Eikonal Solution Using Physics-Informed Neural Networks for Global Seismic Travel Time Prediction
Abstract
Neural networks with physics constrained formulation to predict solutions has shown a reasonable interest across science and engineering. The method addresses challenges that have long been unsolved via the conventional numerical approaches. We introduce a physics informed neural network scheme geared towards solving the Eikonal equation applied on the global Earth velocity model. The proposed scheme relies on two important networks. The first network acts as an interpolation operator for the global velocity model, which will be useful for both training and further updates on the velocity model. The other main network, which is constrained by the Eikonal equation provides the first-arrival travel time for any given source-receiver pairs. The networks are built using the open-source software TensorFlow. We demonstrate one important benefits of using the proposed scheme compared to any numerical methods, namely the ability to be free from the upwind limitation. Hence, the scheme will still work on a velocity model with gaps (i.e. no information of velocity between sources and receivers). The ability to perform transfer learning and surrogate modeling on the network will allow the proposed scheme to learn faster the travel time solution. In addition, the trained physics-informed network could utilize transfer learning method for future velocity updates, and it can even be useful in developing the velocity updates themselves. Furthermore, the training time is reduced by applying surrogate modeling. The end result is a neural network model that stores the Earth model information to generate travel times between sources and receivers instantly. This neural network model can be updated efficiently to accommodate any changes on the global velocity model.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.S34A..01T