JuliaCon 2023

Massively parallel inverse modelling on GPUs with Enzyme
07-28, 11:30–12:00 (US/Eastern), 26-100

We present an efficient and scalable approach to inverse PDE-based modelling with the adjoint method. We use automatic differentiation (AD) with Enzyme to automaticaly generate the buidling blocks for the inverse solver. We utilize the efficient pseudo-transient iterative method to achieve performance that is
close to the hardware limit for both forward and adjont problems. We demonstrate close to optimal parallel efficiency on GPUs in series of benchmarks.

Massively parallel hardware architectures such as graphics processing units (GPUs) make it possible to run numerical simulations at unprecedented resolution. Matching the results of simulations to the available observational and experimental data requires estimating the sensitivity of the model to changes in its parameters. The adjoint method for computing sensitivities gains more attention in the scientific and engineering communities. This method allows for computing the sensitivities in the entire parameter space using the results of only one forward solve, in contrast to the direct method that would require runninng the forward simulation for each parameter separately.

Recent breakthrough innovations in compiler technologies resulted in wide adoption of the paradigm of differentiable programming, in which the entire computer program could be differentiated via automatic differentiation (AD). Differentiable programming allows for automated generation of adjoints from the forward model. In this work, we demonstrate the applications of the adjoint method to inverse modelling in geosciences, and share our experiences using the Julia ecosystem for high-performance computing. We developed massively parallel 3D forward and inverse solvers with full GPU support. We used Enzyme.jl for the AD implementation, ParallelStencil.jl to generate efficient computational kernels for multiple backends including GPUs, and ImplicitGlobalGrid.jl for distributed parallelism.

Co-authors: Ludovic Räss¹ ², Samuel Omlin ³

¹ ETH Zurich | ² Swiss Federal Institute for Forest, Snow and Landscape Research (WSL) | ³ Swiss National Supercomputing Centre (CSCS )

GPU computing and geo-HPC. Researcher at ETH Zurich, Switzerland.

This speaker also appears in:

Computational Scientist and Responsible for Julia computing at the Swiss National Supercomputing Centre, ETH Zurich

This speaker also appears in:

Ivan Utkin's scientific record revolves around numerical modelling of natural processes in geosciences with a strong emphasis on resolving coupled multi-physics interactions. Specifically, the record includes studies of fluid flows in porous rocks, including compaction-driven flow focussing, ground displacement due to the elastic response of rocks to the fluid pressure, and the influence of chemical interactions between fluid and rocks on the flow dynamics and observed chemical composition in rocks. A particular focus of Utkin's research is the development of numerical techniques to resolve processes in high resolution using massively parallel computing architectures such as graphics processing units (GPUs).

This speaker also appears in: