Speaker
Description
Accurate and efficient numerical methods for solving differential equations remain fundamental to modern computational engineering. Challenges arise, e.g., from costly vector field evaluations, high-dimensional states, multirate dynamics, long-term integration, real-time requirements in control, or ensemble forecasting.
This work explores neural network-based approximations for the integration errors inherent in standard numerical methods, as well as the development of more advanced integration schemes. By combining the strengths of traditional physics-based numerical methods with the universal approximation capabilities of neural networks, we seek to enhance both computational efficiency and the mitigation of errors in classical integration techniques. Concretely, we propose NN-enhanced Runge-Kutta schemes as well as structure-preserving enhanced symplectic integration methods.
The approach is applied to the simulation of wind turbines based on open-source automatically generated wind turbine models with parameters derived from the open-source reference simulation framework OpenFAST. These multi-body models generated using CADynTurb contain modal representations of the elastic tower and blades and modular aerodynamics models. Large simulation tasks occur, e.g., in fatigue analysis or when predicting the remaining useful lifetime for which ranges of initial conditions, system parameters, and exogenous inputs have to be considered.
