Fast EHD SiML - EHD Building Block for System Simulation
The goal of the Fast EHD research project is to develop a method for integrating distribution-parametric simulation models into system simulations. The integration shall precisely, quickly, and easily integrate an elastohydrodynamic (EHD) simulation for sealing contacts, which was developed in a previous project, into a system simulation. In doing so, novel neural networks (NN) will be implemented for a robust and accurate representation of the existing EHD.
Targets and Approaches
| Benefit | Procedure |
|---|---|
| Integration of distribution parametric models in system simulations | Acceleration of the underlying pressure distribution equation |
| Precise and computationally efficient calculation of sealing simulations | Acceleration calculation of deformation |
| Possible use in system simulations | Merge both solutions |
| Optimization of the simulative design process | Implementation of the Fast EHD SiML module |
Motivation
There is currently an increasing shift in the design process from purely experimental approaches to greater use of simulation software. This shift offers the potential to save time and money, as time-consuming and expensive experiments need not be integrated until later in the product development cycle. However, the success of more simulation-based design depends largely on the precision of the simulation models, as reliability and accuracy of design results suffer without sufficiently detailed modeling. Although many approaches already exist for describing components and individual contacts, these complex simulation approaches often cannot be integrated into existing simulation infrastructures. The reasons for this are the high complexity of model generation, necessary expert knowledge and high computing times. In fluid technology, for example, this problem occurs in sealing contacts. Since seal friction is highly nonlinear, complex elastohydrodynamic (EHD) calculations are required for accurate computation. However, the computational time required to calculate a single sealing contact often already exceeds the limits allowed for a system simulation. Therefore, the sealing contacts are currently usually considered individually. However, many fluid power components, such as hydraulic cylinders or pneumatic valves, consist of several sealing contacts. These interact with each other during operation due to the pressure or kinematics. In the view of individual sealing points that has prevailed up to now, these interactions can only be inadequately taken into account. The separation of system and contact simulations, as well as the isolated consideration of individual details, currently prevents existing potentials from being exploited in the modeling and design of fluid technology components and systems. A combination of individual contact simulation and system modeling would improve existing approaches to system simulation. This requires a method that accelerates distribution-parametric simulations and provides them with interfaces for connection in system simulations. The benefit is a broader application of complex simulation methods, which speeds up the design process and increases product quality.
Project goal
The aim of the research project is to develop a method to integrate distribution parametric simulation models into system simulations. The integration should precisely, quickly and easily integrate an EHD simulation for sealing contacts, which was developed in a previous project, into a system simulation.
The ifas-DDS EHD simulation model
Acceleration of the seal simulation
Within the project, a Fast-EHD framework will be developed by ifas-DDS. The framework is both data- and physics-based, which allows for later transfer of the method to other EHD contacts. Novel neural networks (NN) will be implemented for robust and accurate mapping of the existing EHD. A weakness of classical NN is the amount of data required for training. In order to find a meaningful relationship between input and output data, the network needs a large amount of different data. Depending on the process under investigation, data acquisition can be particularly difficult. An insufficient amount of data usually leads to incorrect and unusable results of the network. Other weaknesses of NNs are the long training time and the lack of robustness compared to established numerical methods such as EHD. These weaknesses have driven the development of a new type of neural network: The Physics-informed neural network (PINN). In this approach, physical knowledge is incorporated into the PINN in the form of system-describing differential equations or empirically proven rules. Compared to conventional neural networks, the PINN can strongly limit the space of possible solutions by the addition of information, since not only the data-based behavior but also the physical relationships of the system parameters are relevant. This results in a greatly reduced training time and a much more robust behavior of the PINN. Figure 3 shows a schematic comparison of NNs, PINNS and EHD based on training time, computation time and robustness. While the EHD simulation requires no training and has high robustness, its high computation time makes it unsuitable to be integrated into conventional system models. While the neural network provides the desired short computation time, it requires a large amount of time to train and provides only low robustness to the input data. This means that the network can provide qualitatively incorrect solutions if unfavorable training data is chosen. The PINN combines the high computational speed of data-based approaches with the robustness of the model-based approach. In contrast to the EHD, training is still required, but this can be performed in a much shorter time due to the physical correlations it contains.
Acknowledgement
The project is accompanied and supported by an industry-dominated working group of the Research Fund of the Fluid Power Association in the VDMA.