Adsorption capacity prediction on carbon-based materials using deep learning

In this study we performed data-driven modeling of adsorption capacity on carbon-based materials by the use of artificial neural networks on a dataset of 1514 data points. The objective was to predict the adsoprtion capacity of various organic based activated carbon materials towards the elimination of industrial dyes from wastewater. ANNs are composed of neurons and weighted relationships between those neurons. Many of these ANN models have fully-connected layers, which implies that all of their neurons in the ANN architecture are connected to all of the neurons in the layer adjacent to it. we have used MLP for modelling and prediction of adsorption capacity of synthetic dyes on the surface of AC materials, which is a classical feed-forward ANN model. It operates on the idea of mapping given set of input data to the corresponding set of output data. The architecture of MLP is divided into three set of layers: input, hidden and output layer. The input receives the input features while the output layer returns the predicted target value. The hidden layers connect the input layers with the output layer through randomly initialized weights that are further optimized during training of the model. A non-linear activation function is used on the output of fully connected layer which enables the network to model non-linear relationship between input and output. The final value that comes out of the output layer is the predicted target value by the MLP.

Data

The input confined 12 input variables, such as the adsorption time (min), type of adsorbent, calcination temperature (oC), calculation time (min), type of dye, initial dye concentration (mg/L), solution pH, adsorbent loading (g), volume (L), adsorption temperature (oC), particle size (nm), surface area (m2/g) and pore volume (cm3/g). The output variable was the adsorption capacity (mg/g). The entire dataset contains two categorical input features, type of dye and type of adsorbent, which was changed into numerical values using the one-hot encoding technique. Afterwards, the entire dataset was randomly divided into two sets of 70:30, from that 70% (1059 data points) data was applied to train the ANN model, while the remaining 30% (455 data points) of data was used as test data. 30% of the training data was considered as validation set and was used for hyperparameter optimization. A comprehensive analysis of data is given in 1. EDA

Results

various performance metrics including the root-mean-square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) was applied to study the model performance. Pearson correlation results showed that the adsorption capacity prediction was positively correlated with the pore volume, BET surface area, initial concentration, solution pH, and negatively correlated with the volume, adsorbent loading, calcination time and temperature for adsorbent synthesis. Feature importance using SHAP analysis suggested that the adsorption characteristics with 51.4% was the most imported in the ANN prediction followed by the adsorption experimental condition (31.2%) and adsorbent synthesis condition (17.4%). The ANN model prediction performance exhibited outstanding results based on highest R2 values and lowest errors. Moreover, SHAP feature importance analysis suggested that the adsorbent characteristics and adsorption experimental conditions were equally important for model prediction. The SHAP dependency analysis was further applied to model the impact of six most important input features.

Reproducibility

To replicate the experiments, you need to install all requirements given in requirements file . If your results are quite different from what are presented here, then make sure that you are using the exact versions of the libraries which were used at the time of running of these scripts. These versions are given printed at the start of each script. Download all the .py files in the scripts including utils.py (utils) file. The data is expected to be in the data folder under the scripts folder.

Scripts

Paper

Artificial neural networks for insights into adsorption capacity of industrial dyes using carbon-based materials

https://doi.org/10.1016/j.seppur.2023.124891

If you use this work in your research, consider citing it using following BibTeX entry

@article{iftikhar2023artificial,
      title={Artificial neural networks for insights into adsorption capacity of industrial dyes using carbon-based materials},
      author={Iftikhar, Sara and Zahra, Nallain and Rubab, Fazila and Sumra, Raazia Abrar and Khan, Muhammad Burhan and Abbas, Ather and Jaffari, Zeeshan Haider},
      journal={Separation and Purification Technology},
      volume={326},
      pages={124891},
      year={2023},
      publisher={Elsevier}
}