Machine Learning and Corona Effect Coupling, a New Tool for Water Quality Control: Highlighting the Impact of Microwaves on Water Samples
Faniry Raveloson1, Mickaël Gresse1, Laurent Vandanjon2*, Hugues de la Bardonnie1
1LIMEC Laboratory, 24 rue du Général Ferrié, 31500 Toulouse, France
2Laboratory of Marine Biotechnology and Chemistry (LBCM), University of South Brittany (UBS), EMR CNRS 6076, IUEM, Campus Tohannic, 56000 Vannes, France
*Corresponding author: laurent.vandanjon@univ-ubs.fr
Keywords: corona effect, boosting, bagging, weak learner, strong learner, principal component analysis
Submitted: August 9, 2025
Revised: May 4, 2026
Accepted: May 14, 2026
Published: September 24, 2026
Abstract
L’Imagerie Macroscopique à Effet Couronne (LIMEC) was used to investigate whether microwave exposure produces detectable differences in water samples from the same commercial source. Corona-discharge images of untreated and microwave-treated bottled waters (800 W, 30 s) were acquired under controlled conditions, and 102 image-derived descriptors were extracted. Feature selection combined ANOVA, LASSO regularization, correlation analysis, and Random Forest importance ranking, while classification was performed using XGBoost.
The model achieved accuracies of 74% and 78% with ROC–AUC values of 0.85 and 0.86 for the Montcalm and Mont Roucous datasets, respectively. When both water sources were combined, classification accuracy reached 80%. The results demonstrate reproducible discrimination between untreated and microwave-treated samples based on their macroscopic corona-discharge patterns. These findings indicate that LIMEC coupled with machine learning is sensitive to subtle differences associated with distinct treatment histories, without implying permanent molecular-scale modifications of water structure.
Introduction
L’Imagerie Macroscopique à Effet Couronne (LIMEC) was used to investigate whether microwave exposure produces detectable differences in water samples from the same commercial source. Corona-discharge images of untreated and microwave-treated bottled waters (800 W, 30 s) were acquired under controlled conditions, and 102 image-derived descriptors were extracted. Feature selection combined ANOVA, LASSO regularization, correlation analysis, and Random Forest importance ranking, while classification was performed using XGBoost.
The model achieved accuracies of 74% and 78% with ROC–AUC values of 0.85 and 0.86 for the Montcalm and Mont Roucous datasets, respectively. When both water sources were combined, classification accuracy reached 80%. The results demonstrate reproducible discrimination between untreated and microwave-treated samples based on their macroscopic corona-discharge patterns. These findings indicate that LIMEC coupled with machine learning is sensitive to subtle differences associated with distinct treatment histories, without implying permanent molecular-scale modifications of water structure.
The characterization and monitoring of water quality remain major challenges in environmental science, industrial process control, and resource management. Although conventional analytical techniques provide detailed chemical information, they often require large sample volumes, complex preparation procedures, and laboratory-based instrumentation. Consequently, there is increasing interest in complementary approaches capable of rapidly detecting subtle and reproducible variations in water behavior using small sample volumes and non-destructive measurements. Among these approaches, imaging-based methods have attracted growing attention because of their ability to capture integrated and multivariate responses of complex systems under controlled conditions. Gas discharge and corona effect imaging techniques are known to be sensitive to environmental parameters and material properties. When applied to liquid droplets, corona discharge patterns arise from the interaction between a high-voltage electric field, the surrounding gas, and the physicochemical properties of the liquid, providing a macroscopic signature of the global behavior of the system rather than direct access to molecular-scale structures. Recent advances in image processing and machine learning have greatly enhanced the analytical potential of such imaging techniques. By transforming complex visual patterns into quantitative descriptors, supervised classification methods enable the detection of subtle but systematic differences between experimental conditions within a multivariate framework, which is particularly well suited to systems such as liquid water.
Liquid water is commonly described as a highly dynamic hydrogen-bond network whose local organization continuously fluctuates on the picosecond timescale. Recent theoretical and experimental work has proposed that transient cluster-like arrangements may contribute to several anomalous properties of water, while remaining short-lived and strongly temperature-dependent (Andersson, 2023). Importantly, the present study does not attempt to directly probe such microscopic arrangements but rather investigates whether reproducible macroscopic differences can be detected using imaging-based and data-driven approaches. In a previous study published in WATER (Raveloson et al., 2024), we showed that macroscopic corona effect imaging combined with multivariate analysis can discriminate between chemically distinct aqueous solutions as well as between chemically similar bottled waters with low mineral content. These results suggest that corona discharge imaging captures reproducible macroscopic features related to water properties under standardized conditions.
Building on these findings, the present study investigates a more experimentally stringent scenario: the discrimination between two water samples of identical chemical composition originating from the same commercial bottle and differing only by a brief microwave exposure. Microwave radiation at 2.45 GHz interacts strongly with liquid water, primarily through dielectric heating and dipolar relaxation mechanisms, which convert electromagnetic energy into thermal motion of dipoles and ions (Metaxas and Meredith, 1983; Ellison et al., 1996). Although transient reorganizations of the hydrogen-bond network may occur during heating, the existence of permanent or long-lived non-thermal microwave effects in bulk water remains controversial.
Accordingly, the objective of this work is not to claim or demonstrate persistent microscopic structural modifications of water induced by microwave exposure. Rather, we adopt a pragmatic and operational approach aimed at determining whether reproducible and statistically detectable differences can be observed at the macroscopic imaging level between untreated and microwave-treated water samples under controlled conditions. To this end, we combine macroscopic corona effect imaging using the LIMEC/EDS© system with supervised machine learning techniques. Corona images are converted into feature vectors describing complementary aspects of intensity, spatial organization, heterogeneity, and entropy-related complexity. The methodology is evaluated on two independent commercial bottled water brands, Montcalm and Mont Roucous, as well as on a combined classification task, in order to assess the robustness and transferability of the proposed approach.
The Basis of Electrophotonics: Understanding Light and Water
Light: The hypothesis that light is composed of particles and is therefore of a corpuscular nature was advanced by Sir Isaac Newton (Newton, 1952) in 1704. Through his numerous experimental studies describing many physical phenomena, he rejected the wave theories of light proposed by R. Hooke, C. Huygens, and L. Euler at that time.
In the 19th century, however, experimental observations such as T. Young’s interference experiments (Young, 1802) and the discovery of polarization by EL Malus (Malus, 1811) suggested a wave nature to light. Thanks to A. Fresnel’s wave theory (Fresnel, 1826) and, finally, the unification of electricity and magnetism by JC Maxwell (Maxwell, 1873), light was then considered as a wave, more precisely an electromagnetic wave.
At the end of the 19th century, with Max Planck’s work and his quantum energy hypothesis (Planck, 1900), the classical vision of matter underwent a revolution. The work of Poincaré, Lorentz, and then Einstein on special relativity and his mass-energy equivalence relation E = γ.mc² (where c is the speed of light in vacuum and γ is the Lorentz factor) (Einstein, 1905a) questioned the nature of matter. In 1905, Einstein also questioned the nature of light, despite it being so well described by Maxwell through his work on the photoelectric effect (Einstein, 1905b).
This last exploration of wave-particle duality leads us to consider that the nature of physical reality is not intuitive. In the case of light, it can be observed either as a wave structure or as a particle (photon). Under these conditions, the corona effect can be considered the visualization of photonic particle emission caused by the electrical discharge applied to an object.
Water: Understanding water structure and its properties is essential to understand the mechanism by which water contributes to photon emission during an electrophotonic discharge.
Wiggins (Wiggins, 2008) showed that water is composed of two liquids of high and low density. This was validated by international research teams using recent X-ray diffraction techniques (Perakis et al., 2017). At liquid-solid or liquid-gas interfaces, there is a zone where water is comparable to a negatively charged gel, the “Exclusion Zone” (EZ) described by G. Pollack (Pollack, 2013). Indeed, at interfaces, water molecules present a dynamic organization in rings or chains (Vandanjon, 2021).
