TL;DR
This paper presents a machine learning approach to predict the source regions of deep moonquakes (DMQs) using only their occurrence times and lunar orbital parameters. This method is crucial for interpreting data from future single-station lunar seismic missions, offering a way to estimate quake locations even without waveform data.
What Problem Does This Paper Address?
Future lunar exploration missions, like NASA's Farside Seismic Suite (FSS), will deploy seismic stations on the Moon's surface. A key challenge is interpreting data from these new, potentially single-station deployments, especially when trying to relate it to the wealth of data collected during the Apollo missions. Deep moonquakes (DMQs) are the most frequent seismic events on the Moon, and their source regions are known to exhibit temporal and spatial patterns linked to lunar tidal cycles. However, directly comparing new seismic waveforms with old data is difficult due to differences in station locations and instrument characteristics. This paper addresses this by developing a machine learning model that can classify DMQ source regions using only temporal information (occurrence times) and lunar orbital parameters, bypassing the need for waveform analysis.
Methods
The core of this research is the application of a Random Forest (RF) classifier. The authors extracted eight features related to lunar orbital mechanics and the timing of moonquakes. These features include:
- Time differences between a quake's occurrence and key lunar orbital events (ascending node, full moon, perigee).
- The Moon's position in its orbit (true anomaly, represented by its cosine and sine).
- Time intervals between successive quakes within a nest.
- The distance between the Moon and Earth, and the rate of change of this distance.
These features were then used to train an RF model to classify DMQs into their known "nests" (source regions). The model was trained on a subset of Apollo data and evaluated on a separate test set. The authors also explored the impact of feature normalization and the number of nests included in the classification.
Datasets
The primary dataset used is the Apollo lunar seismic catalog, which contains information on deep moonquakes (DMQs), including their occurrence times and previously determined source region classifications (nests). This catalog was updated from Nakamura et al. (1981) and revisited in Nunn et al. (2020a, 2020b). The study also utilizes lunar interior models from various sources (Garcia et al., 2011, 2019; Khan et al., 2014; Matsumoto et al., 2015; Weber et al., 2011) to calculate travel times for feature engineering.
Key Insights and Ideas
- Temporal patterns are powerful predictors: The study demonstrates that the timing of moonquakes, relative to lunar orbital cycles, contains significant information about their source regions, even without waveform data.
- Lunar orbital parameters are key features: Features derived from the Moon's position and motion (true anomaly, tidal events) are highly effective in distinguishing between DMQ source regions.
- Single-station interpretation is feasible: The RF model's success suggests that future single-station seismic networks can contribute meaningfully to understanding lunar seismicity by leveraging these temporal and orbital features.
- Complexity increases with more classes: As the number of distinct DMQ nests to classify increases, the model's performance generally decreases, highlighting the challenge of distinguishing between nests with similar underlying mechanisms.
Key Results
- The Random Forest model achieved accuracy greater than 70% when classifying 4 or fewer DMQ source regions.
- When classifying just two of the largest nests (A1 and A8), the model achieved an accuracy of 89%.
- The most important features for classification were found to be cos(γ) (cosine of true anomaly), ΔtPG (time difference to perigee), d (distance to Earth), ΔtAN (time difference to ascending node), and ti+1 - ti (time between successive quakes).
- The study found that adding more nests to the classification problem led to a decrease in overall accuracy, dropping to 77% when classifying five nests.
- The model's performance was generally better for nests with higher precision, indicating a higher confidence in correctly identifying those quakes.
Limitations and Future Work
- Limited by Apollo data: The study is constrained by the quality and quantity of data from the Apollo missions. The number of events per nest varies significantly, leading to unbalanced datasets.
- Waveform data excluded: The current model intentionally excludes waveform information, which is a limitation for direct comparison with future data. However, this was a deliberate choice to address the single-station challenge.
- Future work suggestions:
- Incorporate amplitude and energy information as additional features.
- Explore sequential classification models that consider the order of events.
- Investigate one-to-one or one-to-all decomposition strategies for multi-class classification.
- Retrain the model with a complete DMQ dataset and new data from future missions to improve accuracy and nest localization.
- The authors suggest that future missions could benefit from this approach to provide preliminary quake source locations.
Summary updated