A New Physics‐Constrained Neural Network Approach to Retrieve CH 4 Vertical Profiles From Hyperspectral Infrared Data

Abstract Methane () is a critical greenhouse gas, and accurate monitoring of its vertical distribution is essential for understanding its radiative impact and identifying emission sources. This paper introduces a novel methodology using a physics‐constrained neural network based on soft constraints to retrieve vertical profiles from hyperspectral Infrared Atmospheric Sounding Interferometer (IASI) data. The methodology combines the advantages of Optimal Estimation and neural networks, leveraging physical constraints to enhance retrieval accuracy while maintaining computational efficiency. The approach integrates dimensionality reduction techniques, such as Principal Component Analysis and Averaging Kernel optimization, to maximize information content while reducing computational complexity. Two models were developed: one for general profile retrieval and another optimized for detecting surface level enhancements. The novel approach was validated on synthetic data and applied to real IASI observations, including the Nord Stream pipeline sabotage event, and demonstrated that the algorithm can confidently detect enhancements as low as 50 ppbv near the surface and reconstruct profiles with average error of ppbv in the mid‐troposphere and >100 ppbv in the stratosphere on the considered test set. While effective in capturing strong enhancements, the technique exhibits slightly reduced sensitivity near the surface due to limited spectral information, which is mitigated by inflating covariance matrices in the cost function. This methodology offers a promising tool for global vertical monitoring with existing sensors and could be potentially extended to other compounds such as .