Water molecules assemble very transiently due to hydrogen bonds (lifespan 10–12 to 10–13 s), forming structures resembling liquid crystals (clusters). Based on quantum electrodynamics theory, Preparata (Arani et al., 1995) describes clusters as coherence domains within which information transfer is possible and explains how a specific electromagnetic wave can remain “confined” in coherent water (Bono et al., 2012). JG Watterson explains that this is a vibrational energy transfer (quantum oscillation of a particle in a crystal lattice, also called a phonon) without any matter displacement in water coherence domains (Watterson, 1987), similar to the proton transfer mechanism described by Theodor von Grotthuss in 1806 (Von Grotthuss, 1806).
An important point to emphasize is that these theories have been experimentally validated in Coudert’s thesis (Coudert, 2007), who observed the electron trajectory in water in confined environments using ultrafast laser spectroscopy. Regarding photon transfer, a similar mechanism has been proposed but is much more difficult to explain (Henry, 2016). However, Voeikov’s work has shown that interfacial water can emit light (Voeikov and Korotkov, 2017) through an oxidative reaction initiated by energy input. Based on this, one could assume that water containing more colloids (more liquid-solid interface) or more nanobubbles (more liquid-gas interface) should theoretically emit more photons or transmit light more easily when an electrical discharge is produced by the Electrophotonic device. Similarly, electromagnetic vibrations have a resonance relationship with hydronium ions (H3O+) in water, which can function as receptors of vibrational information (Sbitnev, 2016).
Microwave Exposure and Water: Microwave radiation at the frequency used in domestic ovens (2.45 GHz) interacts strongly with liquid water due to its high dielectric permittivity and permanent dipole moment (Gabriel et al., 1998). The dominant and well-established interaction mechanism is dielectric heating, resulting from dipolar relaxation and ionic conduction processes that convert electromagnetic energy into thermal energy (Metaxas and Meredith, 1983). Consequently, microwave irradiation leads to volumetric heating of water, with temperature increases that may be spatially and temporally heterogeneous depending on exposure conditions (Meredith, 1998). At the molecular level, liquid water is characterized by a highly dynamic hydrogen-bond network with bond lifetimes on the picosecond timescale (Hill and Roth, 2011; Soper, 2013). This network is strongly influenced by temperature, dissolved ions, gases, and interfaces (Chaplin, 2006). Numerous spectroscopic studies using infrared, Raman, and terahertz techniques have shown that increasing temperature induces progressive weakening and reorganization of hydrogen bonds, manifested as changes in vibrational band shapes and intensities (Sun, 2009; Soper, 2013). From this perspective, many reported modifications of water “structure” following microwave exposure can be attributed to thermal effects alone (Chaplin, 2006; Sun, 2009).
Beyond this classical framework, the possibility of so-called non-thermal microwave effects—i.e., effects not fully attributable to temperature changes—has been explored in the literature (Pakhomov and Murphy, 2000; Kaatze, 1989). However, isolating such effects from thermal influences remains experimentally challenging, and their interpretation is still debated (Foster and Glaser, 2007; Knoepfel, 2000). Differences in experimental protocols, temperature control, container materials, dissolved gases, and post-irradiation relaxation times frequently lead to contradictory conclusions (Bohr and Bohr, 2000). Consequently, no consensus currently exists regarding permanent or long-lived structural modifications of bulk liquid water induced by microwave irradiation under typical conditions.
In this context, an increasing number of studies adopt a pragmatic and operational perspective, focusing not on defining microscopic structural changes, but on identifying reproducible and measurable differences between experimental conditions using sensitive observational techniques and statistical analysis (Guyon and Elisseeff, 2003; Bishop, 2006).
In line with this pragmatic perspective, microwave exposure was selected in the present study as a simple and well-defined electromagnetic perturbation, primarily associated with the dielectric heating of water at 2.45 GHz. The objective was not to investigate microwave effects per se, but to evaluate the sensitivity of the proposed imaging and classification framework to controlled perturbations.
LIMEC, Macroscopic Corona Effect Imaging Studies:
As explained in the introduction, the corona effect has numerous applications, including gas-discharge visualization. Depending on the device used, extensive research has been conducted in the fields of health and well-being, notably by Korotkov (2013). The use of artificial intelligence in the analysis of electrophotonic images has also been explored for disease detection (Janadri et al., 2017) and, more generally, for biomedical applications.
