A Convolutional Neural Network Based Framework for Linear Fluid Dynamics

Kwame Agyei‑Baah 1 , Muhammad Rizwanur Rahman 2 and E. R. Smith 1 · 2026

Read A Convolutional Neural Network Based Framework for Linear Fluid Dynamics by Kwame Agyei‑Baah 1 , Muhammad Rizwanur Rahman 2 and E. R. Smith 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-11946.txt.

Introduction

Kwame Agyei-Baah, Muhammad Rizwanur Rahman, and E. R. Smith frame a neural network route to linear fluid dynamics. Fluid flow rules acts at scales from a breakfast table to cosmic and quantum systems. That vast range is why the field is called a core science.

Their frame rests on convolutional neural networks. Such nets learn traits from grid data, which suits flow fields. A learned tool could solve some flow tasks much faster than step-by-step maths. The work sets out the design and its use for linear flow cases. The frame tests a fast route to explain text model calls.

Catalog record

CREATORKwame Agyei‑Baah 1 , Muhammad Rizwanur Rahman 2 and E. R. Smith 1
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-11946

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Convolutional Neural Network Based Framework for Linear Fluid Dynamics Kwame Agyei‑Baah 1 , Muhammad Rizwanur Rahman 2 and E. R. Smith 1 Abstract 1 Abstract Fluid dynamics, for its strength in describing physical phenomena across vastly different scales from the cheerios effect on the breakfast table to the evolution of cosmic and quantum systems, has been called the ‘queen mother’ of science ( 5 ) . However, a central challenge remains: ensuring the generalisability, interpretability and reliability of the machine learned models when applied to physical systems. To address this, we present a transparent approach that provides insights into how data-driven fluids dynamics and machine learning (ML) work. This is achieved by training a convolutional neural network (CNN), on data from a simple laminar fluid flow, to behave as an operator that exactly matches the finite-difference numerics. Importantly, the model demonstrates strong generalisation capability by reproducing the dynamics for a wide range of distinct and unseen flow conditions within the same flow category. The CNN learns the forward Euler three-point stencil weights, capturing physical principles such as consistency and symmetry despite having only three tunable weights. Going beyond pure numerical training (numCNN), the approach is shown to work when trained on analytical (anCNN) and even molecular dynamics (mdCNN) data. In some cases, the physics is not captured, and thanks to the simple and interpretable form, these CNNs provide insight into the limits, pitfalls and best practice of data-driven fluid models. Because the approach is based on finite‑difference operators and demonstrated with diffusive flow, it naturally extends to many structured‑grid computational fluid dynamics (CFD) problems, including turbulent, multiphase, and multiscale flows. Keywords: Machine Learning, Numerical schemes, Finite difference, Convolutional Neural networks, Fluid dynamics 2 Introduction From molten rock flowing in the volcanic eruptions to the quark-gluon plasma within protons and neutrons, and even to the phenomena at the event horizon of a black hole ( 36 ) , fluid mechanics offer a unifying language to describe how matters flow. At its core, fluid mechanics seeks to understand the motion of liquids and gases, most often through the solution of the Navier–Stokes equations. For decades, computational methods have been employed to solve these equations across a wide range of fluid mechanical problems. However, the application of traditional computational fluid dynamics (CFD) techniques for tackling realistic flows thus particularly those involving complex geometries, chemical agents e.g., surfactants and additives, and the resulting non-linearities, turbulence, or multiphase interactions remains computationally expensive, and hence limited. Recent advances in machine learning (ML) offer a paradigm shift, with algorithms capable of processing and extracting information from large dataset, automating feature detection and augmenting human domain knowledge ( 4 ) . Instead of solely relying on solving the governing equations which are limited by their inherent assumptions, and rare to find for the aforementioned complex flows, machine learning offers a new way to handle these problems by finding patterns in data. This way, machine learning can complement traditional fluid mechanical approaches with data-informed predictions, reveal previously uncharacterised non-linearities, and accelerate the discovery of hidden flow-physics. A substantial amount of research in machine learning continues to advance across various disciplines, including healthcare, finance, marketing, and engineering. And the field of fluid mechanics is no exception. Far from being overlooked, fluid mechanics community has witnessed significant progress as researchers increasingly integrate various machine learning techniques to solve complex fluid flow problems. 