If physics-based simulation models cannot be improved further due to lack of information or exploding development times, NeuralFMUs come into play.
NeuralFMUs are hybrid models that combine the knowledge of existing physics-based simulation models in form of Functional Mock-up Units (FMUs) with the impressive potentials of neural networks. The deviation between simulation model and reality is strongly reduced by a trained NeuralFMU. Since only the missing physical effect needs to be understood, relatively small amounts of data are sufficient to train NeuralFMUs – so existing measurements or simulation outputs can often be recycled for this purpose.
The resulting NeuralFMU is a more predictive simulation model with good extrapolation capabilities. NeuralFMUs understand the missing physical effects instead of just compensating the modeling error consequences. Simulation results not only look good, but are based on physical knowledge gained from data.
Within the UPSIM Brakesystem Use Case, a NeuralFMU of the EC-Motor actuating the brake system was trained on real measurements from complete braking maneuvers. The prediction quality could be significantly enhanced by 62% compared to the original model (on training data). Even with completely unknown validation data, an improved prediction accuracy of more than 40% was confirmed. This enhanced simulation model was created with very little development effort for the end user – the original simulation model and some existing prototyping data where simply reused.
These improvements allow for targeting additional virtual testing and thus a more efficient development process. Further methodical enhancements are in progress.
FMI.jl (open source library): https://github.com/ThummeTo/FMI.jl
FMIFlux.jl (open source library): https://github.com/ThummeTo/FMIFlux.jl
Related Publication: https://doi.org/10.3390/electronics11193202
#PhysicsAI #NeuralODE #Physics-enhanced NeuralODE #NeuralFMU #HybridModelling #LikeABosch