The EDS© Device: EDS© (Electrophotonic DataPhoton System) is an instrument initially designed to perform macroscopic corona imaging applied to a variety of subjects, from living body parts to liquid solution drops and minerals.
This prototype device was manufactured in 3 units and was developed over several years. Since 2009, it has been used for various projects by “Electrophotonique Ingénierie” until its acquisition by “SARL Développement Durable” in 2020 to conduct work combining machine learning and parameterization of corona effect images.
It is designed to pass an oscillating electromagnetic field of 10 mA at a fixed voltage between 8–15 kV and a frequency between 1 and 400 Hz on a pure quartz electrode to generate the corona effect on the experimental object.
The device (see Figure 1) consists of an AEPG© (Advanced Electrophotonic Generator), an EFUSE© (Electrode for Use in Specific Electrophotonic), and a Hamamatsu HD camera (ORCA IIBT 512G2) coupled with optics equipped with a 250–380 nm UV filter. The AEPG© (Advanced Electrophotonic Generator) relies on a specific geometry of its power transformers. Coupled with high-precision electronic components, it generates a pulsed high-voltage electromagnetic field on the electrode plate. The resulting field alternates between positive and negative polarity. In the present study, the excitation frequency was fixed at 90 Hz. The operating parameters of the system, including voltage and frequency settings, as well as synchronization with the Hamamatsu camera, are managed by the self-developed Advanced Electro Photonic Generator© software.
Macroscopic Corona Effect Imaging of Liquid: When applied to liquid droplets, corona discharge patterns reflect complex interactions between the applied electric field, the surrounding gas, and the physicochemical properties of the liquid (Loeb, 1965; Fridman, 2008). Previous studies have shown that corona discharge photography is particularly sensitive to subtle variations in environmental and material conditions, including humidity, pressure, surface properties, and electrical conductivity, making it a potentially powerful—but also experimentally demanding—tool (Raizer, 1991; Peek, 1929; Chen and Davidson, 2002). Consequently, strict control of acquisition conditions such as temperature, hygrometry, atmospheric pressure, and electromagnetic noise is essential to ensure reproducibility (Goldman et al., 1985). Importantly, corona imaging does not provide a direct microscopic probe of molecular structure (Laroussi, 2009). Instead, it offers a macroscopic, integrative response that can be interpreted as a global signature of the interaction between the liquid, the surrounding gas, and the applied electric field under fixed experimental conditions (Chen and Davidson, 2002). In this sense, differences observed in corona discharge images should be interpreted as phenomenological patterns rather than direct evidence of specific molecular rearrangements (Fridman et al., 2008).
Image Descriptors and Supervised Classification: Recent advances in image processing and machine learning enable complex visual patterns to be transformed into quantitative descriptors capturing complementary properties such as global intensity, spatial organization, heterogeneity, and multiscale complexity. Supervised classification models exploit these multivariate representations to detect subtle differences between experimental conditions without relying on a single dominant descriptor. Dimensionality reduction methods, such as principal component analysis (PCA), are primarily used for visualization, while classification is performed in the full feature space.
To this end, we developed the FaHJe–IMEC framework, which integrates controlled corona imaging, feature extraction from image sequences, multi-stage variable selection, and supervised classification. The methodology was evaluated on two independent commercial bottled water brands, Montcalm and Mont Roucous, as well as on a combined classification task grouping different water sources. By using identical acquisition protocols and the same classification model configuration across datasets, we specifically assessed the robustness and transferability of the approach.
Methods
The methodological workflow consisted of five main steps: (i) sample preparation, (ii) corona image acquisition, (iii) image processing and feature extraction, (iv) variable selection, and (v) supervised classification and performance evaluation.
Sample Preparation
Two water qualities were investigated: a commercial bottled water sample and the same water subjected to microwave exposure. For each condition, seven independent batches were prepared, each consisting of a 150 mL sampling tube. Only microliter-scale droplets (20 µL) were used for imaging.
