darcy_physics_informed

Physics Guided models for Darcy flow

This example demonstrates physics informing of a data-driven model using to approaces - the DeepONet approach which computes the gradients using Autograd and an approach using Numerical derivatives (PINO).

Problem overview

This is an extension of the 2D Darcy flow data-driven problem. In addition to the data loss, we will demonstrate the use of physics constraints, specifically the equation residual loss. The physicsnemo.sym module (install with pip install "nvidia-physicsnemo[sym]") has utilities tailored for physics-informed machine learning, and we leverage them here to add physics to an existing data-driven model with ease while keeping a fully explicit training loop.

If you previously used the (now archived) physicsnemo-sym repository, where the Solver / Domain / Constraint abstractions handled the training loop and physics losses implicitly, see the PhysicsNeMo v2.0 Migration Guide for how the equivalent pieces look in this newer, explicit style.

Dataset

The training and validation datasets for this example can be found on the Fourier Neural Operator Github page. The downloading and pre-processing of the data can also be done by running the below set of commands:

pip install -r requirements.txt
python download_data.py

Do demonstrate the usefulness of the Physics loss, we will deliberately choose a smaller dataset size of 100 samples. In such regiemes, the effect of physics loss is more evident, as it regularizes the model in the absence of large data.

Model overview and architecture

In this example we will use a Fourier Neural Operator (FNO). We will demonstrate two cases, in the first case, the FNO is used as the branch net and use a fully-connected network for the trunk net. The input to the branch network is the input permeability field and the input to the trunk network is the x, y coordinates. The output of the model is the pressure field. Having the mapping between the pressure field and the input x and y through a fully-differentiable network will allow us to compute the gradients of the pressure field w.r.t input x and y through automatic differentiation through PhysicsNeMo sym utils.

In the second case, we will use just FNO and then compute the derivatives in a PINO style, using Numerical differentiation. Both approaches are viable ways to introduce physics in the loss function and the use of one over the other can change from case-to-case basis. With this example, we intend to demonstrate both such cases so that the users can compare and contrast the two approaches.

In this example we will use the PDE class from physicsnemo.sym to symbolically define the PDEs and use the PhysicsInformer utility to introduce the PDE constraints. Defining the PDEs sympolically is very convenient and most natural way to define these PDEs and allows us to print the equations to check for correctness. This also abstracts out the complexity of converting the equation into a pytorch representation. physicsnemo.sym also provides several complex, well tested PDEs like 3D Navier-Stokes, Linear elasticity, Electromagnetics, etc. pre-defined which can be used directly in physics-informing applications.

Getting Started

To get started with the example, simply run,

python
darcy_physics_informed_deeponet.py

or

python
darcy_physics_informed_fno.py

Note

If you are running this example outside of the PhysicsNeMo container, install PhysicsNeMo with the sym extra: pip install "nvidia-physicsnemo[sym]"

References

• Fourier Neural Operator for Parametric Partial Differential Equations
• Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators
• Physics-Informed Neural Operator for Learning Partial Differential Equations