Mitigating array tilt mismatch in machine learning-based matched-field processing using waveguide-informed data augmentation

We propose a machine learning-based matched field processing (MFP) method for underwater source range estimation using a vertical line array (VLA). One challenge in MFP is its sensitivity to array tilt, which can significantly degrade performance when tilt conditions differ between the training and test datasets. To address this, we introduce a physics-guided data augmentation technique that generates synthetic training data across multiple tilt angles using waveguide invariant theory. This approach applies tilt-induced phase corrections to sample covariance matrices, enabling the model to learn tilt-robust features without requiring additional measurements. We validate our method using data from the SAVEX15 experiment, where broadband noise (200–500 Hz) from a surface ship was recorded in a shallow-water environment. Results show that our method significantly improves range estimation accuracy under tilt mismatch compared to conventional machine learning-based MFP.