PHYSICS-INFORMED NEURAL NETWORKS FOR MODELING OCEAN POLLUTANT

Karishma Battina, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, Sreedath Panat · 2025

Read PHYSICS-INFORMED NEURAL NETWORKS FOR MODELING OCEAN POLLUTANT by Karishma Battina, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, Sreedath Panat (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-2507-08834.txt.

Introduction

Karishma Battina and co-authors model ocean pollutant spread with a neural net that has the physics baked in, in a paper from two thousand twenty-five. Classic solvers strain at the scale and tangle of ocean flow. Their net learns under the drift-spread law plus edge and start rules.

Training data came from a grid solver with added noise at mixed levels. A blended loss weighs law fit, edge fit, and data fit together. The setup promises scale and bend that grid tools lack. Code ran on a fast science stack in Julia. Readers get a physics-true learner for seas where spills move fast and data run thin.

Catalog record

CREATORKarishma Battina, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, Sreedath Panat
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-08834

Browse the full library · About the Tower of Babel

Full text

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

PHYSICS-INFORMED NEURAL NETWORKS FOR MODELING OCEAN POLLUTANT Karishma Battina, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, Sreedath Panat Abstract Traditional numerical methods often struggle with the complexity and scale of modeling pollutant transport across vast and dynamic oceanic domains. This paper introduces a Physics-Informed Neural Network (PINN) framework to simulate the dispersion of pollutants governed by the 2D advection-diffusion equation. The model achieves physically consistent predictions by embedding physical laws and fitting to noisy synthetic data, generated via a finite difference method (FDM), directly into the neural network training process. This approach addresses challenges such as non-linear dynamics and the enforcement of boundary and initial conditions. Synthetic data sets, augmented with varying noise levels, are used to capture real-world variability. The training incorporates a hybrid loss function including PDE residuals, boundary/initial condition conformity, and a weighted data fit term. The approach takes advantage of Julia’s scientific computing ecosystem for high-performance simulations, offering a scalable and flexible alternative to traditional solvers. 1 Introduction The dispersion of pollutants in ocean currents, from plastic debris to oil spills, poses significant threats to marine ecosystems and coastal communities. Predicting pollutant transport requires solving the advection-diffusion equation, a partial differential equation(PDE) that models the interplay of fluid-driven advection and turbulent diffusion [ 19 ] . However, traditional numerical methods, such as finite-volume or spectral approaches, face critical limitations: computational costs escalate exponentially with domain size, parameter uncertainties degrade accuracy, and resolving boundary effects near coastlines remains challenging [ 14 , 20 , 21 ] . These constraints hinder timely predictions for environmental mitigation efforts, particularly in dynamic, large-scale oceanic systems. Recent advances in scientific machine learning have introduced physics-informed neural networks (PINNs) as a paradigm for solving PDEs [ 12 ] . Unlike conventional solvers, PINNs embed physical laws directly into neural network training, enabling mesh-free simulations and seamless integration of sparse observational or synthetic data. While PINNs have shown promise in fluid dynamics, their application to ocean pollution remains underexplored, especially in addressing real-world complexities such as transient velocity fields, spatially varying diffusion, and irregular coastal geometries. In this work, we develop a PINN-based framework tailored for oceanic pollutant transport. Our key contributions are threefold: 1.1 Review of Pollutant Transport Modeling The modeling of oceanic pollutant transport has traditionally been approached from two main perspectives. Eulerian models, such as the Finite Difference Method (FDM) or Finite Element Method (FEM), discretize the ocean into a fixed grid and solve the governing equations at each grid point [ 15 ] . These methods are robust but become computationally prohibitive for large-scale, long-term simulations. Alternatively, Lagrangian particle-tracking models simulate the trajectories of individual pollutant particles, which is effective for tracking sparse debris but can struggle to represent continuous concentration fields accurately [ 10 ] . Recent years have seen the application of machine learning to augment or replace these methods. Machine learning techniques augment traditional models but face generalization challenges without physical consistency [ 12 , 23 ] . Some studies use data-driven neural networks to predict flow fields or correct biases in numerical models [ 13 ] . However, these purely data-driven approaches require vast amounts of training data and may not generalize well to scenarios outside their training distribution, as they lack inherent knowledge of the underlying physics [ 24 ] . 