Fast Multi-Physics Optimization of a Car (FMOC) is a project led by Renault, in collaboration with the French software editor ESI Group and the Ecole des Mines of Saint-Etienne. Its goal is the optimal design of a vehicle body to reach the next safety objectives (like the ones from EuroNCAP) with representative means and targets that the automakers will use in the next 3-5 years.
Between 2013 and 2014, Renault benefitted from the biggest allocation ever awarded to an industrial by PRACE. Thanks to 42 million core hours on GENCI’s (Grand Equipement National de Calcul Intensif) supercomputer, the research team has simulated, for the first time in the world at this scale, the frontal crash of a vehicle. They used a vehicle in the Renault range, the DACIA LODGY. The objective was double: Defining if it was necessary to more discretize the vehicle and testing the optimization methods by increasing the number of parameters.
Results: 5-year advance and new advanced numerical methods
The automotive engineering gained a lot of benefits from this kind of model. It was quickly used as standard on their internal computing resources. It also shows the great interest to access to PRACE resources for validating, with a 5-year advance, new advanced numerical methods, that would have been otherwise impossible to handle.
Perspectives – from machine learning technique to deep learning methods
The FMOC project has shown a potential benefit to combinatorial optimization using a big and accurate crash model, but that this association is not still compatible with the schedules of the vehicle projects due to the scalability of the crash simulation and the number of iterations of the combinatorial optimization. The explicit solver has progressed but not enough to allow everyday use of a very large crash model. The improvement way is to use a "reduced" crash model. Since the end of FMOC, RENAULT and ESI have been developing crash model reduction methods to replace the response surfaces classically used in optimization tools. The computation of a reduced model is much heavier than a response surface and requires the use of HPC but then makes it possible to carry out a combinatorial optimization in a less expensive way.
- The method developed uses a machine learning technique (Random Forest) which already gives good results and will be improved.
- Evaluation of deep learning methods is also planned.