Many questions regarding transient optimization and control in powertrain development are very difficult to solve using conventional controllers. These instruments should pursue long-term goals in a complex environment, but they have started to come up against their limits. A good example of this is the design of an operating strategy in a hybrid powertrain. At each moment, a decision must be made as to whether it is worth using the electrical energy available in the battery for drive power, or whether it is better to save this for a more favorable moment in the future.

Such tasks often involve incorporating 'human experience' into the solution using a rule-based approach. However, replacing the engineer's 'gut feeling' with artificial intelligence can further increase efficiency and streamline the development process.

Most artificial intelligence methods require "big data", i.e. a large pool of data to learn from. Although there is a lot of data available in powertrain development, the term "a lot" must be used with care. If, for example, the EGR and heating strategy for RDE requirements were optimized using artificial intelligence, 100 RDE test runs would not be enough for an approach like AlphaGo Zero's "reinforcement learning". In this context, big data means 100,000 or 1,000,000 test runs.

Models can be used to boost the potential for combining data from the engine and powertrain test bench with artificial intelligence methods. In this case, these act as a multiplier for big data. Calibrated using 50 real test runs, models can simulate 50,000 or 500,000 virtual test runs – generating big data for artificial intelligence.

Caution: big data cannot just be a huge mass of data. The models used to generate big data must provide high-quality results and a high prediction capability, with acceptable calculation times. 

Many simulation models used for product design are based on a very fine, local discretization. This is absolutely essential, especially for tasks such as designing and evaluating geometries. This results in high calculation times, which can quickly become 10² to 10⁶ times slower compared to real time.

For transient processes where the real-time process already lasts between 2 and 120 minutes, optimization using these approaches would take several years. At the same time, solving these kind of questions with AI approaches based on big data would exceed the computing power of today's high-performance clusters by several orders of magnitude.

 AI approaches which do not require big data are therefore of extreme interest for such applications. For example, heuristic searches, which are also applied to route planning.

At the same time, FKFS has a long tradition of programming and developing simulation software, as well as developing and dimensioning powertrains and engines. To some extent, we have always been an interface between computer scientists and engineers. 

We also feel at home in both areas when using artificial intelligence technologies. It is precisely this interface function which enables us to advise you on the possibilities when using new technologies, as well as limits and uncertainties. Our team will support you with strategic orientation and investment planning, as well as with developing the methodology.

 We look forward to hearing from you!