Neural Network Modeling of Water Quality in the Doce River Basin in Mariana (MG) with Computational Projections through 2030

Objective: This work aimed to model the dynamics of water quality parameters in the Bacia do Rio Doce Mariana (MG) after the dam rupture, integrating neural networks and differential equations from data (1997-2015) made available by the Minas Gerais Water Management Institute (IGAM), enabling computational projections up to 2030 that help in inferring about the recovery of water quality. Theoretical Framework: A statistical analysis of water quality parameter data was conducted, covering the periods before and after the 2015 dam failure. The database was made publicly available by the Minas Gerais Institute for Water Management (IGAM), containing 58 water quality parameters from the Rio Doce basin (Mariana, MG) for the period from 1997 to 2023. The most impactful variables in the dataset were identified as suspended solids concentration, water turbidity, and heavy metals. System modeling was performed using differential equations and feedforward Artificial Neural Network (ANN) models, utilizing the Levenberg-Marquardt training algorithm implemented in the MATLAB platform. The network was configured with two hidden layers, 10 neurons, and a tansig activation function for computational projections of water quality parameters through 2030. Results and Discussion: neural network training resulted in a correlation coefficient of 0.85, evidence of the model's accuracy in relation to real-world data. The simulations performed demonstrated that turbidity values (NTU) are decreasing over the years. The reduction of turbidity in the river basin acts as a fundamental catalyst for water quality recovery, as it decreases the load of pollutants and organic matter frequently associated with suspended particles. By mitigating these solids, an immediate reduction in COD (Chemical Oxygen Demand) and BOD (Biochemical Oxygen Demand) occurs, since there is less substrate for chemical oxidation and less organic material to be decomposed by microorganisms. Consequently, with the lower biochemical oxygen demand and greater sunlight penetration, which stimulates the photosynthesis of aquatic plants, Dissolved Oxygen (DO) levels tend to rise, creating a favorable environment for the restoration of biodiversity and the resilience of the fluvial ecosystem. This results in a forecast of a return to baseline parameters for turbidity, suspended solids, COD, BOD, and DO starting from 2030.Regarding the concentration of heavy metals, computational modeling predicts a more extensive recovery horizon, given that the tailings deposited on the bed of the Rio Doce basin are subject to continuous processes of leaching and sediment resuspension, which may perpetuate the contamination of the water column for decades. Research Implications: The application of dynamic modeling and ANNs proved effective in modeling and analyzing multivariable chemical processes in transient regimes, enabling computational projections up to 2030 that assist in inferring about the recovery of water quality. Originality/Value: The main contribution of this work lies in demonstrating the ability of artificial neural networks (ANNs) to perform computational projections, aiding in inferences about the recovery of water quality in the Bacia do Rio Doce Mariana (MG).