Transport systems are experiencing dramatic changes worldwide. The challenges to the development of environmentally sound solutions that are economically and technically feasible on the way to the mobility of the future require intensive efforts both in basic and application-oriented research. The global trends of urbanization, modularization, electrification, and high levels of automation require the systematic development of new vehicle concepts and bodies with new degrees of freedom and requirements. Part of the Next Generation Car (NGC) family of new road vehicle concepts, developed by DLR, is the Urban Modular Vehicle (UMV), the focus of which is on urban mobility, electrification, the introduction of highly automated vehicle systems and safe body structures for urban use. The Urban Modular Vehicle’s modular system is suitable for a wide range of models from small cars to fully autonomous delivery vans. The project is aimed at developing various vehicles that incorporate the trends, technologies, and development methods of future-generation vehicles. The modularity is meant to reduce the production cost, and allow flexibility for producing different kinds of cars, from standard passenger vehicles to small transporters. The need for reducing the weight and at the same time maintaining safety, for this modular type of vehicle requires innovative solutions, as far as crash safety is concerned. Structural crashworthiness requirements are mainly met by means of crumple zones at the front and rear of the vehicle. During a collision, these zones are supposed to absorb most of the kinetic energy by means of energy absorbers such as crash boxes. Depending on the usage of the main platform and the final derivative, these energy absorbers have to be designed and optimized accordingly. The optimization of the structural behavior of the crash box requires the combination of numerous variables, and standard techniques for optimization come to their limits for solving this problem. Moreover, as crash is a highly non-linear phenomenon, it is very complex to find a crash box design that leads to an optimized behavior (absorbing a high level of energy, while minimizing the weight and satisfying the manufacturing constraints). Hence, the design process is manual, and still mostly relies on engineers’ experience and best practices established throughout the years. In parallel, Deep Learning has evolved radically over the last few years. Neural Concept has developed unique algorithms, based on geodesic convolutional neural network (GCNN), which allows the training of deep learning models by using raw 3D geometrical and simulation data. This new generation of software allows shortcutting any simulation chain through a predictive model that outputs post-processed results right from the geometry design. Moreover, this Deep Learning model is independent of any underlying parametrization and allows very high flexibility to explore the design space and find the optimum design. This work intends to use the Deep Learning methods described above, for improving the performance of the crash box, by shortening at the same time the development cycles.