The evolution of powertrain design is an increasingly complex problem in automotive engineering. Traditional methods can no longer cope with the multitude of parameters and constraints associated with the increase in myriad hybrid and electric vehicle configurations. We’re building solutions that will address these challenges using machine learning that is inherently suited to solve them.
Accelerated ECU Calibration using machine learning.
Our ECU Calibration solution automatically designs the engine data collection process, simultaneously developing a deep understanding of that engine’s performance characteristics and real-world operating constraints to ensure safe operation.
By using our machine learning technology and optimisation algorithms, we handle vast amounts of high complexity, highly noisy data to find the optimal operating profile. As a result we can calibrate an Engine and EV Control ECU in a fraction of the time of existing approaches, saving our customers time and resources, whilst making the entire process more efficient and sustainable.
Mazda: Powertrain optimisation using Secondmind
We’re collaborating with Mazda to reduce ECU calibration time and manage the ever increasing engine complexity for their industry-leading fuel efficient engines.
A cornerstone of prototyping, used in domains such as aerodynamics or stress tests and commonly by automotive manufacturers to assess and refine the chassis design of vehicles. Simulation can accurately predict results of crash tests while incurring a fraction of the cost compared to building and damaging prototypes. It allows for more data collection at less cost, for parameter space to be explored without the constraints of prototype manufacturing, and offers various fidelity vs cost tradeoffs. Our approach leverages these specificities to maximise the impact numerical simulators can have on overall design and calibration.
We use Gaussian process models to capture the input-output relationship between the parameters of a system and the measurement of interest. Gaussian process models are well-suited for problems with a large number of parameters (say up to 50), and they offer good predictive accuracy and uncertainty quantification. Our specific breeds of Gaussian process models support non-stationary noise and can handle data sets of any size. Furthermore, they can adapt to the complexity of the phenomenon being modelled. This is particularly relevant in powertrain optimisation, where simpler models sometimes better describe the physics of a subset of the system under study.