 |
Center for Nonlinear Studies |
Efficient multiscale modeling strategies are useful for several problems of engineering interest. Some examples include dynamic failure of fiber-reinforced composites, additively manufactured materials, detonation in heterogeneous energetic materials, response of systems involving multiphase turbulent flows, problems involving shock-droplet and shock-particle interactions, among others. This work presents a multiscale modeling framework using machine learning techniques. Closures to the macroscale computational model are provided by a Bayesian Kriging algorithm, trained using ensembles of high-fidelity meso-scale computations. The machine-learning based multiscale modeling approach is demonstrated using two example problems, viz. the interaction of a cloud of particles with a shock and shock-induced detonation of heterogeneous explosives. In both cases, macroscale models closed using machine-learned sub-grid scale information are shown to be in good agreement with experimental predictions. Additional strategies such as multi-fidelity techniques to reduce the computational burden of generating training data will also be discussed. In addition to scale-bridging methods, quantification of the propagating uncertainty from the meso to the macro-scales will also be discussed. In conclusion, the presentation will demonstrate a machine-learning based multiscale modeling framework, applicable to a wide variety of problems with engineering application.
Host: Michael Chertkov