Editing Earthquake prediction (section)

==== Machine learning ====
Rouet-Leduc et al. (2019) reported having successfully trained a regression [[random forest]] on acoustic time series data capable of identifying a signal emitted from fault zones that forecasts fault failure. Rouet-Leduc et al. (2019) suggested that the identified signal, previously assumed to be statistical noise, reflects the increasing emission of energy before its sudden release during a slip event. Rouet-Leduc et al. (2019) further postulated that their approach could bound fault failure times and lead to the identification of other unknown signals.<ref>{{Harvnb|Rouet-Leduc|Hulbert|Lubbers|Barros|2017}}.</ref> Due to the rarity of the most catastrophic earthquakes, acquiring representative data remains problematic. In response, Rouet-Leduc et al. (2019) have conjectured that their model would not need to train on data from catastrophic earthquakes, since further research has shown the seismic patterns of interest to be similar in smaller earthquakes.<ref>{{cite web|url=https://www.quantamagazine.org/artificial-intelligence-takes-on-earthquake-prediction-20190919/|title=Artificial Intelligence Takes on Earthquake Prediction|last1=Smart|first1=Ashley|website=Quanta Magazine|date=19 September 2019|access-date=2020-03-28}}</ref>

Deep learning has also been applied to earthquake prediction. Although [[Bath's Law|Bath's law]] and [[Omori's law]] describe the magnitude of earthquake aftershocks and their time-varying properties, the prediction of the "spatial distribution of aftershocks" remains an open research problem. Using the [[Theano (software)|Theano]] and [[TensorFlow]] software libraries, DeVries et al. (2018) trained a [[Artificial neural network|neural network]] that achieved higher accuracy in the prediction of spatial distributions of earthquake aftershocks than the previously established methodology of Coulomb failure stress change. Notably, DeVries et al. (2018) reported that their model made no "assumptions about receiver plane orientation or geometry" and heavily weighted the change in [[shear stress]], "sum of the absolute values of the independent components of the stress-change tensor," and the von Mises yield criterion. DeVries et al. (2018) postulated that the reliance of their model on these physical quantities indicated that they might "control earthquake triggering during the most active part of the seismic cycle." For validation testing, DeVries et al. (2018) reserved 10% of positive training earthquake data samples and an equal quantity of randomly chosen negative samples.<ref>{{Harvnb|DeVries|Viégas|Wattenberg|Meade|2018}}.</ref>

Arnaud Mignan and Marco Broccardo have similarly analyzed the application of artificial neural networks to earthquake prediction. They found in a review of literature that earthquake prediction research utilizing artificial neural networks has gravitated towards more sophisticated models amidst increased interest in the area. They also found that neural networks utilized in earthquake prediction with notable success rates were matched in performance by simpler models. They further addressed the issues of acquiring appropriate data for training neural networks to predict earthquakes, writing that the "structured, tabulated nature of earthquake catalogues" makes transparent machine learning models more desirable than artificial neural networks.<ref>{{Harvnb|Mignan|Broccardo|2019}}.</ref>