/PaML_PiML

PiML is a widely used approaches to incorporate physical constraints, which can be trained from additional information obtained by enforcing the physical laws (for example, designing loss functions (regularization))

Physics-informed Machine Learning

In scientific computing, physical phenomena are often described using a strong mathematical form consisting of governing differential equations as well as initial and boundary conditions. PiML is a widely-used approach that integrates physical laws into ML models (for example, designing loss functions or regularization), facilitates the accurate capture of dynamic patterns and concomitantly diminishes the search space for model parameters. This approach is sometimes referred to as imposing differentiable constraints in loss functions. It integrates (noisy) data and mathematical models, and implements them through neural networks (physics-informed neural networks) or kernel-based neural operators (physics-informed neural operators).

Challenges of vanilla PINNs

(a) Challenges of vanilla PINNs. (b) An example of vanilla PINNs failing to converge on high-frequency and multi-scale PDEs. 1-D convection equations with high-frequency (first row): Train and infer on a  spatial-temporary resolution 256 \times 100; 2-D steady incompressible Navier-Stokes equations (second row): Train and infer on a  spatial resolution 49 \times 77. Analytical solution and computational fluid dynamics (CFD) represent ground truth.

Model Zoo

Physics-informed Neural Networks

image

Physics-informed Neural Operators

image
Continuously updating
More information can be found in the PAPER