37 described rapid advancements in CFD driven by machine learning, highlighting techniques such as proper orthogonal decomposition and autoencoders that help manage the complexity of fluid flow data. In a comprehensive review, 4 discussed a variety of frameworks aimed at understanding, modelling, optimising, and controlling fluid flows. 38 further elaborated these concepts examining techniques such as Physics-Informed Neural Networks (PINNs), Deep Operator Networks (Deep O-Nets), Convolutional Neural Networks (CNNs), and Transformers; all of which are getting increasing acceptance and popularity in computational fluid dynamics. A practical application of machine learning is seen where a firefly machine learning algorithm is used as a deep learning flood forecasting tool for a watershed area in Tennessee, USA ( 11 ) . In particular the introduction of PINNs by 29 has attracted considerable attention for their ability to embed the governing equations (PDEs) directly into the neural network’s loss function. Combining neural networks with the laws of physics promises a paradigm shift in unearthing numerous problems across scales concerning fluid flows. Among numerous investigations in the recent years concerning PINNs, 8 employed a transformer-based encoder-decoder network for the prediction of transonic flow over supercritical airfoils, ingeniously encoding the geometric input with diverse information points. 31 focused on turbulence modelling through data assimilation and machine learning for separated flows over airfoils. 1 combined the concept of auto-encoders with other machine learning algorithms in studying the flow behind a cylinder providing a low-dimensional model (a probabilistic flow prediction). 6 used PINNs framework to tackle a heat transfer problem which cannot be handled by traditional computational methods. 18 used neural operators via graph kernel networks to learn mappings between function spaces, enabling mesh-free generalisation across discretisations. The use of Nyström approximation, enabled efficient handling of unstructured grids that outperformed mesh-dependent neural networks. The method supported sparse data, semi-supervised learning, and geometry transfer; e.g., across airfoil shapes without interpolation. 26 proposed a framework that relies on Gaussian Process, a powerful tool that is used for probabilistic inference over functions that balances between model complexity and data fitting. This approach is applied to canonical problems including the Navier-Stokes equations, the Schrödinger equation and the time dependent linear fractional equations. Their work ( 27 ; 28 ) facilitated the development of models capable of handling complex domains without requiring large quantities of data. Initial exploratory investigations in the present study employed artificial neural networks (ANNs) to solve the one-dimensional (1D) heat equation for a range of viscosities as a baseline to understand how neural networks capture flow dynamics. Feature engineering was applied to embed physical information into the model. However, the model’s marked failure to predict beyond trained viscosity ranges highlighted the limitations of purely data-driven approaches. To address this, Physics-Informed Neural Networks (PINNs) were explored by incorporating residuals of the 1D heat equation, along with boundary and initial conditions, directly into the loss function. While this improved physical consistency, the model still exhibited instability and failed to generalise reliably beyond the training range, an observation consistent with previous work ( 10 ) . With a fully converged PINNs architecture, 3 too obtained accurate predictions only within the immediate proximity of the training domain, and the model poorly performed outside the domain. These underscore a broader gap in the field where most machine learning models for CFD remain tailored to specific flow conditions or geometrical settings, limiting their general applicability. Worse still, the black box nature means we do not know how general our model is. Hybrid approaches coupling machine learning with traditional numerical solvers despite their potential to enhance accuracy and efficiency in solving complex flows are only at their initial phase of development. This can be seen in 25 where the authors proposed a tool for mixing the information of physical laws and scientific models with data-driven machine learning approaches for universal differential equations (UDEs). In

Read the complete text