Learned closure of the five-moment kinetic system

Learn a state-dependent regularisation of the five-moment closure of the 1D BGK equation from kinetic data, and compare the closed moment system with a reference BGK solution (closure error, hyperbolicity, explicit time-step restriction)

Quick Facts

Type

Demonstrator

Status

Planned

Work Packages

WP2

Frameworks

PyTorch

1. Overview

Learn a state-dependent regularisation of the five-moment closure of the 1D BGK equation from kinetic data, and compare the closed moment system with a reference BGK solution (closure error, hyperbolicity, explicit time-step restriction)

1.1. Repository

private, access on request (remy.hosseinkhan@polytechnique.edu)

2. Technical Stack

2.1. Frameworks Used

  • PyTorch

2.2. Parallel Frameworks

  • Sequential

3. Methods & Algorithms

3.1. WP1: Discretization

  • finite volume

3.2. WP2: Model Order Reduction & SciML

  • NN closure of a moment model

4. Data Flow

4.1. Inputs

  • in-house (BGK reference data)

4.2. Outputs

  • JSON reports

  • PyTorch checkpoints

5. Benchmarking

5.1. Metrics

  • Benchmark Verification

  • Time To Solution

5.2. Benchmark Scope

  • Single Node

6. Timeline

Milestone Date

Specification Due

N/A

Prototype Due

N/A

7. Team

Partners: IPP

Responsible: M. Massot

8. Notes

Contributors: R. Hosseinkhan-Boucher, M. Massot. Validation: error against the reference BGK solution, real-spectrum test at evaluation states. Planned: execution time of the closed solver.