Interpretable Scientific ML with KAN
Physics-guided Kolmogorov–Arnold Networks that turn learned predictors into inspectable retrieval formulas.
Many scientific machine-learning models improve prediction while remaining difficult to inspect. This work combines a radiative transfer model with a Kolmogorov–Arnold Network (KAN) to learn an explicit soil-moisture retrieval formula from physically consistent simulation data.
The approach connects physics-based modeling with modern machine learning: KAN training discovers compact symbolic relationships among brightness temperature, surface temperature, vegetation optical depth, and soil moisture. Once derived, the formula can be applied directly without retraining the network or iteratively solving the radiative transfer model.
On non-forest in-situ data, the method achieved a 0.98 Pearson correlation and improved both interpretability and generalization over a multilayer perceptron.
Published paper · JIF 12.3 (2025) · JCR Q1 (SCIE) · Open PDF
