Optimization of sensor locations for source tracking of chemical leaks using numerical simulations and physics-informed neural networks

Abstract

Inverse modeling or inverse problem is applied in leak source detection in dispersion cases. It is the process of calculating the circumstances that produce a set of observed or measured outcomes. However, inverse modeling methods are computationally expensive and numerically unstable. Recent literature has suggested that machine learning methods can be used to solve this inverse problem. Physics informed neural network was implemented to predict gas leak location. Our results showed that this physics + data-driven approach leads to better prediction performance while requiring lesser data. In a real-world scenario, the reduced amount of data required would also correspond to a need for fewer sensors and, consequently, cost savings. Based on this methodology, the number of sensors and sensor locations was also optimized to save costs and achieve the optimal performance of the model. This project showed the potential of physics informed models in solving scientific inverse problems, especially in the context of leak source detection. Additional preliminary evaluations of different types of sensor placements also showed the potential for further optimization to realize further cost savings without compromising predictive performance.