Threat Zones Identification and Consequence Modeling Approach of Ammonia Dispersion via Machine Learning

Abstract Release of toxic chemicals from process industries is a threat to our environment, and simulating such scenarios provides first-hand information to make decisions regarding measures to tackle such situations. The available ALOHA software for such simulations is time-consuming. Therefore, this study presents a new framework for ammonia dispersion using a multioutput artificial neural network (ANN) surrogate model trained on 4989 scenarios simulated via the readily available simulation software to save crucial time, enabling quick decisions. The simulation is based on a real-world industrial site in Sheikhupura, Pakistan, and results in the generation of a comprehensive data set. The data set has taken into consideration various input variables and studied their effect on target output variables. Next, the ANN surrogate model is optimized via hyperparameter tuning, and its performance is evaluated using standard regression metrics, demonstrating excellent agreement with ALOHA simulation outputs, with R2 values exceeding 0.98 and mean absolute percentage errors below 2.6% for all output parameters. These performance metrics represent the agreement between the ANN surrogate model and ALOHA simulation outputs. Comparing the speed of the trained ANN surrogate over the ALOHA simulation software, a superiority of 27,000 times is observed, that is, predictions are made in milliseconds. This significant reduction in inference time demonstrates the capability of the proposed ANN surrogate model to provide a rapid approximation of ALOHA simulation outputs for rapid scenario evaluation and preliminary consequence assessment.