• Deep Learning (DL)
• Earthquake Analysis
• Neural Networks (NNs)
• Python
• Seismology

The innovation of Artificial Intelligence (AI) and Machine Learning (ML) in the field of geophysics
(and not only) is grown exponentially in the last decade and proved the effectiveness of these
techniques to ease the burden of operational tasks, complementing it with fast computational power.
This Master's thesis focuses on applying and improving machine learning techniques to
observational seismology tasks, specifically the detection and location of seismic events. I’ve
contributed to the development of a graph-based, deep-learning (DL) method, using the Linux
Operating System, the Python programming language, and the PyTorch library. The algorithm used
is part of a subcategory of machine learning, Deep Learning (DL), widely explained in a dedicated
chapter of this thesis. The algorithm uses an Artificial Deep Neural Network (ADNN) called Graph
Neural Network (GNN), specifically able to work with graphs.
The dataset analyzed in this work comes from a seismic sequence that occurred in the Kumamoto
region (Japan) in April 2016. More precisely, we analyzed with the algorithm a couple of weeks
around the main events, the 14th and 16th of April.Starting from the theory of earthquake physics, we firstly tested the approach using a fully-based
synthetic approaching, calculating the Green’s Functions of the region by using velocity models
reported by local seismic agencies. We then created waveforms to utilize in the training of our
models. We then start to preprocess the data using 4 characteristic function channels needed to
bridge the synthetic and real-data worlds. Though, while the synthetic realm allows us full control
over the training of DL models, it is also true that a direct correspondence with real data is
unfeasible, no matter the effort. Therefore, we adopted a fine-tuning of our model’s weights using a
handful of events extracted from the real data. This passage better allowed the convergence of our
training and its generalization capability to infer from new data.
The goal is to make the algorithm able to learn and recognize seismic events, as well as the arrivals
of P and S waves, therefore discerning them from other physical signals (e.g., noise), and provide a
preliminary location of the seismic events in the 3D volume.
Throughout the thesis, I’ve documented, software utilization through the usage of Jupyter
Notebooks, and build several bash and Python scripts to handle the code library. This work helped
me expand my experience in computational geophysics, critical thinking, and problem-solving.