Transfer of multivariate calibrations for potentiometric multisensor systems to account for non-calibrated...

The deterioration in the predictive ability of calibration models when transitioning between different experimental conditions is a common analytical problem. Such circumstances as low long-term instrument stability, instrument replacement, or changes in the matrix composition of the analyzed samples (appearance of non-calibrated interferences) lead to necessity of sensor system recalibration, which is a long and tedious process. Moreover, in industrial settings where multisensor systems tend to be applied this recalibration is often impossible at all due to specific installation requirements and technological process regulations. Thus, there is a need for some mathematical correction for novel measurement conditions with unknown interference.

We have studied four multivariate calibration transfer protocols: Dynamic Orthogonal Projection (DOP), unsupervised Dynamic Orthogonal Projection (uDOP), Domain-Invariant iterative Partial Least Squares (DIPALS), and domain-invariant Partial Least Squares (diPLS) in extending multisensor calibration model applicability. These methods were applied to transfer the multivariate calibrations for predicting Nd³⁺ concentrations in a ternary nitric acid solution system (Nd³⁺, Eu³⁺, Er³⁺) mimicking the industrial media for lanthanides reprocessing. The initial conditions were further changed by addition of two non-calibrated interferences to the working solutions: Cs⁺ and Ba²⁺_. As expected, the initial regression models intended for quantification of Nd³⁺ failed in these new conditions, the root mean squared error in cross validation (RMSECV) increased from 0.24 to 1.86 (logC(Nd³⁺) [-5; -3]). Application of model transfer algorithms allowed for substantial improvement of prediction in the new conditions returning the RMSECV values almost back to the initial model (0.21 logC(Nd³⁺) in case of DIPALS).

The findings of this study indicate that non-calibrated interferences can be successfully handled in the framework of already developed calibration models in multisensor potentiometry. This can significantly enlarge applicability domains for multisensor systems such as e-noses and e-tongues.