Solving Riemann Problems with Physics-Informed Neural Networks: An Analysis of the riemaNN Project

Solving Riemann Problems with Physics-Informed Neural Networks: An Analysis of the riemaNN Project

riemaNN is an open-source project developed by gusbeane (released on GitHub on May 26, 2026). Its core is using Physics-Informed Neural Networks (PINN) to solve classic Riemann problems in Computational Fluid Dynamics (CFD) and astrophysics. The project directly predicts star region pressure via neural networks, offering an efficient alternative to traditional numerical methods. Its tech stack is based on JAX, Flax, and optax, supporting GPU/TPU acceleration, with a modular architecture and differentiable features.

Background: Definition of Riemann Problems and Challenges of Traditional Methods

Riemann problems describe phenomena like shock waves and rarefaction waves generated by the interaction of fluids in two different states, and are core problems in CFD and astrophysics. They need to be solved in large numbers in simulations of stellar evolution, supernova explosions, etc. Traditional methods require iterating nonlinear equations, which have high computational costs—especially in large-scale parallel simulations, where millions of solutions may be needed per time step.

Overview of the riemaNN Project and Its Tech Stack

Project goal: Predict the normalized star region pressure p*/p_ref from 3D gas state differences (drho, dp, du). The tech stack includes Python3.12, JAX (high-performance computing/automatic differentiation), Flax (neural network definition), and optax (optimization library). Modular architecture:

• physics.py: Defines physical formulas like gas constants and critical velocities
• data.py: Uniform/quasi-random data sampling
• model.py: MLP model that takes state differences as input and outputs pressure in log space
• train.py: Multi-stage training pipeline (supports checkpoints)
• plot.py: Visualization of loss curves, error distributions, etc.

Core Mechanism: Application of Physics-Informed Neural Networks

riemaNN embeds physical constraints into training:

1. Input normalization: Symmetrically process density difference (drho=(rhoR-rhoL)/(rhoR+rhoL)), pressure difference (dp=(pR-pL)/(pR+pL)), and velocity difference (du=uRL/ducrit) to ensure numerical stability
2. Output positivity constraint: Exponential transformation (p*/p_ref=10^model(x)) to ensure pressure is positive
3. Loss function: Minimize the squared residual of the Riemann jump function f(p*) so that the solution satisfies physical conservation laws

Multi-Stage Training Strategy and Practical Value

Training strategy: Progressive difficulty increase in stages, supporting checkpoint-based resumption and combination of multi-stage results. Practical significance:

• Efficiency improvement: After training, a single forward pass solves the problem, which is much faster than traditional methods
• Differentiable simulation: Based on JAX, it supports automatic differentiation and is suitable for solving inverse problems
• Methodology generalization: Combining physical priors with data-driven approaches can be applied to fields like materials science and climate modeling

Project Usage and Summary Insights

Usage method: The command-line toolchain supports experiment execution (venv/bin/python run.py experiments/smoke_test.py), retraining (--retrain), batch runs, and report generation.

Summary: riemaNN represents an important direction in scientific machine learning, demonstrating the potential of combining physical constraints with deep learning. It provides a reference template for computational physics developers, and PINN will drive scientific computing into a new era in the future.