After several decades of breakthroughs in numerical modelling of complex physical phenomena, including calibration and validation on real case studies, attempts to quantify uncertainties and solve ill-posed inverse from experimental data, Machine Learning techniques appear as a general-purpose substitute to these numerical techniques in our fields of expertise. Meanwhile, most of our industrial partners are eager not to miss this AI turn, though a bit reluctant about making decisions on AI-based predictions. Finally, most of our students feel they cannot ignore these techniques, which are likely to be their routine job soon.
This lecture will report on some experiments done in earthquake engineering and non-destructive testing attempting at combining high-fidelity Physics-Based Simulations with Machine Learning methods such as Generative Adversarial Networks and Variational Auto-Encoders. This talk will not enter technical details but will try to emphasize generic tools, achievements, difficulties, and prospects for such combined approaches, especially when bridging the gap between experimentalists and modellers.
