Are neural networks or neural network based finite elements superior to traditional finite elements ?

Traditional modeling tools such as finite elements are being challenged by modern machine learning algorithms and software tools. What is best?

Being a master’s student

List of projects

Neural network framework such as TensorFlow, PyTorch and JAX have changed the way simulations systems are being made. Further, finite element system such as TorchMesh, JAX-FEM provide implementations of finite element methods often in a more user-frieldly system and a more direct GPU compatible implementation. In this project we aim to compare such systems with traditional systems such as FEniCS in terms of in particular efficiency on moderate to large scale problems.

Goal

Compare traditional computational modeling tools such as FEniCS with modern neural network tools in terms of efficiency, accuracy and robustness.

Learning outcome

Deep knowledge of modern numerical methods and simulation tools
Experience with contemporary neuroscience challenges and models
Team work in an active research group

Qualifications

A background in applied mathematics and/or computer science
Knowledge of partial differential equations and numerical methods
Experience with Python programming.

Supervisors

Kent-Andre Mardal

Collaboration partners