Full-Waveform Modeling and ML- based Tomography of Volcanic Edifices

Abstract

Seismic imaging is an essential method for exploring the structure of volcanic features beneath the surface, such as conduits and magma chambers. These insights are crucial for a more profound understanding of volcanic processes. However, the complex and heterogeneous nature of the subsurface, along with the limited availability of data, presents substantial challenges in accurately mapping the internal structure of volcanoes, often leading to interpretive uncertainties. This study explores the application of machine learning, specifically artificial neural networks (ANNs), in seismic tomography and full-waveform inversion to improve volcanic imaging. The training is performed on synthetic data generated from randomized velocity structures. In the initial phase, a feedforward neural network (FNN) is trained on travel-time data to predict velocity profiles. Tests in the 2D domain for simplified cases demonstrate that the ANN approach outperforms traditional linear inversion techniques, especially when seismic data is sparse. Furthermore, the method’s efficiency in the 3D domain is demonstrated. However, in real-world scenarios, such inversions would significantly benefit from using full-waveforms to accurately determine the velocity profiles. An ANN with an encoder followed by a fully connected stack is utilized. The encoder, composed of convolutional layers, extracts features from the waveforms, which are then used by the fully connected stack to predict velocity profiles. Initial tests show promising results, suggesting that encoders can capture key characteristics of the waveforms, beyond just arrival times, which can be leveraged for inversion. The ANN architecture can be further optimized to enhance the inversion process and better account for complex velocity profiles.