4. Deep Data Assimilation for Source Detection and Localization

Characterizing the source of chemicals is a significant challenge with relevance in security, environmental monitoring, disaster mitigation, and search/rescue missions. The main objective of this project is to create a data-centric approach for localizing and characterizing chemical sources, free from the constraints imposed by current state-of-the-art methods’ simplifications. This innovative technique aims to capitalize on the affordability and low power consumption of emerging distributed wireless sensor systems to effectively identify and trace chemical sources. The outcomes of this research will have implications for security, pollution monitoring, mitigation, and firefighting. The project is divided into two primary goals:

1. Explore a data-driven approach to characterize sources based on sensor data, employing a unique physics-constrained deep learning framework.
2. Validate, confirm, and enhance the physics-constrained deep learning method through numerical simulations and real-world field experiments

Overall methodology

Usually, in these data driven models a large amount of data is required to achieve a significant accuracy. However, as the dispersion and flow of the transport medium adhere to physical laws (advection-diffusion for chemical dispersion and Navier-Stokes for transport medium flow), the solutions exist within a low-dimensional manifold. In other words, transport medium velocities, pressures, and chemical species concentrations are interconnected over time and space through governing differential equations. This suggests that solving the problem of chemical source characterization in a purely data-driven fashion necessitates a relatively small amount of data. Moreover, introducing some of the physics equations mentioned above as constraints in the form of loss functions will enable even more accurate predictions with less data (as shown in figure above).

A preliminary deep learning model has already been developed using numerical simulations for data generation where it was simulated in steady state conditions (no change in properties with time).

Simulation domain for chemical source characterization numerical experiments. All distance units are in meters

The results of this preliminary work show great potential of this method where the deep neural network was able to accurately predict the concentration given the basic and preliminary conditions (as shown in figure below)

Comparison of concentration and source magnitude map predicted by the physics-constrained deep learning algorithm with the reference.