Researchers have developed Extended Physics-Informed Neural Networks (Ex-PINNs), a new machine learning method designed to solve partial differential equations (PDEs) in domains with spatially heterogeneous parameters. Ex-PINNs address a key limitation of traditional PINNs, which struggle to handle the spatial variability of physical properties such as conductivity or permeability. This advancement is crucial for modeling complex systems in fields like geophysics, materials science, and biology, where medium properties vary significantly from point to point.

The Ex-PINN method introduces a neural network architecture that can simultaneously learn the PDE solution and the spatial distribution of heterogeneous parameters. This is achieved by incorporating additional basis functions or auxiliary neural networks that represent the parameter variation within the domain. By training the network with observational data and the physical equations themselves, Ex-PINNs can infer hidden system properties and predict their behavior with greater accuracy than previous approaches.

The ability of Ex-PINNs to handle spatial heterogeneity is a significant step towards more realistic simulation of natural and engineering phenomena. Standard PINNs often assume uniform or known parameters, which restricts their applicability to many real-world problems. By overcoming this barrier, Ex-PINNs open new avenues for characterizing complex media, optimizing material designs, and enhancing the understanding of dynamic systems where local properties play a fundamental role.