Knowledge-Guided Deep Learning Framework for Source Apportionment and Mitigation of Perchlorate Contamination in River Basins

Fireworks production leads to severe perchlorate contamination in aquatic systems, yet traditional approaches struggle to resolve the spatially coupled industrial pollution dynamics and governance fragmentation in river basins, fundamentally limiting effective perchlorate mitigation. Herein, we investigate the spatiotemporal patterns of perchlorate contamination in China's largest fireworks production regions, observing a significant ″accumulation corridor″ aligning with the industrial clustering zones. We developed a knowledge-guided framework, integrating process-based hydrology with deep learning, to predict perchlorate concentration dynamics at the basin scale. The model outperforms traditional data-driven approaches in capturing long-term trends and episodic peaks while leveraging Mixture-of-Experts (MoE) gates to distinguish point versus nonpoint sources. We demonstrate that industrial point sources contribute ∼65% of total corridor loads, whereas nonpoint loads are seasonally amplified, with rainy-season fluxes 1.7-2.8× higher than dry seasons. Attention analysis identifies that average fireworks factory density, precipitation, and soil properties are dominant drivers for nonpoint perchlorate sources. Furthermore, multiscenario simulations indicate that coordinated controls on point, nonpoint, and upstream boundary inputs substantially reduce exceedance frequency within the simulated ensembles. This work provides a transferable framework that converts fragmented observations into basin-scale diagnosis and scenario-based decision support for industrial contaminant risks, contingent on local monitoring, industrial structure information, and regulatory setting.