For the microwave-treated condition, the water was poured into a glass beaker and exposed to microwave radiation at a nominal power of 800 W for 30 s. After irradiation, the water was transferred into sampling tubes, sealed, and allowed to cool naturally to room temperature. All subsequent handling, storage, and imaging conditions were identical for both water qualities.
A minimum two-hour relaxation period was applied prior to image acquisition in order to allow transient thermal gradients and short-lived hydrogen-bond reorganizations to relax, thereby minimizing the contribution of immediate thermal effects to classification.
Corona Imaging Device and Acquisition Protocol
Corona discharge images were acquired using the EDS© device. For each batch, twenty droplets were generated and analyzed independently. Droplets were deposited using a pipette equipped with a conductive nozzle, ensuring reproducible electrical contact with a transparent conductive electrode. A custom-designed pipette holder was used to minimize background intensity and stray light.
Each droplet was stimulated by applying a high-voltage electric field at a frequency of 90 Hz and a voltage of 11 kV for a duration of 2000 ms. Corona emission was recorded using an optical camera between 30,000 ms and 35,000 ms after stimulating onset. For each droplet, between 20 and 22 successive stimulations were recorded, corresponding to distinct configurations of the droplet over time. Approximately 400 images were acquired per batch, resulting in a dataset of about 2800 images per water quality. Images were stored in 16-bit TIFF format to preserve dynamic range and enable quantitative analyses.
Image Processing and Feature Extraction
Each corona image was transformed into a numerical feature vector. Extracted descriptors covered complementary families of properties, including global intensity statistics, spatial dispersion measures, structural organization indicators, streamer-related features, and entropy-based multiscale descriptors. All images were thus converted into a data matrix in which each row corresponds to an observation and each column to a variable. A binary class label was associated with each observation to indicate the experimental condition.
Variable Selection
Given the relatively high dimensionality of the feature space compared to the number of observations, feature selection was necessary to reduce model complexity, limit overfitting, and improve generalization performance.
A multi-stage feature selection strategy was implemented to reduce dimensionality and limit redundancy among variables. First, a univariate screening based on analysis of variance (ANOVA) was applied to identify variables exhibiting statistically significant differences between classes. Variables with a p-value below the significance threshold (α = 0.05) were retained.
In parallel, an embedded multivariate selection was performed using L1-regularized logistic regression (LASSO). This method enables simultaneous classification and feature selection by promoting sparse models through the shrinkage of coefficients associated with redundant or weakly informative variables. This approach relies on the minimization of a penalized logistic loss function as seen in Equation 1.
The variables retained by ANOVA and LASSO were subsequently examined through correlation analysis. Several groups of strongly correlated variables were identified, reflecting redundancy within the feature space. To mitigate multicollinearity, only a limited number of representative variables were retained from each correlated group, based on their statistical relevance and stability across selection procedures.
In addition to these statistical and regularization-based methods, an independent feature selection procedure based on Random Forest variable importance was applied to the full set of descriptors. Random Forests are particularly well suited to high-dimensional and heterogeneous data, as they capture non-linear relationships and feature interactions without assuming linear separability. Feature importance was quantified using the mean decrease in impurity criterion. Given the distribution of importance values, a relative ranking strategy was adopted, and the most informative variables were identified based on their importance ranking rather than on an absolute threshold.
Classification and Performance Evaluation
Given the expected non-linear and multivariate nature of corona discharge patterns, ensemble boosting methods were considered particularly well suited for the present classification task. Among the evaluated models, boosting-based classifiers consistently exhibited the best overall performance. In particular, the XGBoost algorithm achieved the highest classification accuracy and Receiver Operating Characteristic Area Under the Curve (ROC–AUC) value, together with the most stable results across stratified five-fold cross-validation. This superior performance can be explained by several properties that match the characteristics of the dataset.
First, the extracted image descriptors define a heterogeneous and partially correlated feature space, combining statistical, spatial, and entropy-related information. Unlike linear classifiers, XGBoost does not assume linear separability and is capable of modeling complex non-linear relationships as well as higher-order interactions between variables. Second, the gradient boosting framework on which XGBoost is based iteratively emphasizes difficult-to-classify observations, which is especially advantageous when class differences are subtle and distributed across multiple descriptors rather than driven by a single dominant feature.
