The development of simulation tools to solve engineering problems has traditionally relied on the resolution of the physics equations in industrial problems by means of different numerical techniques, such as the finite element or the finite volume methods. Over the last 30 years, it has been difficult to identify real breakthrough technologies that go clearly beyond the state of the art of these well-known numerical tools. Generally speaking, these improvements can be classified as refinements of the original methods and the main process accelerations have been achieved by means of parallelization and massive hardware usage. In recent times, the usage of ML for CAE has arisen as a fresh proposal that can be applied to model industrial problems and some examples can be found in the literature. These references include aerodynamic surrogate models, physics-informed neural networks or reduced-order models; however, these AI applications have been developed as a proof of concept and their application has not gone beyond academic research.
The consortium has got inspiration from many academic sources and has implemented some of the most promising solutions to solve real industrial problems. The main challenges that the UPSCALE consortium has had to overcome include the development of new procedures based on ML algorithms, the integration of these toolkits in the current simulation frameworks, the generation of the training datasets and the application of these tools to real industrial problems, such as a full vehicle crash simulation including the complete battery pack or vehicle aerodynamics simulation in real time.
The main innovation of the UPSCALE project consists of applying ML to model real industrial problems and fulfilling the automotive industry standards, regarding the quality of the results, cost and robustness. Some of the ML techniques deployed in the UPSCALE project, such as ROM for non-linear mechanics, PINN or surrogate aerodynamics modelling, have been proven up to a proof of concept level, which means a TRL of 4 or 5. The consortium has the goal of developing these solvers for industrial application, which means that the consortium will bring them up to a TRL of 6 or 7.
The deployment of AI solvers for both crash and aerodynamics can have a huge impact on the performance of both simulation processes, and the time-reduction of those simulations by orders of magnitude can lead to significant improvement of the whole development process. On the one hand, real-time aerodynamics solvers will make it possible to actually use CFD as an interactive design tool and thus avoid costly design-simulation loops of days or weeks and on the other hand, ROM of batteries will allow for the first time ever the behavior of all the battery cells to be assessed in a crash event and avoid the unnecessary over-dimensioning of the battery pack. Thus, the UPSCALE project is not only accelerating the development process, but also impacts on vehicle efficiency, with lighter and more aerodynamic vehicles.
At the end of the project, some of the main project outcomes, such as: a battery ROM for crash simulation, a universal parametrezation tool for objects characterization or an aerodynamics simulator that can provide the aerodynamic forces in real time have become a reality and have shown successful results in industrial applications such as complete vehicle simulations for aerodynamics and crash.
The final implementation of this ML toolkit by the automotive industry allows a lead time reduction of 20%. Two main time savings have been identified, related with the battery pack design and the exterior vehicle design. On the one hand, front-loading virtual battery pack crash analysis allows skipping unnecessary re-design loops after the testing phase and on the other hand, vehicle design with interactive aerodynamics monitoring tools, allows a natural convergence of the design esthetic goals and the aerodynamic targets in parallel.
