Prediction of acoustic field in 1-D uniform duct with varying mean flow and temperature using neural networks

Veerababu Dharanalakota † and Prasanta K. Ghosh , ‡ · 2025

Read Prediction of acoustic field in 1-D uniform duct with varying mean flow and temperature using neural networks by Veerababu Dharanalakota † and Prasanta K. Ghosh , ‡ (2025) 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-2507-22370.txt.

Introduction

Dharanalakota and Ghosh used neural networks in twenty twenty-five to predict sound fields inside a one-dimensional duct. The duct carries a medium whose flow and temperature vary along its length, changing how sound travels. Physics-based neural networks offer an alternative to classic numerical solvers here.

The authors first derive the governing equation for sound in such a duct. They then turn the problem into an optimization task the network can solve. The network predicts both sound pressure and particle velocity, the two state variables of the field. Its results were checked against a traditional Runge-Kutta solver. The study also shows how temperature gradients reshape the sound field, and demonstrates transfer learning and automatic differentiation for acoustics.

Catalog record

CREATORVeerababu Dharanalakota † and Prasanta K. Ghosh , ‡
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-22370

Browse the full library · About the Tower of Babel

Full text

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

Prediction of acoustic field in 1-D uniform duct with varying mean flow and temperature using neural networks Veerababu Dharanalakota † and Prasanta K. Ghosh , ‡ Abstract Neural networks constrained by the physical laws emerged as an alternate numerical tool. In this paper, the governing equation that represents the propagation of sound inside a one-dimensional duct carrying a heterogeneous medium is derived. The problem is converted into an unconstrained optimization problem and solved using neural networks. Both the acoustic state variables: acoustic pressure and particle velocity are predicted and validated with the traditional Runge-Kutta solver. The effect of the temperature gradient on the acoustic field is studied. Utilization of machine learning techniques such as transfer learning and automatic differentiation for acoustic applications is demonstrated. \catchline keywords Duct acoustics; Unconstrained optimization; Transfer learning; Complex-valued acoustic field. {history} 1 Introduction Traditionally, acoustics has relied heavily on theoretical models and empirical data to understand and predict sound behaviors in various environments [ 1 , 2 ] . However, the advent of big data and advanced machine learning algorithms has opened new avenues for research and application, fundamentally transforming how acoustic phenomena are analyzed and utilized. Data-driven methods in acoustics leverage vast amounts of data to train models that can uncover patterns and relationships that are often too complex for traditional analytical methods. For instance, machine learning algorithms can be trained on large datasets of acoustic signals to identify and classify sounds with high precision, leading to advancements in fields such as speech [ 3 , 4 ] , architectural acoustics [ 5 ] , environmental noise monitoring [ 6 , 7 ] , active noise control [ 8 , 9 ] , and structural health monitoring [ 10 , 11 ] . Despite these advancements, the application of data-driven methods in acoustics is not without challenges. Without high-quality and diversified training data, one cannot build a reliable machine learning model [ 12 ] . In scenarios where the required data are not available enough, physics-informed neural networks (PINNs) can be used as a supplement. PINNs learn the underlying physics from the governing differential equations rather than the data from either the simulations or experiments, and predict the required quantities at par with the classical analytical and numerical methods [ 13 ] . The researchers successfully solved the Schrodinger equation [ 14 , 15 ] , Korteweg-de Vries equation [ 14 , 16 ] , Burger’s equation [ 14 , 17 ] , Poisson equation [ 18 , 19 ] , Navier-Stokes equations [ 20 , 21 ] , etc. The solution of the Helmholtz equation using PINNs can also be evidenced in the literature [ 22 , 23 ] . However, most of the proposed formulations focus on solving the governing equations with low-fidelity physics. Very limited research is available on the application of PINNs to real-world problems where it requires the incorporation of high-fidelity physics into the modeling. This paper demonstrates the neural network-based solution to predict the acoustic field inside a one-dimensional (1-D) uniform duct carrying a fluid whose properties change with respect to position. This work finds extensive applications in the aerospace industry, especially in the design of gas turbine combustors, where the rapidly changing medium properties introduce combustion instabilities into the system through the constant feedback loop between the acoustics of the combustor and the heat release rate [ 24 , 25 , 26 ] . Predicting acoustic field in such a heterogeneous medium is a challenging task even in the 1-D settings. Researchers are able to overcome the difficulties and provided solutions by analytical means [ 25 , 27 , 28 ] . However, developing a neural network-based solution (despite its success in recent times) is still a challenging problem for the following