Project description
The student will begin with simple isotropic models and gradually extend the approach to anisotropic media (VTI, TTI). The core idea is to train a neural network, starting with a so-called multilayer perceptron (MLP), to predict the position and slowness of an ‘intermediate’ ray given its neighbors. A possible later extension is to use a graph neural network (GNN).
Training data will be generated using dense WFC simulations, from which rays are selectively removed to create supervised learning targets.
The project includes:
- Implementing ray tracing in isotropic and anisotropic media.
- Designing ML architectures for local interpolation.
- Incorporating physics‑informed constraints (e.g., Hamiltonian consistency).
- Evaluating accuracy, stability, and computational efficiency.
- Integrating the ML operator into a WFC prototype.
The project is well‑bounded but offers room for innovation. It is suitable for a student with interest in computational geophysics, numerical methods, and machine learning.