Surrogate Models

Speeding Up Simulation for Automotive Industries

The numerical simulation of parts of car engines may be quite time consuming. For running hardware-in-the-loop systems, computation times must be reduced by a factor 100.

Challenge overview

By connecting automotive hardware and simulation software in a testbed environment, the requirements on computation time get extremely high, as every engine cycle should be replicated by numerical simulation of the parts under consideration within real-time. The idea how to achieve this is to introduce surrogate models in the form of support vector machines.

The problem

For the various components of a car engine, the powertrain and the virtual driver system, sophisticated software tools are available to study, e.g. fuel consumption, exhaust gas aftertreatment or optimal gearing. If a hardware-in-the-loop-system is used, some components are realized in hardware and some of them as simulation programs to study, e.g. various designs of hybrid engines.

In such a combined testbed environment, it is essential that the simulation software runs at least as fast as the hardware meaning that every millisecond of real time must be simulated in not more than a millisecond.

Results and achievements

The approach which was realised in the project was based on surrogate models, here in the form of so-called support vector machines. These surrogate models aim to evaluate a function which is easily to calculate instead of solving numerically a partial differential equation. They require a training phase during which the shape of the ansatz function and its parameters are determined. After this training phase, which can be carried out offline, the surrogate model may achieve speed-ups of a factor 1000 and more compared to the full numerical simulation.

Torque measurements (black dots) and surface obtained from the surrogate model using a support vector machine (SVM).

Further Reading:

Roman Heinzle: Machine learning methods and their application to realtime engine simulation. Ph.D. thesis, Johannes Kepler University, 2009.