2025-10-02 –, Jean-Baptiste Say Amphitheater
Language: English
Computational Earth Science relies on modelling to explore complex physical systems that are difficult to observe directly and where data is limited. This is especially crucial in geodynamics and ice sheet dynamics, both of which require large-scale simulations on next-generation HPC systems. These simulations benefit significantly from automated data-driven model constraints, enhancing accuracy and predictive capability.
We present our recent advancements in large-scale differentiable modelling for high-performance geophysical simulations, focusing on Earth’s largest ice sheets and high-resolution 3D geodynamic models of magmatic systems and mountain formation.
Our goal is to develop differentiable multi-physics solvers for extreme-scale geophysics simulations, leveraging GPU acceleration and automatic differentiation (AD) with Julia. To achieve this, we integrate Enzyme.jl, an advanced AD tool, enabling highly efficient reverse-mode differentiation at near-theoretical peak performance. This extends accelerated pseudo-transient solvers with differentiable modelling capabilities, significantly enhancing computational efficiency.
A key aspect of this effort is the enhancement of the GPU4GEO software stack—a modular ecosystem of Julia-based solvers designed for geophysical simulations. These solvers provide customisable building blocks and optimised subroutines, facilitating rapid development and high-performance execution across multi-core CPUs and AMD/NVIDIA GPUs. By integrating Enzyme.jl, we enable high-performance AD on GPUs, ensuring scalability through MPI-based distributed parallelism on next-generation supercomputers. We will discuss latest advance and report about scalability and simulation results on the ALPS and LUMI supercomputer at CSCS and CSC, respectively.
Computational geoscientists with Earth science background. Julia GPU and HPC enthusiast.