1.2 Physics-Informed Neural Networks (PINNs) To address these limitations, recent advances in scientific machine learning have introduced Physics-Informed Neural Networks (PINNs) as a paradigm for solving PDEs [ 1 , 8 , 9 ] . Unlike conventional solvers or purely data-driven models, PINNs embed the physical laws directly into the neural network’s loss function [ 22 ] . This enables mesh-free simulations and seamless integration of sparse observational or synthetic data [ 13 , 18 ] . While PINNs have shown promise in fluid dynamics [ 18 ] , their application to ocean pollution remains underexplored, especially concerning the systematic analysis of how network design choices impact solution accuracy for complex, noisy systems [ 3 , 5 , 6 ] . In this work, we develop a PINN-based framework tailored for oceanic pollutant transport. Our key contributions are: • The development of a robust, data-augmented PINN framework in Julia that accurately enforces sharp initial conditions through a hybrid, weighted loss function. • A systematic investigation into the role of neural network architecture (depth, width) and hyperparameters (learning rate, optimizers) on solution accuracy and computational cost. • A comprehensive performance benchmark that identifies an optimal configuration capable of achieving a relative L_{2} error of approximately 8.25% against a high-resolution FDM. 2 Methodology 2.1 Mathematical Formulation of the Advection-Diffusion Equation The advection-diffusion equation governs the transport of pollutants in oceanic environments, a partial differential equation that captures the combined effects of advection (directional transport by currents) and diffusion (random spreading of particles) [ 11 , 19 ] : \frac{\partial u}{\partial t}+\mathbf{v}\cdot\nabla u=D\nabla^{2}u (1) Where: u(x,y,t) Represents the concentration of pollutants at spatial location (x,y) and time t . \mathbf{v}=(v_{x},v_{y}) is the velocity field representing ocean currents, with components v_{x} and v_{y} . In this study, we set: v_{x}=0.5,\quad v_{y}=0.5. D is the diffusion coefficient, quantifying the rate of pollutant dispersion [ 2 , 16 ] . For this work, we use D=0.01. The advection term \mathbf{v}\cdot\nabla u\quad (2) models the transport of pollutants by ocean currents, while the diffusion term \quad D\nabla^{2}u (3) accounts for the spread of pollutants due to random molecular motion and turbulent eddies. When expanded in two dimensions, this equation becomes the following: \frac{\partial u}{\partial t}+v_{x}\frac{\partial u}{\partial x}+v_{y}\frac{\partial u}{\partial y}=D\left(\frac{\partial^{2}u}{\partial x^{2}}+\frac{\partial^{2}u}{\partial y^{2}}\right) (4) This formulation serves as the foundation for our PINN-based modeling approach, enabling the simulation of pollutant transport across oceanic domains with varying current patterns and diffusion characteristics. 2.2 Physics-Informed Neural Network Framework Our approach leverages a PINN to approximate the solution u(t,x,y) of the advection-diffusion equation. The framework is designed to be expressive, physically consistent, and robust to noisy data through a hybrid training strategy. 2.2.1 Boundary and Initial Conditions To ensure a well-posed problem, we define the spatio-temporal domain as t\in[0,\,0.25],\quad x\in[0,\,1.0],\quad y\in[0,\,1.0]. The final time is set to T_{\mathrm{final}}=0.25 to ensure the reference solution remains non-trivial within the domain. The initial condition represents a sharply localized pollutant release at the center of the domain. This Gaussian peak is a standard analytical choice for modeling an idealized, instantaneous point-source spill before significant dispersion occurs [ 7 , 9 ] . Dirichlet boundary conditions are imposed, enforcing approximately zero pollutant concentration at the domain edges. This setup represents a localized pollution event within a much larger body of water, where the concentration at the far-field boundaries is assumed to be negligible. The inclusion of stochastic noise in these conditions aims to simulate the minor fluctuations and measurement uncertainties present in real-world oceanic systems. The following boundary and initial conditions, incorporating stochastic noise to simulate realistic data, are imposed: Boundary Conditions (Dirichlet): u(t,\,x,\,0)\approx 0.0+\epsilon_{\mathrm{BC}} (5) u(t,\,x,\,1)\approx 0.0+\epsilon_{\mathrm{BC}} (6) u(t,\,0,\,y)\approx 0.0+\epsilon_{\mathrm{BC}} (7) u(t,\,1,\,y)\approx 0.0+\epsilon_{\mathrm{BC}} (8) These conditions enforce approximately zero pollutant concentration at the domain boundaries [ 10 ] . \epsilon_{\mathrm{BC}} represents additive Gaussian noise with a standard deviation of \sigma_{\mathrm{BC}}=0.01 . Initial condition (IC): u(0,x,y)\approx\exp\left(-100\left[(x-0.5)^{2}+(y-0.5)^{2}\right]\right)\cdot\left(1+\epsilon_{\mathrm{IC}}\right) This initial condition represents a sharply localized pollutant release at the center of

Read the complete text