CSRR chemical sensing in uncontrolled environments by PLS regression

Complementary Split Ring Resonators (CSRRs) have been extensively studied as planar sensors in the last two decades. However, their practical use remains limited to controlled environments and classification problems. Their performance reliance on high-end Vector Network Analyzers (VNAs), highly repeatable laboratory conditions, and special sample holders or microfluidic circuits hinders its regular use in chemistry laboratories as analytical tool. Temperature drifts and humidity variations during measuring, uncertainties in the electromagnetic properties of the sample containers, and careles sample handling introduce significant uncertainties in measurements, leading to unreliable results. Therefore, the prediction of target compounds concentration in samples have been out of the research focus up to now. Machine Learning algorithms can help to mitigate these uncertainties and open the applicability of CSRR sensors to quantification problems, where is necessary to determine the amount of a substance in a liquid (or solid) sample. This work presents a novel approach that takles this issue, combining a CSRR sensor with well stablised ML algorithms that enhances its quantification performance. For ilustration purposes, a low-cost, benchtop CSRR-based system is proposed to predict ethanol concentration in water solutions. Ethanol samples from 10% to 96% concentration were prepared in commercial vials, generating 450 randomized measurements. Principal Component Analysis (PCA) was employed for data exploration, while a Partial Least Squares regression model (PLS), tuned with Leave-One-Group-Out Cross-Validation, was trained for ethanol concentration prediction. No feature extraction technique or noise reduction strategy was applied. Although this straightforward workflow is well known in the chemical sensing field, it has not been applied to data acquired with CSRR sensors. The trained model achieved a Root Mean Square Error in Prediction (RMSEP) of 3.7%. Compared with 23.4% RMSEP when using univariate calibration at optimized frequencies, it presentes a prediction performance reduced by a factor of 6. No evidence of underfitting or overfitting was observed during test of the trained model. The low RMSEP achieved by the presented setup demonstrates the potential of CSRR-based sensors when combined with ML techniques for concentration prediction working in realistic, uncontrolled conditions. This pushes forward the applicability of CSRR sensors in the chemical analysis field, which might lead to benchtop, lowcost and relaible analysis devices for many laboratories.