ELPINN: Eulerian Lagrangian Physics-Informed Neural Network

Sukirt Thakur, Maziar Raissi · 2025

Read ELPINN: Eulerian Lagrangian Physics-Informed Neural Network by Sukirt Thakur, Maziar Raissi (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-2504-09053.txt.

Introduction

Sukirt Thakur and Maziar Raissi rebuild flow fields from drifting particles by wedding two viewpoints inside a physics-informed network, in a 2025 paper. Eulerian eyes watch fixed points; Lagrangian eyes ride the particles. Their network takes particle tracks and returns the full velocity field, and, for flat flows, the pressure too, with no pressure gauges anywhere. Three testbeds span flow past a cylinder, flow in a box, and air inside an airplane cabin.

Velocity came back faithfully in all three. Finer time steps mattered far more than sheer particle count. The authors pitch the method as a booster for particle-tracking experiments, with three-dimensional pressure and hidden traits like viscosity as next targets.

Catalog record

CREATORSukirt Thakur, Maziar Raissi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-09053

Browse the full library · About the Tower of Babel

Full text

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

ELPINN: Eulerian Lagrangian Physics-Informed Neural Network Sukirt Thakur, Maziar Raissi Abstract Physics-Informed Neural Networks (PINNs) have gained widespread popularity for solving inverse and forward problems across a range of scientific and engineering domains. However, most existing PINN frameworks are limited to the Eulerian domain, where physical quantities are described at fixed spatial locations. In this work, we propose a novel PINN-based framework that couples Eulerian and Lagrangian perspectives by using particle trajectory data to reconstruct Eulerian velocity and pressure fields. We evaluate the performance of our method across three distinct fluid flow scenarios: two-dimensional external flow past a cylinder, two-dimensional internal flow in a confined geometry, and three-dimensional internal flow inside an airplane cabin. In all three cases, we successfully reconstruct the velocity field from Lagrangian particle data. Moreover, for the 2D external and internal flows, we recover the pressure field solely through the physics-informed learning process, without using any direct pressure measurements. We also conduct a sensitivity analysis to understand the effects of temporal resolution and particle count on the reconstruction accuracy. Our results show that smaller time-step sizes significantly improve the predictions, while the total number of particles has a comparatively smaller influence. These findings establish the potential of our coupled Eulerian-Lagrangian PINN framework as a powerful tool for enhancing experimental methods such as Particle Tracking Velocimetry (PTV). Looking ahead, this approach may be extended to infer hidden quantities such as pressure in three-dimensional flows or material properties like viscosity, opening new avenues for data-driven fluid dynamics in complex geometries. 1 Introduction The application of machine learning (ML) and deep learning (DL) techniques to scientific problems is rapidly expanding. These approaches have been successfully applied across diverse fields, including fluid mechanics [ 1 , 2 ] and epidemiology [ 3 ] . However, many physical and biological systems suffer from sparse, noisy, and difficult-to-acquire data, making traditional ML frameworks prone to overfitting and generalization errors [ 4 ] . Physics-Informed Neural Networks (PINNs) [ 5 ] have emerged as a robust computational framework for solving forward and inverse problems across multiple domains. PINNs effectively handle small datasets by incorporating governing physical equations as a regularization term, reducing the dependency on large labeled datasets. This approach has been successfully applied to fluid dynamics [ 6 , 7 , 8 ] , rheology [ 9 , 10 ] structural mechanics [ 11 , 12 ] , and heat transfer [ 13 , 14 ] . Moreover, PINNs have been extended to tackle high-dimensional PDEs [ 15 ] and multi-physics problems [ 16 ] . Most PINN applications focus on governing equations formulated in the Eulerian domain. In this framework, physical quantities such as velocity, pressure, and temperature are described at fixed spatial locations. This perspective is well-suited for continuum mechanics, where fluid flow and heat transfer are analyzed across a spatial domain. The Eulerian formulation, commonly used to describe systems governed by partial differential equations (PDEs), is particularly effective for fluid dynamics applications, such as solving the Navier-Stokes equations [ 4 , 17 ] . Conversely, the Lagrangian perspective focuses on individual particles or fluid elements, tracking their trajectories over time [ 18 , 19 ] . This approach is fundamental for studying particle-laden flows [ 20 ] , turbulence modeling [ 21 ] , molecular dynamics [ 22 ] , and transport phenomena such as sediment transport [ 23 ] . While PINNs have been widely applied to Eulerian problems, their extension to Lagrangian descriptions remains underexplored. However, many physical systems exhibit hybrid behavior, requiring a coupled Eulerian-Lagrangian approach. For instance, sediment transport in a river involves suspended particles (Lagrangian) influenced by fluid flow (Eulerian) and vice versa [ 24 ] . Similarly, in biological systems, cells within tissues experience fluid forces (Eulerian) while simultaneously migrating and interacting with neighboring cells (Lagrangian). Accurately modeling these interactions necessitates a unified framework that bridges both viewpoints. In this work, we present a PINN framework that leverages Lagrangian particle data to make predictions in the Eulerian domain. Specifically, we track passive particles in the Lagrangian domain to reconstruct the velocity field in the Eulerian domain. Using this velocity field, we further infer the pressure distribution by minimizing the residuals of the Navier–Stokes equations. The governing equations and the PINN architecture for Eulerian–Lagrangian coupling are described in Section 2 . We evaluate our framework on three test cases of increasing complexity: a two-dimensional external flow past a cylinder, a two-dimensional internal flow in a confined domain, and a three-dimensional internal flow representing airflow within an airplane cabin. In all three cases, we successfully reconstruct the velocity field from particle trajectory data. For the 2D external and internal flow problems, we additionally recover the pressure field using physics-informed learning. The corresponding results and a detailed sensitivity analysis are presented in sections 3.1 and 3.2 for the 2D cases and in Section 3.3 for the 3D airplane geometry. We conclude with a summary and discussion of the broader implications of our findings in Section 4 . 2 Problem definition 2.1 Governing equations Our objective is twofold: first, to track inert particles and obtain the velocity field using the trajectory of the particles; and second, to use this velocity field in a Lagrangian frame of reference to calculate the pressure field by minimizing the residual of the Navier-Stokes equations. We consider the passive transport of mass-less particles, whose dynamics are governed by the following equation: \boldsymbol{x}_{t}=\boldsymbol{u}(t,\boldsymbol{x}), (1) where the subscript t denotes the time derivative and \boldsymbol{x} is the spatial coordinate that is \boldsymbol{x}=(x,y) in two dimensions and \boldsymbol{x}=(x,y,z) in three dimensions. The velocity field is defined as \boldsymbol{u}=(u,v) in two dimensions and as \boldsymbol{u}=(u,v,w) in three dimensions. Equation ( 1 ) can be written in terms of the individual components as: \frac{dx}{dt}=u,\frac{dy}{dt}=v,\frac{dz}{dt}=w. (2) The conservation of mass for an incompressible fluid is given by: \nabla\cdot\boldsymbol{u}=0, (3) where \boldsymbol{u} is the fluid velocity vector. In terms of individual velocity components, the conservation of mass in two-dimensions can be written as: u_{x}+v_{y}=0. (4) The conservation of momentum for an incompressible Newtonian fluid under isothermal, single-phase, transient conditions in the absence of a body force is given by: \left(\frac{\partial\boldsymbol{u}}{\partial t}+\boldsymbol{u}\cdot\nabla\boldsymbol{u}\right)=-\frac{1}{\rho}\nabla p+\nu\nabla^{2}\boldsymbol{u}, (5) where \rho is the density of the fluid, \boldsymbol{u} is the velocity vector, t is time, p is the pressure, and \nu is the kinematic viscosity. The vector form of the momentum equation in two dimensions in the x and y directions is given by: \begin{split}&u_{t}+uu_{x}+vu_{y}=-\frac{1}{\rho}p_{x}+\nu(u_{xx}+u_{yy}),\\ &v_{t}+uv_{x}+vv_{y}=-\frac{1}{\rho}p_{y}+\nu(v_{xx}+v_{yy}).\end{split} (6) We define the left-hand side of the momentum equations as \boldsymbol{f^{LHS}}=(f^{LHS}_{1},f^{LHS}_{2}) where: \begin{split}&f^{LHS}_{1}:=u_{t}+uu_{x}+vu_{y},\\ &f^{LHS}_{2}:=v_{t}+uv_{x}+vv_{y}.\end{split} (7) The right-hand side of the momentum equations in 2D can be defined as \boldsymbol{f^{RHS}}=(f^{RHS}_{1},f^{RHS}_{2}) where: \begin{split}&f^{RHS}_{1}:=-\frac{1}{\rho}p_{x}+\nu(u_{xx}+u_{yy}),\\ &f^{RHS}_{2}:=-\frac{1}{\rho}p_{y}+\nu(v_{xx}+v_{yy}).\end{split} (8) For the 2D external and internal flows, the velocity field is reconstructed by solving the momentum equations and minimizing the residuals. Additionally, the pressure field is learned by minimizing the residual of the Navier-Stokes equations for these cases. However, in the 3D internal flow case, we reconstruct only the velocity field and do not solve for the pressure field. Figure 1: Neural network setup to learn the Eulerian velocity field by tracking particles in the Lagrangian domain. The arrows represent the flow of information. The peach network has the particle positions x and y as

Read the complete text