Moreover, XGBoost incorporates multiple built-in regularization mechanisms, including constraints on tree depth, learning rate control, and row and column subsampling. These strategies effectively limit overfitting and improve generalization in high-dimensional settings, which is particularly relevant given the relatively large number of variables compared to the number of observations. The observed stability of performance across cross-validation folds further supports the robustness and generalization capability of the model.
Based on both these theoretical considerations and the empirical results, XGBoost was retained as the final classifier for the Montcalm and Mont Roucous datasets. The same model architecture and hyperparameter configuration were applied to both datasets to ensure methodological consistency and to enable direct and fair comparison of results.
Results
Descriptive Analysis and Correlation Structure
A descriptive statistical analysis was performed to characterize the dataset and examine relationships among variables and with the class label. The correlation matrix of the extracted descriptors reveals several groups of highly correlated variables, indicating multicollinearity within the feature space. This behavior is expected, as many descriptors originate from related physical and statistical properties. In contrast, no strong individual linear association is observed between single variables and the class label, suggesting that class discrimination arises from multivariate and potentially non-linear feature interactions rather than isolated effects.
Variable Selection Results
The variables ranked highest by Random Forest included meanStream, hs_DistMean, sumEntropyLog2Line, hs_DistHigh10, firstDecay, and hs_DistHS, highlighting descriptors related to global intensity, spatial dispersion, and entropy-based complexity. These results show a clear convergence at the level of descriptor families with those identified by ANOVA and LASSO, despite differences in the underlying selection mechanisms.
Although the exact variables retained may differ between datasets, they consistently belong to the same conceptual families of descriptors. This behavior is expected in high-dimensional and correlated feature spaces, where multiple variables often capture closely related aspects of the same underlying phenomena. Feature selection algorithms may therefore select different but functionally equivalent descriptors without affecting classification performance.
The final feature set was obtained by combining the outputs of the ANOVA, LASSO, and Random Forest selection procedures, while explicitly removing redundant variables identified through correlation analysis. This process resulted in a compact subset of ten non-redundant and informative variables: nbCorrelSup95, l1, varRatioSlopeMean, hs_DistHS, Biodiv, sumEntropyLog2Circle, hs_DistMean, nbStreamer, hs_ValHS, and meanStream.
It should be noted that, in correlated and high-dimensional settings, feature selection is not unique. Small variations in selection criteria or regularization strength may lead to different but functionally equivalent subsets of variables, without significantly affecting predictive performance. This behavior reflects the distributed and multivariate nature of the information underlying class discrimination in the present study.
Although the final subsets of selected variables differ between datasets, they exhibit strong conceptual similarities. In both cases, the retained descriptors belong to comparable families describing global intensity, spatial dispersion, heterogeneity, and entropy-related complexity. In high-dimensional and correlated settings, different but functionally equivalent variables may be selected without affecting predictive performance. A detailed comparative analysis of the variables selected by ANOVA, LASSO, and Random Forest, as well as their overlaps and complementarities, is provided in Appendix 1.
Classification Performance
Using the selected feature subset, several supervised classifiers were evaluated. Boosting-based models achieved the best performance, and XGBoost was retained as the final classifier after hyperparameter optimization.
Performance was assessed using stratified five-fold cross-validation. The optimized XGBoost model achieved a mean ROC–AUC of approximately 0.85 (Figure 4) with low variability across folds and an overall classification accuracy of 74%. As shown by the confusion matrix (Appendix 2), 412 out of 550 Montcalm samples and 452 out of 616 Montcalm Microwave samples were correctly classified.
Complementary metrics (Appendix 3) indicate balanced performance, with a precision of 0.77, a recall of 0.73, and an F1-score of 0.75, confirming the robustness and generalization capability of the model.
