PHYSICS-INFORMED DEEP LEARNING APPLIED TO STEADY-STATE METHANE DISPERSION

ABSTRACT – The calculation of gas dispersion is relevant for many applica-
tions in process safety. Although the gas flow can be described by the set of

Navier-Stokes equations, the numerical solution is demanding as far as the com-
putational time is concerned. The solution also requires a closure model for the

turbulence problem. We address these issues by developing a physics-informed
neural network (PINN) that considers the underlying physics of the problem by
means of the Navier-Stokes embedded in the architecture of the neural network.
An in-house CFD (Computational Fluid Dynamics) code is used to generate the
dataset. We show that the physics-informed deep learning is robust to the low

quantity of the dataset. It also generates quick solutions for gas dispersion pro-
blems, a feature of interest compared with the long time required to perform

CFD calculations.