(372aa) Fixed/Mobile Detector Placement Optimization and Validation System Using Differentiable Physics-Based Surrogate Models for Chemical Leak Dispersions

Accidental chemical leaks can lead to large-scale industrial disasters such as fires, explosions, or toxic dispersions. The detection and appropriate response to chemical leaks are crucial in preventing, mitigating and minimizing loss by such incidents. Leak detection by pre-installation of fixed detectors for major gases handled in the process is a common technology, but installation of these sensors in the process is mainly accomplished by arranging them at regular intervals according to legal codes. However, whether these deployment results guarantee detection and risk control performance such as appropriate detection time and risk minimization requires a detailed evaluation compared to massive dispersion simulations according to major leakage scenarios.

This study introduces an open-source, leak detector placement optimization and validation system that is being developed with support from the Ministry of Environment, especially to support small and medium-sized enterprises. Various objective functions can be selected out of the toolset, such as minimizing risk, detection time, or the cost of sensor placement, allowing for the selection of the most appropriate objective function and constraints based on the characteristics of the process and risk policy of the enterprise. The proposed system enables the achievement of more accurate and effective sensor placement and the following validation and verification, tailored to potential risks of the process and capable of robust operation in possible leak scenarios. The prototype system has been validated against a couple of field tests.

The proposed prototype system for optimal placement of chemical leak detectors combines a differentiable physics-based surrogate model and a sensor placement optimization tool. The proposed system solves MINLP optimization problems by using massive dispersion simulation results or by directly integrating the surrogate models into formulations. The traditional process of generating massive data out of Computational Fluid Dynamics (CFD) simulations is time-consuming, and it becomes a bottleneck during the placement optimization process. The surrogate model, developed based on differentiable physics, offers high-throughput capabilities by providing faster results compared to traditional CFD simulations. The surrogate model is also combined with a CNN-based deep learning model to provide high-resolution outputs from low-resolution inputs. To enhance the accuracy of the surrogate model, a sequential correction process is conducted, where the surrogate model is adjusted to be similar to the CFD model results obtained from commercial CFD software such as COMSOL Multiphysics and FLACS while reducing computation time.