Abstract Moonquakes provide critical observations for probing the lunar interior, yet their analysis is hindered by the limited number of recordings and their inherently low signal‐to‐noise ratio (S/N). Conventional detection methods such as Short‐Term Average/Long‐Term Average (STA/LTA) perform poorly on lunar data, while standard deep learning models (e.g., CNN, LSTM) require large volumes of clean, fixed‐length inputs and often generalize weakly across domains. To address these challenges, we investigate the Fourier Neural Operator (FNO) for moonquake detection. To our knowledge, this study presents the first application of Fourier Neural Operators to planetary seismic event detection, leveraging operator learning to remain data‐efficient under limited labels and variable acquisition settings. Our training data set includes waveforms and spectrograms of events and noise windows. Cross‐domain generalizability is assessed under two regimes: (a) training on earthquakes only, and (b) training on earthquakes augmented with 64 labeled moonquake examples. Both models are evaluated on independent Apollo Passive Seismic Experiment (PSE) and Lunar Seismic Profiling Experiment (LSPE) records. Despite limited training data, the 1D and 2D models achieve high detection performance, with F1‐scores of 0.96 and 0.99, respectively. Furthermore, the resolution‐invariant nature of FNO enables application to waveform or spectrogram inputs of arbitrary length and sampling rate. Compared to CNN‐based approaches, FNO models require fewer parameters, fewer training examples, less computation time, and yield superior cross‐domain generalization while maintaining competitive accuracy. These results highlight FNO as a flexible, lightweight framework for real‐time lunar seismic monitoring, with clear potential for extension to other planetary data sets such as Mars InSight.
Al‐Qadasi et al. (Sun,) studied this question.