Rafael Teloli | Assistant Professor, Department of Applied Mechanics
FEMTO-ST Institute, Besançon, France
June 21, 2023 | 11h00-12h00 (Lisbon Time) | Room U.0.6 | https://videoconf-colibri.zoom.us/j/93285885760 | Lusófona University, Lisbon, Portugal
Abstract: Model identification plays a crucial role in understanding complex systems, yet it remains a challenging task due to uncertainties and the presence of multiple potential models. In this seminar, we propose two innovative approaches that harness the use of artificial intelligence (AI) methods to enhance model identification. The first approach combines Reinforcement Learning with Approximate Bayesian Computation (RL-ABC) to address the challenges of identifying the optimal model from a family of possibilities using data corrupted by measurement uncertainties. By calibrating models within confidence intervals and accounting for uncertainties, the methodology enables the selection of the model that best represents the available data. This approach is demonstrated using viscoelastic materials characterized by frequency-temperature dependent properties obtained from Dynamic Mechanical Analysis (DMA) data. In the second method, we employ a Physics Informed Neural Network (PINN) framework to identify a neural network that not only captures the physics of a Euler-Bernoulli beam but also estimates model parameters. By incorporating the partial differential equation into the loss function, the PINN framework seamlessly integrates physics-based constraints into the neural network training. We validate our methodology using numerical data from a finite element model and subsequently verify its effectiveness using experimental data. At the end, this seminar highlights the opportunities and advancements offered by artificial intelligence in the realm of model identification.