Predicting deep moonquake source regions using their temporal and spatial patterns and machine learning

In the near future, lunar exploration should be enhanced by deploying a new seismic sta- 16 tion on the farside of the Moon, the Farside Seismic Suite (FSS), a package of two seis- 17 mometers recently selected by NASA to fly on a commercial lander. The new data should 18 provide us with new insight into Moon’s seismic activity, its interior and composition. 19 To fully benefit from the new data, we need to take advantage of data acquired during 20 the Apollo missions. The problem of relating new and old data is complex due to the 21 single-station nature of the future deployment. In this study we tackle this issue by de- 22 veloping machine learning model in the context of the deep moonquake (DMQs) clas- 23 sification problem. The DMQs form the largest group of detected events from Apollo data, 24 their source regions have been located and are known to exhibit temporal and spatial 25 patterns connected with the monthly lunar tidal periods. Therefore, we propose to uti- 26 lize a machine learning (ML) algorithm named Random Forest to identify DMQs source 27 regions without using waveform information, and only using the lunar orbital parame- 28 ters related to DMQs time occurrences. We show that ML models perform well (with 29 an accuracy > 70%) when they are trained to classify 4 or fewer different source regions. 30 This approach gives us a good first location approximation of the DMQs source regions 31 and opens up a new approach to their location estimate when captured by the future 32 FSS single-station seismometers.