Machine-learning-based models in particle-in-cell codes for advanced physics extensions

Author of this repository: Óscar Amaro (Feb 2024)

Authors of paper: Chiara Badiali, Pablo J. Bilbao, Fábio Cruz, Luis O. Silva

Link to paper: https://www.cambridge.org/core/journals/journal-of-plasma-physics/article/abs/machinelearningbased-models-in-particleincell-codes-for-advanced-physics-extensions/9D34BB83508AF220EC60EF892079D053

Pre-print: https://arxiv.org/abs/2206.02937

Abstract: In this paper we propose a methodology for the efficient implementation of Machine Learning (ML)-based methods in particle-in-cell (PIC) codes, with a focus on Monte-Carlo or statistical extensions to the PIC algorithm. The presented approach allows for neural networks to be developed in a Python environment, where advanced ML tools are readily available to proficiently train and test them. Those models are then efficiently deployed within highly-scalable and fully parallelized PIC simulations during runtime. We demonstrate this methodology with a proof-of-concept implementation within the PIC code OSIRIS, where a fully-connected neural network is used to replace a section of a Compton scattering module. We demonstrate that the ML-based method reproduces the results obtained with the conventional method and achieves better computational performance. These results offer a promising avenue for future applications of ML-based methods in PIC, particularly for physics extensions where an ML-based approach can provide a higher performance increase.

Notes:

• OSIRIS simulations shown here were run with the Compton-ML branch described in this paper. To know more about the Osiris Consortium see here

• OSIRIS simulation data shown in the notebooks can be downloaded here (~240Mb compressed, ~920Mb uncompressed)