Accelerated estimation of chemical and sensory liquid attributes using an AI-assisted electrochemical electronic tongue

Abstract

Electronic tongues based on arrays of potentiometric sensors bears promises for enabling portable and fast analysis of complex liquids by leveraging low-selectivity and high-sensitivity of the inherent sensing materials. Thus, they have been studied and developed to target multiple applications, including analysis of chemical compositions or for untargeted quality control in the food industry. A key element to enable practical use of electronic tongues, however, has often been the analysis and interpretation of their combinatorial response due to the cross-sensitivity of the employed sensors. Indeed, disentangling the effect of multiple liquid constituents on the potentiometric response is the main associated challenge. The present contribution aims at studying the feasibility and accuracy of employing this class of sensors to both the quantification of multiple analytes in water as well as prediction of sensory attributes in coffee samples. A proof-of-concept system comprises an integrated array of miniaturized polymeric sensors and an electronic board records 15 differential voltages, produced by the interaction with both ionic and organic constituents of liquids, and can transmit them via Bluetooth to an external device. An automated pipeline performs pre-processing, hand-crafted feature extraction and training of machine learning models that can be then deployed on the cloud or edge device (e.g., smartphone). The same array can be then reconfigured to address a new specific task by leveraging new potentiometric data collection and applying the automated data analysis. In particular, the device presented herein employed 16 polymeric sensors that were electrodeposited on conventional electroless nickel immersion gold (ENIG) electrodes. The conductive polymers (PEDOT, PPy, PANI and PAPBA) were synthesized by chronoamperometry or cyclic voltammetry and enriched with doping agents for enhanced sensitivity. Differential voltages between these polymeric sensors were measured during the transition of the sensor array from a reference solution to a test solution, thereby obviating the need for a conventional reference electrode. Indeed, the use of low-selective sensors for potentiometric measurements does not require integration of reference electrodes, which are known to be unpractical for remote sensing applications. The analysis of the potentiometric response revealed deviations from linear sensitivities, showing non-Nernstian voltage trends when varying concentration of target analytes. Training data were obtained by alternatively immersing the sensor array in reference (120 s) and test (60 s) solutions continuously using an automated test rig. In addition to the conventional approach of measuring the equilibrium potential after a certain settling time, the complete evolution of the potentiometric signal in time could provide richer information that could be useful for building calibration models. Hence, hand-crafted features were extracted from time series recorded during sensor training in order to reduce the dimension of data and describe the complete voltage perturbation of the sensor during transition between reference and test solution. If exposed to enough “known liquid examples”, the sensor array can learn patterns to perform accelerated qualitative and quantitative analysis. The device has demonstrated enhanced discrimination of various mixtures and was proved able to predict beverage sensory properties after intensive training with samples with known sensory attributes. Firstly, the sensitivity to multiple ionic species was estimated in model mixtures obtained following an Orthogonal Experimental Design (OED) for a set of four metal ions, including $Al^{3+}$, $Cu^{2+}$, $Na^+$ and $Fe^{3+}$, and another set comprising $Ca^{2+}$, $Mg^{2+}$ and $Na^+$. The former experiment demonstrated that non-linear regression models, such as a Random Forest of Extra Trees, can cope with non-linearity of the sensor response and yielded better performances than widely used multivariate regression model i.e., Multiple Linear Regression (MLR). The mean relative quantification error varied between 1% for $Fe^{3+}$ and 44% for Na+ in the concentration range 1-10 mg/L. The second mixture set was used instead to unveil the cross-sensitive response of the sensor array, build a quantification model to predict major cations in mineral water and proving the importance of feature selection for data-driven sensors. Overall, the sensor showed comparable accuracy to ICP-MS (MRE 12-24%) but with a drastic reduction of cost and processing time. Finally, the same sensor array configuration was also used to test 12 coffee samples, which were also evaluated by a sensory panel through 5 taste attributes. Regression models were built to map instrument response to sensory data and Extra-Tree algorithm yielded the highest performances with a MRE varying from 8% to 21% depending on coffee sample and validation scheme. Thus, it was demonstrated that data-driven sensor arrays could be also valuable alternative tools for accelerated evaluation of beverage sensory characteristics and support food innovation processes.