reasons. {arabiclist} Presence of mean flow make the coefficients the governing differential equation complex-valued. Hence, acoustic pressure becomes a complex-valued function, which ultimately makes the loss function complex-valued. However, the loss function must be a real-valued function to perform the optimization procedure. Developing a framework that can predict the acoustic field at different frequencies without altering the network architecture is not a trivial process. These problems are addressed in the current work, and the development of the formulation is presented step-by-step. The article is organized as follows: Section 2 describes the derivation of the governing equation from the fundamentals of fluid dynamics and thermodynamics. In Section 3 , a primer on PINNs and conversion of solving the derived differential equation into an optimization problem are presented. The prediction of acoustic pressure and particle velocity from the formulation developed is presented in Section 4 . The effect of the temperature gradient on the acoustic field is also studied in the same section. The article is concluded in Section 5 with final remarks and future scope. 2 Derivation of the governing equation Let us assume that a duct is carrying a fluid whose mean temperature and mean velocity change along the axial direction. Also, assume that the fluid is inviscid and obeys the perfect gas law. The governing equation that represents the propagation of sound inside the uniform duct with heterogeneous medium properties can be derived from the fundamental equation of fluid dynamics. According to Li et al. [ 25 ] , the continuity and momentum equations that govern fluid flow inside the one-dimensional duct can be written as \displaystyle\frac{\partial\rho}{\partial t}+\frac{\partial}{\partial x}(\rho u) \displaystyle=0, (1) \displaystyle\rho\frac{\partial u}{\partial t}+\rho u\frac{\partial u}{\partial x}+\frac{\partial p}{\partial x} \displaystyle=0, (2) where \rho , u , and p are the density, velocity, and pressure of the fluid medium, respectively. If \rho^{\prime} , u^{\prime} , and p^{\prime} represent their fluctuating components, then, \rho , u , and p can be expressed as \displaystyle\rho(x,t) \displaystyle=\overline{\rho}(x)+\rho^{\prime}(x,t), (3a) \displaystyle p(x,t) \displaystyle=\overline{p}(x)+p^{\prime}(x,t), (3b) \displaystyle u(x,t) \displaystyle=\overline{u}(x)+u^{\prime}(x,t). (3c) Here, \overline{\rho} , \overline{u} , and \overline{p} are the steady-state components of \rho , u , and p , respectively. Substituting the above expressions in Eqs. ( 1 ) and ( 2 ), and subtracting the mean components produce the linearized continuity and momentum equations as follows: \displaystyle\frac{\partial\rho^{\prime}}{\partial t}+\overline{u}\frac{\partial\rho^{\prime}}{\partial x}+u^{\prime}\frac{\partial\overline{\rho}}{\partial x}+\overline{\rho}\frac{\partial u^{\prime}}{\partial x}+\rho^{\prime}\frac{\partial\overline{u}}{\partial x} \displaystyle=0, (4) \displaystyle\overline{\rho}\frac{\partial u^{\prime}}{\partial t}+\rho^{\prime}\overline{u}\frac{\partial\overline{u}}{\partial x}+\overline{\rho}u^{\prime}\frac{\partial\overline{u}}{\partial x}+\overline{\rho}\,\overline{u}\frac{\partial u^{\prime}}{\partial x}+\frac{\partial p^{\prime}}{\partial x} \displaystyle=0. (5) The assumptions that the fluid medium is a perfect gas and the propagation of sound inside the duct is an isentropic process yields [ 2 ] \displaystyle\left(\frac{\partial p^{\prime}}{\partial\rho^{\prime}}\right)_{s} \displaystyle=\overline{c}^{2}, (6) \displaystyle\frac{\partial\overline{\rho}}{\partial x} \displaystyle=\frac{\overline{\rho}}{\gamma\overline{p}}\frac{\partial\overline{p}}{\partial x}, (7) where \overline{c}=\sqrt{\gamma R\overline{T}} is the speed of sound, \gamma is the specific heat ratio, R is the universal gas constant, and \overline{T} is the steady-state mean temperature. The subscript s denotes the is isentropic process. Refer to Appendix A for the derivation. Upon using Eqs. ( 6 ) and ( 7 ), and assuming that the time dependency is harmonic in nature, that is, \rho^{\prime}(x,t)=\widehat{\rho}(x)\,e^{-j\omega t} , p^{\prime}(x,t)=\widehat{p}(x)\,e^{-j\omega t} , and u^{\prime}(x,t)=\widehat{u}(x)\,e^{-j\omega t} , where \omega is the angular frequency, the continuity and momentum equation can be written as \displaystyle\left(j\omega+\gamma\frac{d\overline{u}}{dx}\right)\widehat{p}+\overline{u}\frac{d\widehat{p}}{dx}+\frac{d\overline{p}}{dx}\widehat{u}+\gamma\overline{p}\frac{d\widehat{u}}{dx} \displaystyle=0, (8) \displaystyle\left(j\omega+\frac{d\overline{u}}{dx}\right)\widehat{u}+\overline{u}\frac{d\widehat{u}}{dx}+\frac{\overline{u}}{\gamma\overline{p}}\frac{d\overline{u}}{dx}\widehat{p}+\frac{1}{\overline{\rho}}\frac{d\widehat{p}}{dx} \displaystyle=0. (9) The steady-state continuity and momentum equations gives the relations between \overline{\rho} , \overline{p} , and \overline{u} as follows \displaystyle\frac{d\overline{u}}{dx} \displaystyle=-\overline{u}\left(\frac{1}{\overline{\rho}}\frac{d\overline{\rho}}{dx}\right), (10) \displaystyle\frac{d\overline{p}}{dx} \displaystyle=-\overline{\rho}\,\overline{u}\frac{d\overline{u}}{dx}. (11) If \alpha=\frac{1}{\overline{\rho}}\frac{d\overline{\rho}}{dx}, (12) then Eqs. ( 10 ) and ( 11 ) can be written as \displaystyle\frac{d\overline{u}}{dx} \displaystyle=-\overline{u}\alpha, (13) \displaystyle\frac{d\overline{p}}{dx}

Read the complete text