Bridging Physics and Machine Learning: Exploring SEIR Epidemiological Modeling using Physics-Informed Neural Networks

Soni, Abhishek

Permalink

https://hdl.handle.net/2142/128600 Copy

Description

Title

Bridging Physics and Machine Learning: Exploring SEIR Epidemiological Modeling using Physics-Informed Neural Networks

Author(s)

Soni, Abhishek

Issue Date

2025-05-06

Keyword(s)

Physics

Date of Ingest

2025-06-02T14:03:44-05:00

Abstract

Scientific Machine Learning (SciML) represents a transformative intersection of artificial intelli- gence, physics, and computational science, offering new approaches to solving complex physical sys- tems. At its forefront are Physics-Informed Neural Networks (PINNs), which integrate physical laws directly into machine learning frameworks to enhance accuracy, efficiency, and interpretability. This paper explores the principles of SciML, contrasting it with traditional machine learning and scientific computing. It then delves into the workings of PINNs, their advantages in addressing data scarcity, and their ability to incorporate domain knowledge into predictive models. This study also explores the application of PINNs as an innovative approach to enhance the flexibility and accuracy of SEIR Epi- demiological modeling. Epidemiological modeling is crucial for understanding and predicting disease transmission dynamics. Traditional SEIR (Susceptible-Exposed-Infected-Recovered) models rely on ordinary differential equations with fixed parameters, limiting their adaptability to complex real-world scenarios. We implemented a PINN architecture using PyTorch to model COVID-19 spread dynamics based on a dataset from South Korea. Four separate neural networks were constructed to represent each SEIR compartment, with physics-informed loss functions enforcing adherence to epidemiological constraints. The model was trained on 489 days of confirmed and recovered case data using Adam op- timizer with L2 regularization and adaptive learning rate. Model performance was evaluated through comparison with traditional ODE solvers using identical parameters (β=0.132, γ=0.06, σ=1/5.2). The PINN-based model successfully captured SEIR dynamics while maintaining physical consistency with epidemiological principles. Loss convergence analysis demonstrated stable learning behavior, with final loss values reaching 10−5 scale. Comparative analysis with classical ODE solutions showed comparable prediction accuracy, while the PINN approach demonstrated superior robustness when handling noisy data points.

Type of Resource

text

Language

eng

Owning Collections