Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Air Flow Predictions in Converging-Diverging Nozzles Using Fourier Neural Operators

Karlsson, Max LU (2025) MVKM01 20251
Department of Energy Sciences
Abstract
This thesis investigates the potential of using Fourier neural operators to predict subsonic and supersonic flows in converging-diverging nozzles, which includes complex flow phenomena such as shock waves. Two different data sets were generated using the Euler equations for fluid flow, one based on analytical relations for a quasi one-dimensional nozzle and another one which was created by numerically solving the governing equations on a planar two-dimensional geometry. Both data sets were generated with a wide range of input conditions, including diverse nozzle shapes and various combinations of boundary conditions. Based entirely on the set of input conditions, the model could be trained to output the solution fields for temperature,... (More)
This thesis investigates the potential of using Fourier neural operators to predict subsonic and supersonic flows in converging-diverging nozzles, which includes complex flow phenomena such as shock waves. Two different data sets were generated using the Euler equations for fluid flow, one based on analytical relations for a quasi one-dimensional nozzle and another one which was created by numerically solving the governing equations on a planar two-dimensional geometry. Both data sets were generated with a wide range of input conditions, including diverse nozzle shapes and various combinations of boundary conditions. Based entirely on the set of input conditions, the model could be trained to output the solution fields for temperature, pressure and velocity.

The trained Fourier neural operator's final results showcase great generalization across various combinations of input conditions and nozzle shapes, while accurately distinguishing between different flow regimes in both one and two dimensions. Furthermore, the presence of normal shock waves as well as their magnitude and location in the diverging section of the nozzle were predicted with high accuracy.

Lastly, by including the conservation of mass and energy as a part of the loss function, it was demonstrated that the generalization could be drastically improved for small data sets, essentially filling potential gaps in data with prior knowledge of the problem. However, for larger data sets, the addition of physics proved to be rather ineffective in terms of improving the generalization. (Less)
Popular Abstract
In many high-speed aerodynamic applications such as rocket engines and wind tunnels, a so called converging-diverging nozzle is used, which is a tube-shaped part where the cross-sectional area first decreases and then increases. Such nozzles are in particular used to accelerate air to supersonic speeds, which means that it flows faster than the speed of sound at roughly 343 m/s. At such high velocities, the flow can exhibit complex phenomena such as shock waves, which are drastic changes of flow properties such as temperature, pressure and velocity over an extremely thin region of the gas (typically around one ten-thousandth of a millimeter for air).

The flow of air through the nozzle is primarily driven by a set of input conditions... (More)
In many high-speed aerodynamic applications such as rocket engines and wind tunnels, a so called converging-diverging nozzle is used, which is a tube-shaped part where the cross-sectional area first decreases and then increases. Such nozzles are in particular used to accelerate air to supersonic speeds, which means that it flows faster than the speed of sound at roughly 343 m/s. At such high velocities, the flow can exhibit complex phenomena such as shock waves, which are drastic changes of flow properties such as temperature, pressure and velocity over an extremely thin region of the gas (typically around one ten-thousandth of a millimeter for air).

The flow of air through the nozzle is primarily driven by a set of input conditions which includes a difference in pressure over the nozzle which pushes the gas forward, an initial temperature of the gas and finally the shape of the nozzle. In order to design fully optimized nozzles without unwanted effects such as shock waves, it is crucial to be able to predict how the flow will develop based on the initial settings. In most cases there exist no exact solutions to the flows and the classical method of producing a solution instead involves dividing the geometry into very tiny parts called cells. The solution can then be approximated in each cell separately by using numerical methods on a set of physical laws that govern the flow of any gas. These traditional simulation methods yield approximate solutions that become more accurate as the number of cells increases, thus making them computationally expensive and extremely time-consuming to perform.

This thesis explores an alternative approach to solving these problems using newly developed methods within artificial intelligence (AI), specifically a subclass called neural operator. Such a model aims to predict the air flow within the nozzle by learning patterns from a large data set that contains a wide range of input conditions, geometries and their corresponding solutions. The generation of the data itself has to be performed with classical methods and can therefore be quite time-consuming, however once the model has been trained, it can make predictions for various combinations of inputs and geometries within a few seconds.

The results showcase that neural operators are capable of making flow predictions with great accuracy, while remaining flexible with regard to initial conditions and various nozzle shapes. Furthermore, the model precisely predicts the presence of complex flow phenomena like shock waves and captures key features such as magnitude and location. The promising results highlight a potential alternative to classical methods, which when necessary could be used to facilitate fast feedback loops for design and development of nozzles. (Less)
Popular Abstract (Swedish)
I många aerodynamiska höghastighets-applikationer som t.ex. raketmotorer och vindtunnlar används ett så kallat konvergerande-divergerande munstycke, vilket är en rörformad del där tvärsnittsarean först minskar och sedan ökar. Sådana munstycken används ofta för att accelerera luft till supersoniska hastigheter, vilket innebär att det flödar snabbare än ljudets hastighet som är ungefär 343 m/s. Vid så pass höga hastigheter kan det förekomma komplexa flödesfenomen inuti munstycket som exempelvis chockvågor, där egenskaper hos flödet som t.ex. temperatur, tryck och hastighet drastiskt förändras över en väldigt tunn region av gasen (typiskt runt en tiotusendels millimeter för luft).

