Design optimization of dual-circulation wells using deep learning

• Published in: Stochastic environmental research and risk assessment Vol. 39; no. 5; pp. 1899 - 1914
• Main Authors: Sikai, Yang, Zhang, Fang, Yanling, Ma, Hao, Ke, Rui, Zhou, Sadam, Muhammad
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.05.2025; Springer Nature B.V
• Subjects: Aquatic Pollution; Aquifers; Artificial neural networks; Back propagation networks; Chemistry and Earth Sciences; Circulation; Computational Intelligence; Computer Science; Computer simulation; Deep learning; Design; Design optimization; Earth and Environmental Science; Earth Sciences; Environment; Genetic algorithms; Groundwater; Groundwater flow; Groundwater treatment; Hydraulic conductivity; Machine learning; Math. Appl. in Environmental Science; Mathematical analysis; Mathematical models; Multiple objective analysis; Neural networks; Optimization models; Optimization techniques; Original Paper; Particle tracking; Permissible error; Physics; Pollutant removal; Probability Theory and Stochastic Processes; Recurrent neural networks; Remediation; Simulation; Simulation models; Solute transport; Sorting algorithms; Statistics for Engineering; Structural design; Waste Water Technology; Water Management; Water Pollution Control; Wells; Deep learning; NSGA-II; Groundwater; Multiobjective optimal design; Convolutional neural network; Dual-circulation well
• ISSN: 1436-3240; 1436-3259
• DOI: 10.1007/s00477-025-02947-9

Optimizing groundwater circulation wells (GCWs) is crucial for their effective use in groundwater remediation at polluted sites. Therefore, this study developed a deep learning-based simulation–optimization coupling technology for the design of dual-circulation wells. Groundwater simulation modules based on the Python package were used to construct models for groundwater flow, particle tracking, and solute transport. Operational influencing factors, including hydraulic conductivity, aquifer thickness, extraction and injection tubing lengths, and well locations, were the input variables. Indicators, such as the circulation efficiency ( P r ) and pollutant removal rate ( η ), were the output variables in the developed dataset. The convolutional neural network (CNN) model, the recurrent neural network (RNN) model, and the backpropagation neural network (BP) model were applied as surrogates to the numerical simulation model for computation. Integration occurs with the multiobjective optimization model, which employs the nondominated sorting genetic algorithm II. The influencing parameters and optimal design of the GCWs were subsequently obtained at a test site located in Hefei, Anhui Province, via the coupled model for assessment. The findings indicate the model’s ability to achieve an optimal arrangement of the GCWs. Among the three models, the surrogate CNN model exhibited superior computational precision and higher time efficiency. This approach can serve as an alternative to the numerical simulation model in typical design procedures. The highest relative errors of P r and η were 0.034 and 0.038, respectively, both of which are significantly lower than the permissible error margins. Following the optimization process, four different structural designs were proposed for the remediation site. The proposed integrated simulation–optimization technique serves as a reference for the optimal configuration of GCWs.