The dynamic behavior of physical systems is often described by conservation and constitutive laws, expressed as systems of partial differential equations (PDEs). A classical task involves the use of computational tools to solve such equations across a range of scenarios, e.g., different domain geometries, input parameters, and initial and boundary conditions. Solving these so-called parametric PDEs using traditional tools (e.g., finite element methods) comes with an enormous computational cost, as independent simulations need to be performed for every different domain geometries or input parameter.

In the noteworthy paper [Wan21L], the authors propose
the framework of *physics-informed DeepONets* for solving parameterized PDEs. It
is a simple yet remarkably effective extension of the DeepONet framework
[Lu21L] (cf. Learning nonlinear operators: the DeepONet architecture). By constraining the
outputs of a DeepONet to approximately satisfy an underlying governing law,
substantial improvements in predictive accuracy, enhanced generalization
performance even for out-of-distribution prediction, as well as enhanced data
efficiency can be observed.

Operator learning techniques have demonstrated early promise across a range of applications, but their application to solving parametric PDEs faces two fundamental challenges: First, they require a large corpus of paired input-output observations and, second, their predicted output functions are not guaranteed to satisfy the underlying PDE. Motivated by the fact that the outputs of a DeepONet model are differentiable with respect to their input coordinates, one can use automatic differentiation to formulate an appropriate regularization mechanism in the spirit of physics-informed neural networks [Rai19P]. The target output functions of the DeepONet are biased to satisfy the underlying PDE constraints by incorporating these into the loss function of the network.

When the collection of all trainable weights of a DeepONet is denoted by $\theta$, the network is optimized with respect to the loss function $$ \mathcal{L}(\theta) = \mathcal{L}{\small\text{operator}}(\theta) + \mathcal{L}{\small\text{physics}}(\theta) $$ where $\mathcal{L}{\small\text{operator}}$ fits the available solution measurements and $\mathcal{L}{\small\text{physics}}$ enforces the underlying PDE constraints.

The authors demonstrate the effectiveness of the physics-informed DeepONets in solving parametric ordinary differential equations, diffusion-reaction and Burgers’ transport dynamics, as well as advection and Eikonal equations. In the diffusion-reaction example, the physics-informed DeepONet yields ∼80% improvement in prediction accuracy with 100% reduction in the dataset size required for training. For Burgers’ equation, notably, a trained physics-informed DeepONet model can predict the solution up to three orders of magnitude (1000x) faster compared to a conventional solver.

Despite the promise of the framework demonstrated in the paper, the authors acknowledge that many technical questions remain. For instance, for a given parametric governing law: What is the optimal feature embedding or network architecture of a physics-informed DeepONet? Addressing these questions might not only enhance the performance of physics-informed DeepONets, but also introduce a paradigm shift in how we model and simulate complex physical systems.