Inverse least-squares modeling of vapor descriptors using polymer-coated surface acoustic wave sensor array responses

In previous work, it was shown that, in principle, vapor descriptors could be derived from the responses of an array of polymer-coated acoustic wave devices. This new chemometric classification approach was based on polymer/vapor interactions following the well-established linear solvation energy relationships (LSERs) and the surface acoustic wave (SAW) transducers being mass sensitive. Mathematical derivations were included and were supported by simulations. In this work, an experimental data set of polymer-coated SAW vapor sensors is investigated. The data set includes 20 diverse polymers tested against 18 diverse organic vapors. It is shown that interfacial adsorption can influence the response behavior of sensors with nonpolar polymers in response to hydrogen-bonding vapors; however, in general, most sensor responses are related to vapor interactions with the polymers. It is also shown that polymer-coated SAW sensor responses can be empirically modeled with LSERs, deriving an LSER for each individual sensor based on its responses to the 18 vapors. Inverse least-squares methods are used to develop models that correlate and predict vapor descriptors from sensor array responses. Successful correlations can be developed by multiple linear regression (MLR), principal components regression (PCR), and partial least-squares (PLS) regression. MLR yields the best fits to the training data, however cross-validation shows that prediction of vapor descriptors for vapors not in the training set is significantly more successful using PCR or PLS. In addition, the optimal dimension of the PCR and PLS models supports the dimensionality of the LSER formulation and SAW response models.