2024-07-10 –, Else (1.3)
We introduce MeshGraphNets.jl, an open-source library that provides the neural network architecture of MeshGraphNets by Google DeepMind in the context of NeuralODEs. It supports full computation on the GPU for accelerated training and evaluation of the system and provide an interface for customizing MeshGraphNets for your own use case. We show that the network has an improved understanding of the underlying phenomena on a network derived from a real world hydraulic brake system.
Simulating physical systems with a spatially distributed domain, such as those encountered in fluid dynamics or structural mechanics, requires solving intricate and computationally demanding differential equations. By taking advantage of the mesh-like structure of these methods and incorporating machine learning methods in place of the equations we can significantly improve the computation time during system evaluation. This is the core concept of MeshGraphNets and will be discussed in the first part of the talk.
The second part will continue with the specific details of the design choices of the Julia packages MeshGraphNets.jl and its underlying core package GraphNetCore.jl. Designing MeshGraphNets as a NeuralODE allows you to incorporate the full suite of ODE solvers from DifferentialEquations.jl into both the training and the evaluation process. For example, you can easily switch from a derivative-based collocation training to a solver training to enable the network to more accurately learn the actual physics behind the modeled system. In order to efficiently train the NeuralODE we designed our packages with full compatibility and optimised execution on the GPU in mind.
We will conclude our talk by presenting the improvements of a NeuralODE based design in the context of an industrial application in the form of a hydraulic brake system.
Diploma in Mathematics. PhD in Mechatronics. Worked in Bosch Corporate Research now head of the Chair for Mechatronics at the University of Augsburg.
Research scientist @ University of Augsburg, chair of mechatronics
Github:
- JulianTrommer
- Chair of Mechatronics