Envisioning the Role of Physics-Informed Neural Networks in Atmospheric Science: Advancements, Challenges, and Future Prospects

Since the inception of physics-informed neural networks (PINNs) by Raissi et al. in 2019, it has been seen as a promising approach to outperform conventional algorithms in terms of computational efficiency, reduced costs, and improved prediction accuracy, especially in small data regimes.PINNs incorporate known physical governing equations in the form of partial differential equations (PDEs) or ordinary differential equations (ODEs) into neural networks, and occasionally the governing equations are derived from observational or simulated data, allowing PINNs to address specific atmospheric systems.Moreover, depending on the problem being solved, most work adds the physical constraints directly into the loss or cost function, while others enhance performance using modified architectures or preprocessing techniques.In the realm of atmospheric sciences, challenges remain, including a heavy reliance on simulated data and limited use of observational datasets, which does not show the real-world applicability of PINNs. A detailed review of available results shows critical gaps in scalability, hybrid data integration, and standardization in atmospheric science.We identified a hybrid methodology by combining simulated and observational data, which includes optimizing hybrid loss functions to balance physics-based and observational accuracy, applying adaptive training techniques, and standardizing preprocessing schemes to handle multi-scale atmospheric phenomena.Results demonstrate the ability of PINNs to deliver faster computation, enhanced prediction accuracy, and robustness in sparse data environments. This highlights the transformative advantages of PINNs over traditional methods and suggests future directions for leveraging their capabilities in atmospheric science applications.