Title: Utilizing Physics-Encoded Inverse Modeling to Predict Arctic Snow Depth

Abstract: Estimating variables in time-dependent inverse problems is a persistent hurdle in scientific research, particularly when data is sparse and limited. A prime example is determining snow depth, a task that necessitates deducing hidden parameters related to sea ice physics, a process that can be significantly enhanced through physics-informed encoding. To tackle this issue, we present Physics-Encoded Inversion (PhysE-Inv), a new framework designed to solve inverse problems in real-world scenarios characterized by sparse observations. This approach merges deep sequential learning with physics-informed inference, utilizing an LSTM encoder-decoder architecture to model temporal dependencies. Additionally, it employs contrastive learning regularization to ensure that latent representations remain invariant to noise. The system learns latent parameters that, when paired with observational data, facilitate the reconstruction of snow depth while adhering to physics-based guidance. Our results show that PhysE-Inv surpasses all tested baseline models, delivering an average Mean Squared Error (MSE) reduction of 24.7% across the board, and a 17.3% performance boost over the best-performing baseline in parameter estimation tasks. Ultimately, this study highlights a versatile inverse modeling approach for data-scarce fields, demonstrating the value of integrating physics-informed guidance with sparse observational data.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC