Modelling environmental data, including temperature and air pollution, has been tackled by both physical methods and deep learning models. The former approach builds upon physical laws albeit using computationally complex algorithms. Alternatively, deep learning methods are purely data-driven, exploiting diverse spatio-temporal data correlations. While the latter delivers good performance and fast inference, it does not leverage knowledge about the physical phenomena. Recently, deep equilibrium networks (DEQs), i.e., networks that find the fixed point of some iterative procedure, have shown outstanding performance when compared with their equivalent deep models. In this work, we propose a physics-guided DEQ model for environmental data that can be modeled as a graph. Our approach builds on a graph convolutional operator, incorporating the partial differential equation that defines the convection-diffusion physical process. We evaluate the effectiveness of our approach in the task of air quality estimation based on sensor measurements. Experiments on real-world air quality data show the improved performance of our model with respect to state-of-the-art approaches.
Rodrigo Bonet, E & Deligiannis, N 2024, Physics-guided Graph Convolutional Deep Equilibrium Network for Environmental Data. in 32nd European Signal Processing Conference, EUSIPCO 2024 - Proceedings. European Signal Processing Conference, IEEE, pp. 987-991. https://doi.org/10.23919/EUSIPCO63174.2024.10715398
Rodrigo Bonet, E., & Deligiannis, N. (2024). Physics-guided Graph Convolutional Deep Equilibrium Network for Environmental Data. In 32nd European Signal Processing Conference, EUSIPCO 2024 - Proceedings (pp. 987-991). (European Signal Processing Conference). IEEE. https://doi.org/10.23919/EUSIPCO63174.2024.10715398
@inproceedings{7aaf14919a4b4c1ca0059e4903c10813,
title = "Physics-guided Graph Convolutional Deep Equilibrium Network for Environmental Data",
abstract = "Modelling environmental data, including temperature and air pollution, has been tackled by both physical methods and deep learning models. The former approach builds upon physical laws albeit using computationally complex algorithms. Alternatively, deep learning methods are purely data-driven, exploiting diverse spatio-temporal data correlations. While the latter delivers good performance and fast inference, it does not leverage knowledge about the physical phenomena. Recently, deep equilibrium networks (DEQs), i.e., networks that find the fixed point of some iterative procedure, have shown outstanding performance when compared with their equivalent deep models. In this work, we propose a physics-guided DEQ model for environmental data that can be modeled as a graph. Our approach builds on a graph convolutional operator, incorporating the partial differential equation that defines the convection-diffusion physical process. We evaluate the effectiveness of our approach in the task of air quality estimation based on sensor measurements. Experiments on real-world air quality data show the improved performance of our model with respect to state-of-the-art approaches.",
author = "{Rodrigo Bonet}, Esther and Nikos Deligiannis",
note = "Funding Information: This work was supported in part by the Research Foundation - Flanders (FWO) through the Ph.D. Fellowship Strategic Basic Research under Project 1SC4521N, in part by IMEC under the AAA Project AI-based Air Quality Map and Analytics and in part by the Flemish Government, under the \u201COnderzoeksprogramma Artifici\u00EBle Intelligentie (AI) Vlaanderen\u201D programme. Publisher Copyright: {\textcopyright} 2024 European Signal Processing Conference, EUSIPCO. All rights reserved.",
year = "2024",
doi = "10.23919/EUSIPCO63174.2024.10715398",
language = "English",
isbn = "978-9-4645-9361-7",
series = "European Signal Processing Conference",
publisher = "IEEE",
pages = "987--991",
booktitle = "32nd European Signal Processing Conference, EUSIPCO 2024 - Proceedings",
}