Machine learning and response surface methodology forecasting comparison for improved spray dry scrubber performance with brine sludge-derived sorbent

• Pulp and paper sludge waste contains a beneficial material composition relevant to flue gas desulfurization (FGD). • Spray-dry systems present more economical and operational benefits than wet FGD systems. • Machine learning fosters output predictions, which augments the optimization of industrial processes. In this study, hydrated lime (Ca(OH)₂) sorbent was prepared from industrial brine sludge waste using simple laboratory procedures and utilized in a laboratory-scale spray dry scrubber for desulfurization tests. The effects of key process parameters in spray drying (sorbent particle size, inlet gas phase temperature, and Ca:S ratio) on desulfurization efficiency were investigated using central composite design (CCD). Three machine learning (ML) models, multilayer perceptron (MLP), support vector regressor (SVR), and light gradient boosting machine (LightGBM), were assessed for their output estimation accuracy and compared to the CCD prediction model. The computational framework utilized experimental variables structured by CCD software as input metadata. Model performance was evaluated through generalization and accuracy measurements, including the coefficient of determination (R²), root mean square error (RMSE), mean square error (MSE), and mean squared logarithmic error (MSLE). Analysis of variance revealed that the Ca:S ratio had the most significant influence on SO₂ absorption. A quadratic model correlating the process variables with desulfurization efficiency was developed, yielding an R-squared value of 93.47%. Characterization of the final desulfurization products, particularly using XRD, showed the emergence of new phases such as hannebachite (CaSO 3 .0·5H 2 O), while FTIR analysis identified unreacted portlandite and calcite. Among the ML models, the MLP demonstrated superior performance over SVR and LightGBM, highlighting its efficacy in extracting and decoding information from the input data. The response surface methodology (RSM) model also proved to be a reliable forecasting tool, indicating its potential as a practical alternative to complex algorithmic computations in scenarios with limited raw data.