Subproject M2: Learning of Dynamical Process Models based on Data and Expert Knowledge

Professor Dr.-Ing. Gerhard Neumann
Professor Dr.-Ing. Luise Kärger
Dr.-Ing. Julius Pfrommer

This project is concerned with the learning of dynamical process models from the input/output time series delivered by the simulation models (provided by T2) and the process instrumentation (provided by T1/M1). These learned models will be used in M3 for model-based parameter optimisation, and in M4 for designing feedback controllers that can account for the inherent observation uncertainty and actuation noise once the over-instrumentation of the process is reduced.

Our dynamics models have to predict the resulting geometry as well as additional spatially distributed features of the product. This will be achieved by employing a mesh-based neural architecture that uses message passing in a Graph Convolutional Neural Network (GCNN) to propagate the state information spatially throughout the mesh. Such network architectures have already been shown to predict accurately the dynamics of complex Finite Element (FE) simulations, while being considerably cheaper to evaluate, allowing a potential use of these models for data-driven optimisation techniques (M3 and M4) in a feasible amount of computation time.

However, such dynamics models can (so far) only be trained from simulation data, and have not been conditioned on parameters or distributed actuators; it is, therefore, unclear whether the long-term prediction accuracy is sufficient for subsequent model-based optimisation of the process parameters or feedback controllers. Besides data-driven learning, expertise and knowledge about the physics of the process will be exploited in order to obtain high-quality models, even if few data are available.

In addition to real data obtained from the process, we will leverage several virtual data sources from physically motivated simulations with different levels of fidelity, and we will transfer models learned solely from simulation data to the real world using sim-2-real techniques. We will therefore attack the following research challenges in this project: (i) obtain neural mesh-based models that are fast enough for optimisation; (ii) go beyond the accuracy of the simulated data by adapting to real datasets; (iii) condition dynamics models on distributed parameters and/or actuators, while maintaining those models; (iv) work with a low number of training samples by combining learning approaches with analytical physics priors; (v) transfer already learned models to new, slightly adapted, structures of the process.

Further, this project requires the use of the following AI methods: Geometric Deep Learning, Graph Convolutional Neural Networks (GCNN), Recurrent Neural Networks (RNN), Meta Learning, Sim-2-Real Transfer, Physics-Informed Neural Network Architectures and Variational Inference.