Python embedding in scientific computing to accelerate the development and execution of machine learning models
The influence and impact of machine learning (ML) in scientific computing (SC) is rapidly expanding. To harness the swift pace of algorithmic and hardware advances in ML, computational methods and tools are required to seamlessly integrate ML models into SC codes. The challenge to this seamless integration lies in the predominate use of Python within the ML community, while the SC community has traditionally developed code in Fortran or C/C++. The ease of use, robustness, active development communities of ML Python packages enable ML modelers to quickly and easily develop and train ML models using state-of-the-art ML architectures, optimizers, and hardware. In this work, we demonstrate how Python-based ML models can be incorporated directly into compiled SC code to leverage the benefits of both programing languages within large-scale SC simulations. We present these capabilities in the context of ML-based material models for computational structural mechanics. We show significant computational efficiency within robust, widely used finite element analysis codes by leveraging the native support for GPU accelerations in PyTorch.
IMPACT
This work is 6.1 research on computational approaches to integrating ML surrogate models in scientific computing to expand computational design capabilities across a wide range of DoD R&D efforts.
PRESENTER
Crone, Joshua
joshua.crone.civ@army.mil
520-691-5754
DEVCOM Army Research Lab
CATEGORY
Surrogate Modeling for HPC
SECONDARY CATEGORY
AI/ML for HPC
SYSTEM(S) USED
Nautilus, Narwhal