Response surface methodology and support vector machine for the optimization of separation in linear gradient elution

Abstract

Experimental design methodology was used to optimize the linear gradient elution chromatography. The effect of initial mobile phase composition (phi(in)), initial isocratic time (t(in)), and gradient time (t(G)) on the retention times of phenyl thiohydantoin aminoacids (PTH-amino acids) was investigated. The experiments were performed according to Box-Behnken experimental design to map the chromatographic response surface. Then multiple linear regression and support vector machine (SVM) methods were used to fit the retention times of solutes. Results shows the SVM models have better ability to predict retention time and used for grid search in factor space. The SVM models were verified, as good agreement was observed between the predicted and experimental values of retention time in the optimal condition.

Response surface methodology using Gaussian processes: Towards optimizing the trans-stilbene epoxidation over Co2+–NaX catalysts

“…An alternative approach is support vector machine (SVM) that belongs to the family of kernel modelling methods [20]. SVM employs a structural risk minimization scheme to improve the prediction accuracy, and it has been successfully applied to predictive modelling of catalysis processes [21,22] and other chemical systems [23].…”

Interpretation of non-linear empirical data-based process models using global sensitivity analysis

“…Since the model is restricted to the polynomial form, it is not capable of approximating complex factor-response relationship. To address this issue, more flexible data-based models have been adopted, including ANN [6,10], SVM [7] and GP regression [9]. Unlike polynomial regression, these complex models do not allow a transparent interpretation as to which process factors contribute the most to the response variable.…”

“…logarithmic or logistic) transformation is particularly attractive if it is known a priori to result in linear factor-response relationship, and thus linear regression can be used. However in general situation, the prediction accuracy of the polynomial regression is unsatisfactory if the chemical process is complex and does not conform to the restrictive functional form [6,7,8,9]. Consequently, the model-based process understanding and optimization may be unreliable .…”

“…Consequently, the model-based process understanding and optimization may be unreliable . To address this issue, flexible non-linear models have been applied to provide a more accurate approximation of the process behavior, such as artificial neural network (ANN) [6,10], support vector machine (SVM) [7], and Gaussian process (GP) regression 1 [9]. Here "flexible" refers to the property that the models are not restricted to a certain functional form, but capable of approximating any function to any accuracy should sufficient data are available.…”

Model-aided optimization and analysis of multi-component catalysts: Application to selective hydrogenation of cinnamaldehyde

“…The current trend is to replace the traditional polynomial regression with more flexible models that can approximate the actual process more accurately. In the past few years, artificial neural network (ANN) (Huang et al, 2001;Omata et al, 2006;Coleman and Block, 2007), support vector regression (SVR) (Hadjmohammadi and Kamel, 2008) and Gaussian process regression (GPR) (Omata, 2011;Yuan et al, 2008;Yan et al, 2011b,a) have emerged as popular choices. A good model should not only provide accurate prediction, but also faithfully quantify its own prediction uncertainty, which is an important measure to guide process optimization (Jones, 2001).…”