Interpretation of the Selected Variables
To improve the interpretability of the classification results, the retained image descriptors were analyzed from a phenomenological perspective rather than as a purely machine-learning-driven classification exercise. Although these variables do not provide direct access to molecular-scale properties of water, they capture complementary macroscopic characteristics of the corona discharge patterns generated around liquid droplets under fixed experimental conditions. All selected variables are explicitly defined and associated with their computational formulation and image-processing origin. Each descriptor is further linked to a specific morphological aspect of the corona discharge, such as global intensity distribution, spatial organization of luminous structures, heterogeneity of emission patterns, or entropy-related complexity across scales. This explicit linkage allows the reader to assess both the physical plausibility of the descriptors and their relevance to the observed discriminative behavior. Taken together, the retained variables describe distributed and multivariate features of the electrophotonic response rather than a single dominant effect. This supports an interpretation in which class discrimination arises from subtle but systematic differences in global discharge organization, consistent with the macroscopic and integrative nature of corona imaging. The mathematical formulas are given in Appendix 4.
Comparison with Mont Roucous Water
To further validate the robustness and generalizability of the proposed methodology, the same analysis pipeline was applied to an independent dataset consisting of commercially available Mont Roucous bottled water samples and Mont Roucous microwave samples. This additional experiment aims to assess whether the discriminative patterns identified in the previous analysis are reproducible across different water sources.
Variable Selection for Mont Roucous
For consistency and reproducibility, the same feature extraction and selection framework used for the Montcalm dataset was applied to the Mont Roucous dataset. A total of 5,813 observations were analyzed.
This multi-stage selection process resulted in a final subset of twelve non-redundant and informative variables:
deuxiemeDemiVie, richness, l2, h1, hs_DistHigh10, meanStream, sumEntropyLine, hs_DistMean, lineHSStdToCircleHS, stdLineMean, stdRatiosZones30.
The variables selected by ANOVA + LASSO are detailed in Appendix 5 and for Random Forest in Appendix 6.
Classification Results for “Mont Roucous” and “Mont Roucous Microwave”
Using the selected feature subset, classification was performed following the same modeling strategy as for the Montcalm dataset. In particular, the XGBoost classifier—an optimized implementation of gradient boosting—was retained and applied with the same hyperparameter configuration as previously determined for Montcalm. This choice was made to ensure methodological consistency and to enable a direct and fair comparison between the two datasets.
Model performance was assessed using stratified five-fold cross-validation. The resulting ROC curves (Figure 5) exhibit a mean ROC–AUC of approximately 0.86, with limited variability across folds, indicating good discriminative power and stable generalization behavior.
The model achieved an overall classification accuracy of 78%. Specifically, 455 out of 559 Mont Roucous samples and 454 out of 604 Mont Roucous microwave samples were correctly classified, as summarized in the confusion matrix (Appendix 7).
Additional performance metrics further confirm the robustness of the classifier (Appendix 8), with a precision of 0.81, a recall of 0.75, and an F1-score of 0.78. Performance remained well balanced between the two classes, indicating the absence of significant classification bias.
Interpretation of the Selected Variables
Although individual variables differ between datasets, they consistently describe the same families of structural and statistical properties, supporting a unified interpretation framework.
Representative Image Analysis and Combined Water Class Classification
To visualize the differences between the “Montcalm” and “Montcalm Microwave” classes, a principal component analysis (PCA) was applied to the selected feature set. The first two principal components were used as projection axes, and the centroid of each class was computed in this reduced space. The Euclidean distance between each observation and its corresponding centroid was calculated to identify the most representative samples. For each class, the two observations closest to the centroid were selected and displayed as representative images (Figures 6 and 7), providing a qualitative illustration of the characteristic patterns captured by the extracted descriptors.
For the combined “Water vs. Microwave” analysis, dimensionality reduction using PCA was also performed, and the decision boundary of the SVM classifier was projected onto the reduced space (Figure 9). The classifier achieved an overall accuracy of approximately 0.80, with a precision of 0.78, a recall of 0.85, and an F1-score of 0.81. The ROC curve (Figure 8) indicates good discriminative ability between the two classes. The confusion matrix shows that 41 water samples and 46 microwave-treated samples were correctly classified, with only a limited number of misclassifications, and a relatively symmetrical distribution of errors.