Flödet av luft genom munstycket drivs primärt av en... (More)
I många aerodynamiska höghastighets-applikationer som t.ex. raketmotorer och vindtunnlar används ett så kallat konvergerande-divergerande munstycke, vilket är en rörformad del där tvärsnittsarean först minskar och sedan ökar. Sådana munstycken används ofta för att accelerera luft till supersoniska hastigheter, vilket innebär att det flödar snabbare än ljudets hastighet som är ungefär 343 m/s. Vid så pass höga hastigheter kan det förekomma komplexa flödesfenomen inuti munstycket som exempelvis chockvågor, där egenskaper hos flödet som t.ex. temperatur, tryck och hastighet drastiskt förändras över en väldigt tunn region av gasen (typiskt runt en tiotusendels millimeter för luft).

Flödet av luft genom munstycket drivs primärt av en uppsättning ingångs-tillstånd som inkluderar en tryckskillnad över tuben vilket trycker gasen framåt, en initial temperatur hos gasen och till sist också den geometriska formen på munstycket. För att kunna designa fullt optimerade munstycken utan oönskade effekter som till exempel chockvågor, är det viktigt att kunna förutspå hur flödet kommer utvecklas beroende på de initiala förutsättningarna. I de flesta fall finns inga exakta lösningar till utvecklingen av flödet och de klassiska metoderna för att ta fram en lösning går istället ut på att dela upp geometrin i väldigt många små delar som kallas celler. Lösningen kan då approximeras i varje enskild cell genom att utnyttja numeriska metoder för att försöka lösa en uppsättning fysiska ekvationer som styr flöden av gaser. Dessa traditionella simuleringsmetoder ger lösningar som blir mer noggranna desto fler celler som används, vilket innebär att de också ofta är beräkningsmässigt dyra och därmed väldigt tidskrävande att utföra.

Denna uppsats utforskar ett alternativt tillvägagångssätt till dessa problem, genom att använda nya metoder inom artificiell intelligens (AI), mer specifikt något som kallas neurala operatorer. Modellen försöker förutspå luftflödet genom att lära sig mönster från en stor uppsättning data som innehåller många olika ingångsvariabler, geometrier och deras motsvarande lösningar över hela munstycket. Genereringen av all data måste i sig utföras med hjälp av klassiska metoder vilket kan bli väldigt tidskrävande, men när modellen väl tränats kan den prediktera lösningar för en rad olika ingångsvariabler och geometrier inom ett par sekunder.

Resultaten visar att neurala operatorer kan uppskatta flödeslösningar med god noggrannhet, samtidigt som modellen är extremt flexibel och kan anpassa sig väl till olika ingångsvariabler samt former på munstycken. Dessutom kan modellen också med stor säkerhet fånga mer komplexa flödesfenomen som chockvågor och förutspår var de kommer uppstå samt hur drastiska de är. Dessa resultat är väldigt lovande och lyfter fram potentialen för alternativ till klassiska metoder, som vid behov kan främja snabba återkopplingsloopar för design och utveckling av munstycken. (Less)
Please use this url to cite or link to this publication:
author
Karlsson, Max LU
supervisor
organization
course
MVKM01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Fourier Neural Operator, Deep Learning, Supersonic Flow, Subsonic Flow, Normal Shock Wave, Converging-Diverging Nozzle, Computational Fluid Dynamics (CFD)
report number
ISRN LUTMDN/TMPH-25/5637-SE
ISSN
0282-1990
language
English
id
9200074
date added to LUP
2025-06-17 08:23:58
date last changed
2025-06-17 08:23:58
@misc{9200074,
  abstract     = {{This thesis investigates the potential of using Fourier neural operators to predict subsonic and supersonic flows in converging-diverging nozzles, which includes complex flow phenomena such as shock waves. Two different data sets were generated using the Euler equations for fluid flow, one based on analytical relations for a quasi one-dimensional nozzle and another one which was created by numerically solving the governing equations on a planar two-dimensional geometry. Both data sets were generated with a wide range of input conditions, including diverse nozzle shapes and various combinations of boundary conditions. Based entirely on the set of input conditions, the model could be trained to output the solution fields for temperature, pressure and velocity.

The trained Fourier neural operator's final results showcase great generalization across various combinations of input conditions and nozzle shapes, while accurately distinguishing between different flow regimes in both one and two dimensions. Furthermore, the presence of normal shock waves as well as their magnitude and location in the diverging section of the nozzle were predicted with high accuracy.

Lastly, by including the conservation of mass and energy as a part of the loss function, it was demonstrated that the generalization could be drastically improved for small data sets, essentially filling potential gaps in data with prior knowledge of the problem. However, for larger data sets, the addition of physics proved to be rather ineffective in terms of improving the generalization.}},
  author       = {{Karlsson, Max}},
  issn         = {{0282-1990}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Air Flow Predictions in Converging-Diverging Nozzles Using Fourier Neural Operators}},
  year         = {{2025}},
}