In many areas of life, the safety and reliability of components are of immense importance. Whether in the construction of buildings or bridges, cars or in aerospace technology - hazards in use should be ruled out right from the design stage. Time and again, however, accidents show that there is still some catching up to do. Scientists at the Institute of Mechanics and Fluid Dynamics are therefore working on the simulation of crack propagation and structural failure as well as the behaviour of materials on an atomic and microscopic level.

These types of analyses help to identify potential weak points at an early stage and assess the impact on the overall structure. Simulation tools make it possible to optimise the design of components and structures. Various materials, shapes and configurations are evaluated in order to maximise the performance and service life of materials. The key benefit of simulations is the ability to analyse the effects of crack propagation and failure in a variety of scenarios using a computer and therefore very cost-effectively - without having to carry out real experiments. This is because testing materials and structures in the laboratory or using prototypes is expensive and time-consuming.

Professor Bernhard Eidel from the Chair of Micromechanics and Multiscale Material Modelling goes one step further. He is working with artificial neural networks, a branch of artificial intelligence. These computer programmes can be used, for example, to create a link between heterogeneous materials and their properties. Heterogeneous means that these materials consist of different components, the so-called phases, which are present in different microstructures. In order to establish this connection, the neural network must be trained.

For this purpose, microstructure geometries and the associated properties derived from experimental measurements or computer calculations are read in. By adjusting its parameters, the neural network learns to adapt its predictions more and more precisely to the learning objective until it is able to accurately predict the as yet unknown stiffnesses for new microstructures. With a new generation of neural networks, this is possible for any phase properties in any combination. Previously, this was only possible for phases with fixed properties. Materials science is now able to determine the effective stiffness of new materials in fractions of a second. Expensive and energy-intensive processes are no longer necessary.