Accurate estimation of polyethylene glycol density via machine-learning based techniques

Polyethylene glycol (PEG) has been globally recognized as an environmentally-friendly chemical solvent used in many disciplines for various purposes. In this work, intelligent models are constructed based upon least squares support vector machine (LSSVM) and adaptive neuro-fuzzy inference system (ANFIS) methodologies optimized with either genetic algorithm (GA), coupled simulated annealing (CSA) or particle swarm optimization (PSO) to estimate PEG density in terms of PEG molecular weight, temperature, and pressure based upon data gathered from experimental works delineated in the published literature. Leverage method is performed on the acquired dataset to explore it in terms of outlier datapoints, and relevancy factor is used to perform sensitivity analysis. Graphical and statistical indexes are used to evaluate the authenticity of the developed models. The results show that nearly all intelligent models are accurate, with LSSVM-CSA being the most accurate model, which outperforms the modified Tait equation as outlined by the calculated mean square error, average absolute relative error, and R-squared values. In addition, the performed sensitivity analysis indicates that temperature is the most effective input variable with an indirect relationship. The developed intelligent models, particularly the LSSVM-CSA model, are highly capable of predicting PEG density without needing experimental approaches that are known to be arduous and laborious.