Abstract: In this talk we will review some recent developments around thermodynamics-informed (graph) neural networks. Unlike other approaches (particularly, PINNs) this approach is useful when no previous knowledge is available on the physics of the phenomenon under scrutiny. In the absence of such a knowledge, thermodynamics constitutes the most useful tool for the development of inductive biases. The resulting method has shown to improve the accuracy over black-box neural networks. We will discuss its application to closed as well as open systems and how it could become a useful tool in the development of cognitive twins.
Bio: Elías Cueto is a professor of continuum mechanics at the University of Zaragoza. His research is devoted to the development of advanced numerical strategies for complex phenomena. In particular, in the last years, he has worked in model order reduction techniques and real-time simulation for computational surgery and augmented reality applications. His work has been recognized with the J. C. Simo award of the Spanish Society of Numerical Methods in Engineering, the European Scientific Association of Material Forming (ESAFORM) Scientific Prize, and the O.C. Zienkiewicz prize of the European Community on Computational Methods in Applied Sciences, ECCOMAS, among others. He is a fellow of the EAMBES and IACM societies and the president of the Spanish Society of Computational Mechanics and Computational Engineering (SEMNI).