Development of a Chemical Accident Inverse Tracking Framework for River Systems Using Generative Artificial Intelligence

Accelerating urbanization has made toxic chemical sources in river systems increasingly complex, making their identification and control progressively more challenging. Toxic chemical source tracking is essential for rapid emergency response and effective water quality management. Existing source tracking approaches, such as statistical methods, numerical models, and deep learning, face critical limitations. Statistical methods have limitations in capturing the non-linear transport dynamics and causality of toxic chemicals in river systems. Numerical models require high computational cost and time to achieve high accuracy, while deep learning models suffer from critical data scarcity, as actual toxic chemical accident datasets are limited. This study aims to develop a hybrid framework that combines the high accuracy of numerical models with the computational efficiency of deep learning-based generative artificial intelligence, specifically a Generative Adversarial Network (GAN), enabling near real-time inverse tracking of chemical accidents. To generate training data for the GAN model, we established an automated scenario generation algorithm coupled with the Environmental Fluid Dynamics Code (EFDC), a three-dimensional hydrodynamic and water quality model. For the Geum River basin in South Korea, we conducted EFDC simulations under scenarios varying in source locations, release amounts, and spill timing for phenol, generating a high-quality synthetic dataset. The synthetic dataset is used to train a GAN for inverse problem solving. During training, the Generator learns to map upstream source information to downstream toxic concentration time series, while the Discriminator evaluates whether the generated source-concentration pairs are consistent with EFDC transport mechanisms. In this process, the Generator aims to produce realistic downstream concentration time series to deceive the Discriminator, whereas the Discriminator aims to distinguish these generated outputs from the synthetic training data. Through this adversarial mechanism, the Generator progressively produces more refined downstream concentration time series. In the event of a real chemical accident, the trained GAN model enables rapid inference of the corresponding source information from observed downstream concentrations through inverse problem solving, without the need for iterative numerical simulations. This approach is expected to overcome the limitations of high computational cost in numerical models and data scarcity in deep learning. This rapid inverse tracking framework provides sufficient time to effectively respond to chemical accidents and helps protect critical downstream infrastructure such as drinking water treatment plants from toxic chemicals.