ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-09272021-185602


Tipo di tesi
Tesi di laurea magistrale
Autore
BOZZI, EMANUELE
URN
etd-09272021-185602
Titolo
A COMPARISON OF MACHINE LEARNING ALGORITHMS FOR THE INVERSION OF RAYLEIGH WAVE DISPERSION CURVES
Dipartimento
SCIENZE DELLA TERRA
Corso di studi
GEOFISICA DI ESPLORAZIONE E APPLICATA
Relatori
relatore Aleardi, Mattia
Parole chiave
  • Transfer learning
  • Artificial neural networks
  • Machine learning
  • Dispersion curves
  • Surface waves
  • MASW
Data inizio appello
22/10/2021
Consultabilità
Completa
Riassunto
In this thesis the Rayleigh Wave dispersion curves are inverted to recover 1D near-surface Vs-Vp profiles, exploiting machine learning techniques. The geophysical investigation under study is the Multichannel Analysis of Surface Waves (MASW).

The MASW problem is both non-linear and ill-conditioned, usually tackled with deterministic or stochastic methods, which however suffer of multiple minima and/or high computational costs. Here, a machine learning-based approach is proposed as an alternative method for the inversion.

Various Artificial Neural Networks (ANNs) are developed to model the non-linear inverse map between the data and model spaces. In this framework, the inputs to the ANNs are regularly sampled dispersion curves, while the outputs are subsurface models, parametrized as layer thicknesses-Vp-Vs. Specifically, five architectures are compared in this study, including few (i.e., shallow networks) or several (i.e., deep networks) hidden layers: FeedForward Neural Network (FFNN), Cascade Forward Neural Network (CFNN), Deep Fully Connected Neural Network (DFCNN), Convolutional Neural Network (CNN) and Residual Neural Network (RNN). To exploit them, their hyperparameters are properly set and 1) a generation, 2) a training and 3) an extensive test phases are adopted. The Thomas-Haskell method is the forward modeling for the generation of the training/validation/test datasets, based on the

definition of a gaussian prior model. The root mean square error (RMSE) between the predicted and generated (or target) models quantifies the ANNs prediction accuracy. Eventually, once given a dispersion curve, these architectures can provide model estimations in near real-time.

After two major comparative tests among the investigated ANNs, the CNN is selected as the reference network and exploited for additional analysis and a field test, since it guarantees stable/accurate predictions and reasonable training costs. Moreover, the transfer learning strategy is tested on the updating of an already-trained CNN, to be suitable for a new task (i.e., predicting models with a different Vs mean prior model assumption). To this end, the network is quickly retrained with a smaller amount of target examples, generated under the new model assumption. The updated CNN is successfully tested on this new scenario, showing similar accuracies compared to a CNN specifically trained from scratch for the same task. Eventually, the CNN predictions are compared with those provided by a more standard global optimization method (i.e., Particle Swarm Optimization), which confirms the applicability/reliability of the investigated machine learning-driven inversion framework.

A Monte Carlo approach is used to overcome the ANNs lack of uncertainty estimation, and both the data uncertainties and the network modeling errors are propagated onto the predicted models. To this end, a Posterior Probability Density Function (PPD) is numerically estimated from the ensemble of Monte Carlo samples.

In the concluding part of this study, the CNN is applied to a field dataset (Grenoble, France), provided by twelve co-located shot gathers. Noise contamination of the data results in a challenging dispersion curve picking, especially at low frequencies. The network is trained with target examples generated from a gaussian prior model derived from a previous study. As a result, the CNN provides a model estimation which shows similarities with models yielded by other inversion approaches and borehole measurements. Furthermore, the predicted dispersion curve adequately reproduces the observed data. Even in the field test, transfer learning allows to quickly update a previously trained CNN, for a new task (i.e., different assumption on the Vs mean prior model).

This work verifies the potential of machine learning methods for solving the MASW problem. The implemented CNN provides accurate model estimations and, thanks to the Monte Carlo approach, appropriate uncertainty evaluations. One of the main advantages of the proposed approach is that it can be used to estimate the near-surface Vs-Vp profile and its uncertainties in near real-time
File