Deep neural network algorithm for predicting the reaction performance of Cr(VI) removal by S-ZVI

Sulfidation to zero-valent iron (S-ZVI) is efficient and promising strategy in environmental remediation. Due to the complexity of reaction system, many factors such as the reaction system pH, S-ZVI dosage, etc. can dramatically affect Cr(VI) removal when S-ZVI was used to sequestrate Cr(VI). To obt...

Full description

Saved in:
Bibliographic Details
Published in:Separation and purification technology Vol. 380; p. 135301
Main Authors: Wang, Xiao, He, Bo, Li, Jun, Zhang, Yanshi, Zhang, Yiqiao, Wang, Chunguang, Zhao, Chengxuan, Xu, Chunhua
Format: Journal Article
Language:English
Published: Elsevier B.V 07.02.2026
Subjects:
Deep neural network
Prediction
Pseudo-first-order kinetic constant
Reaction condition
Sulfidated zero-valent iron
Pseudo-first-order kinetic constant
Deep neural network
Reaction condition
Sulfidated zero-valent iron
Prediction
ISSN:1383-5866
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Sulfidation to zero-valent iron (S-ZVI) is efficient and promising strategy in environmental remediation. Due to the complexity of reaction system, many factors such as the reaction system pH, S-ZVI dosage, etc. can dramatically affect Cr(VI) removal when S-ZVI was used to sequestrate Cr(VI). To obtain the optimal performance of Cr(VI) removal, tedious and heavy experimental work must be done in the lab, which would cost a lot of manpower, material resources and time. To solve this problem, a model for predicting Cr(VI) removal performance was constructed. Deep learning was used to build a model which based on the existing experimental data through the use of multiple hidden layers and nonlinear activation functions enable the prediction of unobservable data. In this work, deep neural network (DNN) algorithm was selected to build the model to predict pseudo-first-order kinetic constant kobs of Cr(VI) removal, with reaction conditions as input features. The kobs at different initial pH and S-ZVI dosage were predicted and the R2 and MSE were 0.9960 and 4.1 × 10−5, respectively. The results showed that the model has the optimal performance in predicting kobs with low mean square error (MSE) and high R2 when the number of hidden layers was two and each with 100 and 8 neurons. In addition, the trained DNN model successfully predicted the kobs at unobserved conditions. This study provides a data-driven predictive model which can reduce labor and reagent costs by reducing experimental effort, making S-ZVI more economical and effective in industrial applications. [Display omitted] •DNN algorithm was used to predict pseudo-first-order kinetic constant kobs•Initial pH and S-ZVI dosage were input as dataset for the model with DNN algorithm•Prediction results by DNN were compared with calculated kobs from experimental data•Established and trained DNN model precisely predicted kobs with high R2 and low MSE
ISSN:1383-5866
DOI:10.1016/j.seppur.2025.135301