KAN-Mamba with Physics-Informed Learning for Landfill Methane Leak Localization

Abstract

Landfill methane leakage is spatially sporadic, meteorologically sensitive, and difficult to pinpoint from sparse concentration measurements. This paper introduces a physics-informed KAN-Mamba model for high-resolution leak localization by combining Kolmogorov-Arnold Networks (KANs) to handle nonlinear source-to-receptor mapping, a selective state-space Mamba backbone for long-range temporal dependency, and differentiable plume constraints that penalize mass-balance and advection-diffusion inconsistencies. A simulated field campaign with 64 fixed sensors, two mobile transects, 46 leak events and 18 meteorological regimes was conducted to test the model against LSTM, Transformer, PINN and wind-conditioned graph neural network baselines. The median localization error of the proposed model was 6.8 m, the source-rate mean absolute error decreased to 0.031 kg h-1, and the event-level F1-score reached 0.941 under mixed wind stability; it exceeded the best baseline by 21.6% in distance error and 8.3% in recall. Based on the ablation experiment, the physics loss reduced false alarms near the downwind boundary by 34.7%, and the KAN layer improved the expected calibration error under weak wind from 0.113 to 0.061. Robustness tests further showed a recall of over 0.89 at 35% receptor dropout and only 2.6 m additional error under +/-10-degree wind perturbation. These results indicate that physically constrained sequence learning can convert sparse methane data into high-quality source maps for engineering inspection and landfill emission control.