• Complex Systems
• Fluid Dynamics
• Machine Learning

High-pressure homogenizing valves are critical in a variety
of industrial applications, including food processing and
pharmaceuticals, where precise fluid dynamics under extreme
pressures must be achieved to ensure performance and reliability.
Traditional Computational Fluid Dynamics (CFD) methods, while
essential, often face convergence issues and require extensive
computational resources to accurately simulate the typical small-scale
and high-pressure working regimes of these valves.

In this thesis is proposes a novel approach of this problem that uses
Physics-Informed
Generative Adversarial Networks (GANs) to enhance the accuracy
and efficiency of CFD simulations for these valves. The study
evaluates multiple GAN architectures, including standard GANs,
Conditional GANs (CGANs), and a Conditional Wasserstein GAN with
Gradient Penalty (WGAN-GP), incorporating physics-informed
constraints that embed conservation laws directly into the
learning process. Statistical methods were employed to verify
whether these conservation laws were violated in the generated
data, ensuring in this way the consistency with physical
principles.

The CFD simulations used for this work were provided by the
company overseeing the project. These simulations were rendered
in 2D sections, focusing on pressure and velocity fields, with a
dataset composed of 9 training simulations and 2 simulations
withheld for testing. The quality of the generated simulations
was assessed using the Fréchet Inception Distance (FID) to
compare the real and generated data distributions.

Our results show that the Conditional WGAN-GP, though requiring
longer training times and greater computational resources,
produced simulations that adhered most closely to the physical
constraints, accurately capturing the complex fluid behavior
within the valve. In contrast, standard GANs delivered faster
training times but were slightly less precise, making them better
suited for rapid, preliminary assessments.

Additionally, a non-physics-informed pix2pix model was employed
to generate valve simulations based on specific geometries.
Despite the absence of explicit physical constraints, the pix2pix
model exhibited robust performance, even when generating
out-of-sample simulations. Its speed and adaptability make it
a valuable tool for early design phases, where quick, accurate
simulations are vital.

Although these GAN models are not intended to replace traditional
CFD simulations, they show considerable promise as supplementary
tools. They can streamline the simulation workflow, offering
faster results and meaningful insights that enhance overall
efficiency in industrial CFD applications. Future research will
focus on integrating physics-based constraints into the pix2pix
model to combine computational speed with physical consistency,
as well as expanding its application to 3D simulations and fluid
motion generation using the Navier-Stokes equations as part of
the physics-informed framework.