MMH.20.003 – Towards AI-assisted multiscale analysis and optimization of designed engineering materials
A promising avenue for establishing the next generation of engineering materials lies in the precise tailoring of properties and geometric features at micro- and nanoscales. As the physical and chemical mechanisms behind macroscopic performance stem from small-scale phenomena, adopting a bottom-up approach of designing microstructural features with specific macroscopic needs in mind can lead to unprecedented improvements in material performance. With outstanding advances in high-precision manufacturing techniques such as 3D printing, the technologies necessary to turn this vision into reality are becoming available. However, fabrication advances must go hand in hand with the development of fast and accurate numerical tools for modeling multiscale materials. The accuracy and level of sophistication of computer simulations has been steadily increasing, reinforcing the vision that these virtual testing tools will soon be able to replace experimental testing campaigns. Nevertheless, further model development is needed to deal with complex failure and degradation phenomena. Furthermore, a significant obstacle remains in the way: multiscale simulations are simply too slow to be used in most design situations. We propose to overcome these obstacles by improving current multiscale analysis methods and by integrating deep learning and Bayesian machine learning in the design process. By building smart and efficient surrogate models based on state-of-the-art data assimilation techniques, high-fidelity multiscale simulations will be made fast enough to be used in realistic problems. This will allow for multiscale analysis to predict, for instance, the long-term performance of 3D-printed structures. Moreover, deep learning generative models and Bayesian relevance determination will be used to streamline optimization problems that would otherwise be intractable. In combination with accelerated multiscale modeling, machine learning based optimization will enable microscopic design of materials for optimal macroscopic performance.
Designer materials, machine learning, Multiscale modeling, optimization
|Organisation||Delft University of Technology (TUD)|
|Name||Dr.ir. Frans P. van der Meer|