Towards a machine learning approach of deep moonquake source regions classification using their temporal and spatial patterns - application for a single station mission

In the prospects of the new lunar seismology missions, such as it is a new in-situ seismic experiment called Farside Seismic Suite (FSS) selected by NASA, we explore a novel approach of classifying seismic events using the existing Apollo data. The FSS, onboard CP-12 lander, shall land at the farside of the Moon in Schrödinger Basin, and once in the operational stage should provide the community with the data to further constrain lunar interior and the Moon seismicity. However, the localisation of the newly seismic events shall be challenging, due to the single-station nature of the mission.  Therefore, in this study we develop a pipeline for the deep moonquake (DMQ) source region classification, thus localisation, on the legacy of the data acquired during the Apollo missions. 

DMQs are are a distinctive group of the seismic events that seem to be predominately occurring at the near side of the Moon, at the depths between 700 - 1200 km, in the conditions with high pressure and temperature values. These events are characterised with highly repetitive waveforms, and clustering these waveforms reveals that DMQs are organised in several source regions, called nests. The activation of these nests is closely related to the gravitational solid tides generated due to the monthly motion of the Moon around the Earth.  Thus, the DMQ occurrences exhibit tidal periodicities.

In this study we explore how we can exploit DMQs spatial and temporal occurrence patterns for the purpose of their localisation with future missions. Spatial patterns are determined by the P and S travel time estimates. We can show that deploying a station close to the south pole on the far side, and using the existing lunar models, DMQ nests cover travel time estimates between approximately 160 and 234 seconds. If we consider that travel time estimates of DMQ nests can vary with some standard deviations, such as 5 seconds, then we can from group of nests, called sets, that share similar travel times. Therefore, nests that belong to certain sets cannot be distinguish using just the travel time information. To further distinguish individual nests within a set, we explore the DMQ temporal patterns that are being related to the monthly lunar phases. It has been shown that different nests correspond differently to three lunar months: synodic, draconic, anomalistic. 

By characterising the spatial and temporal patterns of the DMQ occurrences we develop a machine learning (ML) model for nets classification. This is carried out by defining features which are used as an input data to ML model. The input features we use are related to the orbital parameters describing the monthly motion of the Moon around Earth. Eventually, the ML model  learned how to classify between nests that belong to the sam set. We report that models are achieving an accuracy over 70% when those are trained to classify =< 4 nests within the set, and better than 90% when only two DMQ nests are in the same set. This approach opens up a new way to DMQ location estimate, on the near and farside of the Moon, when captured by the future FSS single-station seismometers or other seismic stations on the Moon.