both the defined chemistry and the solvers for use in any compatible application. NETL is seeking to transfer this technology and test it in other CFD codes and believes that CPFD Barracuda could be an ideal candidate for integration
About the speaker. Dirk Van Essendelft obtained a BSE in Chemical Engineering from Calvin College in 2003, a MSE in Chemical and Biochemical Engineering from University of California in 2005, Irvine, and a Ph.D. in Energy and GeoEnvironmental Engineering from The Pennsylvania State University in 2008. He has work experience in the pharmaceuticals, carbon capture and storage, bio-energy conversion, gasification, and high-performance computing. His current research interests are in applied computational chemistry for large scale industrial modeling using super computers. He currently manages the development of the National Energy Technology Laboratory’s Carbonaceous Chemistry for Computational Modeling (C3M) software and is developing unique solutions that allow high fidelity modeling at low computational cost.