Neural network modeling of seismic behaviour of the hellenic Arc: strengths and limitations

The strategy of earthquake-proof construction and seismic risk reduction requires constant improvement of methods of calculation and compilation of increasingly informative normative forecast maps of seismic hazards. Despite the wide range of available methods for fixing deformations of the earth's crust, a reliable seismic forecast is still not possible because local changes in parameters do not always lead to earthquakes, and environmental heterogeneity does not allow to single out any bright shift that can make one think about future earthquakes. The introduction of modern mathematical methods and the development of the newest computer technologies based on artificial intelligence (AI) give a chance to predict the occurrence of natural disasters, in particular, earthquakes. This study aims to build a mathematical apparatus for earthquake prediction, which is based on the use of neural networks (NNs) to process large amounts of information. Artificial neural networks (ANNs) can be used to approximate any complex functional connections. The article presents the results of developing a neural network model (NNM) for forecasting occurrence numbers and sizes of medium-strong earthquakes (Mw >=4 on the Richter scale). To build a forecast NN, data on earthquakes recorded in Greece for the period 2000-2020 (about 2,500 events) were used. The NN receives input from three independent variables: geographical coordinates of the earthquake's latitude, geographical coordinates of the earthquake's longitude, and the earthquake's depth. The construction of a NN to predict strong earthquakes was implemented in the development environment RStudio programming language R. Neuralnet package was used to build the required NN, which contains a very flexible function for training feed-forward neural networks (FFNNs) and allows you to simulate many internal hidden layers and hidden network neurons. We have also used the nnet package, which is a universal tool for building predictive models in NN programming. The result is a NN of the multilayer perceptron type, which includes 2 hidden layers consisting of 5 and 3 neurons, respectively, which generate input data at the output of the network. The NN perceived model of seismicity not only describes the process of occurrence (generation) of earthquakes in Greece, but can also be used to estimate magnitudes of forthcoming seismic events.