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

[Display omitted] •1514 data points including 48 CBMs and 16 dyes used to train and test ANN model.•The applied ANN model showed higher R2 (0.98) and least MAE (20.74) and RMSE (46.95) values.•BET surface area, initial concentration and pyrolysis temperature were the most influential factors for dyes adsorption on CBMs.•Highest adsorption capacity of 2110.8 mg/L was predicted and optimized.•Jupiter notebooks to verify the reproducibility of current are provided. Organic waste-derived carbon-based materials (CBMs) are commonly applied in sustainable wastewater treatment and waste management. CBMs can remove toxic, non-biodegradable and carcinogenic pollutants such as dyes which include indigo, triphenylmethyl, azo, anthraquinone and phthalocyanine derivatives. Nonetheless, their diverse composition, surface properties, presence of numerous surface functional groups and the altering adsorption experimental conditions to which they are applied against the elimination of organic dyes make it challenging to completely understand the removal mechanism. Herein, a dataset of 1514 data points was compiled from various published peer-reviewed journals along with additional adsorption experiments conducted in this study. Artificial neural networks (ANN) based machine learning (ML) model was compared with other ML and a deep learning model named Tab-Transformer and the findings proposed ANN showed superior prediction performance for adsorption capacity as a function of adsorbent synthesis conditions, adsorbent physical characteristics and adsorption experimental conditions. The hyperparameters of ANN model was optimized using Bayesian optimizer and the batch size, activation and units were proven to be more important than the number of hidden layers and learning rate. The ANN model exhibits a higher coefficient of determination (R2 = 0.98) and lower root mean square error (RMSE = 46.95 mg/g) values for test dataset. Feature importance using SHapley Additive exPlanations (SHAP) analysis suggested that the adsorption characteristics with 51.4% was the most important in the ANN prediction followed by the adsorption experimental condition (31.2%) and adsorbent synthesis condition (17.4%). Moreover, the impact of six most important features were individually analyzed. Finally, a detailed discussion on the environmental impact of the presented ANN model is also included.