Conditionally Parameterized, Discretization-Aware Neural Networks for Mesh-Based Modeling of Physical Systems

Jiayang Xu, Aniruddhe Pradhan, Karthikeyan Duraisamy

Simulating complex physical systems—such as the aerodynamic airflow over a wing or the turbulent combustion inside a rocket engine—typically requires solving Partial Differential Equations (PDEs) on intricate, unstructured meshes. While deep learning has shown immense promise in accelerating these computationally heavy simulations, standard neural network architectures suffer from a fundamental "representation gap."

Common tools like Convolutional Neural Networks (CNNs) require data to be structured on perfectly square grids, which rarely exist in real-world engineering geometries. Furthermore, most standard neural networks treat diverse inputs (like temperature, pressure, and local mesh size) as a flat list of numbers, ignoring the hierarchical and mathematical relationships between these variables. For instance, in physics, a "flux" is often the product of a velocity and a gradient. A standard network must "brute-force" its way into learning that multiplication, whereas a physics-aware model should understand that relationship inherently.

To bridge this gap, researchers have developed a new class of Conditionally-Parameterized (CP) neural networks. By making the network "discretization-aware," their proposed CP-GNet can predict complex, unsteady reacting flows on irregular meshes with far greater accuracy and data efficiency than standard graph neural networks.

The Core Method: Conditional Parameterization

The authors introduce Conditional Parameterization (CP) to fundamentally change how network layers operate. In a standard neural network layer, the weights are fixed after the training phase. In a CP layer, the weights are dynamic; they are the output of another small, trainable function that takes physical parameters (like the Reynolds number or local mesh spacing) as its input.

Key technical components of this approach include:

  • CP-Dense Layers: Layers where weights are dynamically calculated based on a conditioning parameter vector.
  • CP-GNet Architecture: A Graph Neural Network (GNN) utilizing an Encoder-Processor-Decoder structure. It treats finite volume mesh cells as nodes and their physical connections as edges.
  • Discretization-Awareness: The network explicitly uses mesh topology features—such as cell volumes, face areas, and normal vectors—to condition the message-passing weights.
  • Ghost Edges: A novel technique to handle boundary conditions (like walls or fluid inlets) by creating virtual edges, allowing the GNN to learn how to "reflect" or "absorb" information at the domain's edges.
The complete CP-GNet architecture, illustrating the flow from finite volume discretization through the Encoder, Processor, and Decoder stages.

Experimental Validation

The researchers evaluated their approach across three distinct challenges of increasing complexity:

  1. 1D Viscous Burgers Equation: A classic benchmark for shock formation, used to test "closure modeling" (predicting missing physics when a simulation runs on a very coarse grid).
  2. Turbulent Channel Flow: Using wall-parallel snapshots of turbulent flow to perform "super-resolution," recovering high-detail flow structures from low-resolution inputs.
  3. Rocket Engine Injector: A highly complex, 2D finite-volume mesh dataset containing over 300,000 cells, involving chemical reactions and unsteady flow—a massive leap from standard "toy" problems.

Key Insights and Results

By using mesh size and physical parameters to generate the weights rather than just concatenating them as input features, the network acts like a local numerical operator. This closely mimics how traditional PDE solvers work. Physical equations are full of quadratic and high-order interactions; the authors demonstrated that CP layers can fit these terms exactly, whereas standard layers require significantly more neurons just to approximate them.

This architectural alignment with physics leads to remarkable data efficiency. Because the CP-GNet is "primed" with geometric intuition, it can be trained on very few data snapshots. By conditioning on mesh features, the model also becomes highly portable across different mesh densities and irregular geometries, rather than being locked to its training grid.

Results on the 1D Burgers equation showing the CP-CNN and CP-DDP models maintaining stability over time, while standard baselines fail or degrade.

The performance improvements are stark:

  • Numerical Stability: In the Burgers equation test, baseline "Data-Driven Parameterization" (DDP) models became numerically unstable (values spiked to infinity), while the CP-variants remained stable and accurate over time.
  • Turbulence Accuracy: For super-resolution tasks, the CP approach significantly reduced training and testing errors, capturing the energy spectra (the distribution of energy across different scales of motion) much more accurately than standard MLPs.
  • Reacting Flows: The CP-GNet successfully predicted the evolution of a reacting flow in the complex rocket injector over hundreds of time steps, maintaining high accuracy across multiple scales even near complex geometry boundaries where standard GNets typically fail.

Limitations and Future Work

While powerful, the CP approach does introduce architectural complexity. Using one network to generate weights for another increases the total number of trainable parameters, though the authors note this is partially offset by the ability to use smaller "latent" vectors. Additionally, dynamically calculating weights for every layer during the forward pass introduces computational overhead compared to using static weights.

Currently, the paper focuses on 1D and 2D cases. Scaling the CP-GNet to massive 3D industrial meshes will require further optimization of message-passing efficiency. Future work will also need to explore how well these conditionally parameterized models generalize to entirely different physical regimes, such as training on low-speed flows to predict high-speed aerodynamics.

Summary updated

Abstract

Simulations of complex physical systems are typically realized by discretizing partial differential equations (PDEs) on unstructured meshes. While neural networks have recently been explored for the surrogate and reduced order modeling of PDE solutions, they often ignore interactions or hierarchical relations between input features, and process them as concatenated mixtures. We generalize the idea of conditional parameterization -- using trainable functions of input parameters to generate the weights of a neural network, and extend them in a flexible way to encode critical information. Inspired by discretized numerical methods, choices of the parameters include physical quantities and mesh topology features. The functional relation between the modeled features and the parameters is built into the network architecture. The method is implemented on different networks and applied to frontier scientific machine learning tasks including the discovery of unmodeled physics, super-resolution of coarse fields, and the simulation of unsteady flows with chemical reactions. The results show that the conditionally-parameterized networks provide superior performance compared to their traditional counterparts. The CP-GNet - an architecture that can be trained on very few data snapshots - is proposed as the first deep learning model capable of standalone prediction of reacting flows on irregular meshes.
BibTeX
@inproceedings{xu2021conditionally,
  author = {Jiayang Xu and Aniruddhe Pradhan and Karthikeyan Duraisamy},
  title = {{Conditionally Parameterized, Discretization-Aware Neural Networks for Mesh-Based Modeling of Physical Systems}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2021},
  url = {https://openreview.net/forum?id=0yMGEUQKd2D},
}
Figure