Hardly any technical system today is created without the aid of computer simulations. With their help, developers can assess the success of their previous work more quickly than would be possible through the construction of gradually improved prototypes. However, if a simulation is to take several physical properties into account at the same time, things become complicated and time consuming. A research team at the University of Stuttgart is using artificial intelligence methods to make such multiphysical simulations for electrotechnical problems better and faster.
Companies can use simulations to significantly reduce the product development times. Although the object in question still goes through all the traditional phases of development - from conceptualization, design and detailing to prototype construction, testing and revision - they are all shorter. Using simulated application scenarios helps developers to assess how a product will behave in various, possibly even extreme, situations at a much earlier stage. However, such simulations can become almost arbitrarily complex, both in terms of the computing effort required, which can quickly become enormous, and for the developers, who need to have an extremely precise understanding of all parameters to ensure that the results actually reflect reality. “This becomes particularly evident in multiphysical simulations,” says Prof. Jens Anders, Head of the Institute for Smart Sensors (IIS) at the University of Stuttgart. “These consider several areas of physics simultaneously, such as a system’s thermal and mechanical behavior.” If, for example, a pump moves its piston, it heats up at the same time. Thus, not only does the shape of the magnetic field in which the rotor rotates have a significant bearing on the behavior of an electric motor, but also the temperature the motor reaches during operation. Moreover, there are interdependencies between the various sub-areas. In terms of the motor, for example, this means that increasing electrical currents also generate more heat.
“For example, to simulate the behavior of such a motor, we would have to calculate the temperature and the magnetic field at many points,” explains Anders. “If an unlimited amount of computing power were available, these points could be placed over the entire model of the engine like a close-meshed network.” But because computing power still has to be used economically today, despite increasingly improving computer technology, this network has to be plotted as coarsely as possible. “Only as accurately as necessary is the rule when simulating,” says Anders. The choice between computing speed and accuracy involves constant compromise - but no important physical effect should be missed.
The experts know, for example, that they have to plot the mesh nodes very closely to calculate the magnetic field, especially along corners and edges. On the other hand, the calculation points in the metal body of the motor can be spaced further apart. The temperature calculation is different: corners and edges or the air gap of the motor are of very little interest in this context, on the other hand, the metal body dissipates a large part of the generated heat, so that a close-meshed network is required for calculations. “Different networks are needed for each physical sub-area, but they have to be adapted to each other in such a way that the results make sense overall,” says Anders.
Due to this complex starting position, science and industry have not yet fully exploited the potential of multiphysical simulations. This was the reason why the IIS and the Institute of Industrial Automation and Software Engineering (IAS) at the University of Stuttgart launched a joint research project a few years ago with funding from the German Research Foundation (DFG). “Most development engineers not only have to deal with multiphysical simulations but also many other tasks,” says Prof. Michael Weyrich, Head of the IAS. So it can happen that they fail to set up complex multiphysical simulations properly, require many attempts or even need external help from specialists. “This costs at least time and money and, in extreme cases can even leads to the decision to abandon multiphysical simulations altogether,” said Weyrich.
The researchers at both institutes want to eliminate this hurdle. “We have provided the user with an intelligent assistant for multiphysical simulations of electrotechnical systems, which executes various tasks independently, so that the user no longer has to have such a profound knowledge of the topic,” Weyrich explains. “Based on available simulation results that make physical sense, the assistant recommends approaches to existing simulation problems.” The project participants created a database for this purpose in which the assistant can search for similar known cases for a given subproblem of the simulation in hand. “If it finds a corresponding match, it uses it as a basis for proposing solutions to the user for the relavant subproblem,” says Weyrich.
Intelligent assistance for experts
In fact, the structure and operation of this assistant are much more complex than described here. It actually coordinates several subordinate assistants, each of which is responsible for a specific subtask such as mechanics or temperature. In addition, there are always various options that the downstream assistants can choose for themselves to select the most resource- efficient assistant for achieving the objective.
The project participants trained the assistance system using machine learning methods, prototypically implemented on the simulation of a microwave oven in which a vessel is to be heated with water. The research team played through various cases based on the hundred or so available solution approaches to find the best solution strategy for the simulation.This resulted in a fourfold decrease in computing times. To optimize the computing time, the software can independently distribute the computing load of a simulation to different computers, if one fails. In addition, the assistant is able to optimize the network mesh sizes of the physical subareas.
“For a development engineer to accept such an assistant, the system must of course only identify the appropriate old cases in its database,” says Weyrich. To achieve this, the project participants defined a set of metrics. I.e., mathematical measures of how well the old cases in the database fit the current problem. “Then we trained the algorithm by considering how well a suggested case fits the current problem,” explains Weyrich. Encouraged by the results, Anders and Weyrich and their teams now want to demonstrate that intelligent assistants are helpful in solving multiphysical simulations in a transfer project using case studies with real data. Warning messages to alert the user to critical points in the simulation and automatic mesh width determination could be the first practical benefits, which could ultimately lead to faster, more reliable results. The two University of Stuttgart Institutes will implement the project in collaboration with Comsol, a Swedish provider of software for multiphysical simulations.
Jens Anders uses a simple numbers game to illustrate the importance of better multiphysical simulations: “Assume 30 variants have to be simulated for the development of a given product. If you only manage to run two simulations a week, you could just as well build a physical prototype and study it in the 15 weeks it would take. If, on the other hand, you could run 15 simulations per week, you would save on prototype construction because the simulations would be completed in just two weeks.” Not only would this shorten the overall development process, but certain individual steps could be eliminated entirely, saving time and money. “But for this to succeed beyond those companies that employ their own multiphysical specialists in their development departments,” says Weyrich, “concepts such as our intelligent assistant must find their way into mainstream industrial practice.”