A state-of-the-art of physics-informed neural networks in engineering

Abstract

Machine learning techniques have gained space in the industrial scenario as a tool to convert the increasing flux of information (data) in process improvement. Among these techniques, neural networks has got much attention due to their universal approximators capacity, of which performance can be improved by providing previous physical knowledge: one has, therefore, the development of the so called Physicsinformed neural networks (PINN). In such context and having noticed a “gap” in the works related on this topics and in the diffusion of this theme in the School of Chemistry, this work proposes a state-of-the-art of the mentioned technique. Particular interesting concerning PINN in fluid mechanics and heat transfer has been noticed. Moreover, PINN have been pointed as important tools for solving forward and inverse problems. Finally, through practical examples, this work has shown the use of neural networks for solving one particular example in chemical engineering without informing the physics of the problem (obtaining the friction factor) and using the differential equation that describes it (solving the 1D heat diffusion equation).