Overall, these results confirm that combining image-derived descriptors with supervised classification enables reliable discrimination between untreated and microwave-treated water samples, even when data from different commercial sources are pooled.
Comparative: Montcalm vs. Mont Roucous
The proposed methodology was applied to two independent commercial water datasets, Montcalm and Mont Roucous, in order to evaluate its reproducibility across distinct sources. While the two datasets differ in their intrinsic properties and in the specific variables retained after feature selection, similar levels of classification performance were observed.
For the Montcalm dataset, the XGBoost classifier achieved an accuracy of 74% with a mean ROC–AUC of approximately 0.85. When the same XGBoost model and identical hyperparameter configuration were applied to the Mont Roucous dataset, an accuracy of 78% and a mean ROC–AUC of approximately 0.86 were obtained. In both cases, stratified five-fold cross-validation yielded stable performance metrics, indicating limited sensitivity to data partitioning.
Although the final sets of selected variables differ between the two datasets, they belong to comparable families of descriptors. In both cases, the retained variables characterize global magnitude, spatial variability, structural dispersion, and entropy-related complexity. This observation indicates that similar types of information are exploited by the classifier, despite dataset-specific differences in the selected features.
Using the same classifier architecture and hyperparameter configuration for both datasets ensures methodological consistency and allows a direct comparison of performance metrics. The observed similarity in classification results across datasets confirms that the proposed pipeline behaves consistently when applied to independent water samples from different commercial sources.
Discussion
Limitations and Interpretative Framework
The present results demonstrate that macroscopic corona imaging combined with supervised learning enables reproducible discrimination between untreated and microwave-treated bottled water samples.
Importantly, discrimination does not rely on a single dominant descriptor but on multivariate patterns combining complementary features related to spatial organization, heterogeneity, and entropy. This supports the idea that classification reflects distributed changes in global behavior rather than isolated effects. However, several important limitations must be emphasized.
First, the approach does not provide direct access to the microscopic molecular structure of water. The extracted descriptors represent global and integrated electro-optical responses arising from the interaction between the liquid, the applied electric field, and the surrounding gas. Consequently, the observed differences should be interpreted as phenomenological signatures rather than as direct evidence of specific molecular rearrangements.
Second, although all samples were commercially available bottled waters compliant with current drinking-water regulations, a detailed chemical characterization beyond standard specifications was not performed. The possible presence of trace organic compounds, microplastics, or other minor constituents below regulatory thresholds cannot be fully excluded. To mitigate this limitation, experiments were conducted on multiple independent batches originating from several bottles, which makes it unlikely that a single uncontrolled contaminant could systematically account for the observed class of separation. In addition, the stability of corona discharge signatures under varying storage conditions, temperature changes, light exposure, or aging has not yet been investigated. Consequently, potential applications such as counterfeit detection or product authentication should be regarded as exploratory and restricted to controlled contexts.
Third, the present study does not aim to demonstrate permanent or long-lived non-thermal microwave effects in water. After microwave exposure, samples were allowed to relax at room temperature prior to imaging in order to minimize immediate thermal contributions. The objective is therefore not to investigate microwave effects per se, but to evaluate whether the proposed framework is sensitive to subtle differences associated with distinct electromagnetic histories.
With the current experimental protocol, acquisition and analysis times remain relatively long, which prevents real-time monitoring of short-term relaxation dynamics. The observed differences therefore reflect integrated effects rather than instantaneous molecular-scale kinetics. LIMEC is not intended to replace established spectroscopic techniques such as NMR, Raman, or THz spectroscopy, but may provide complementary information related to collective and interfacial photonic phenomena.
Finally, although the experimental setup relies on standard components (high-voltage generator, imaging optics, and camera) and requires minimal sample preparation, further engineering developments are needed to improve robustness, automation, and user-friendliness before routine industrial or commercial deployment.
Robustness, Reproducibility, and Methodological Scope
The robustness of the methodology is supported by consistent results obtained on two independent commercial bottled water brands, as well as on a combined dataset pooling different sources. Feature selection relies on complementary statistical and machine-learning-based approaches, and classification performance remains stable across cross-validation folds.
In addition, image acquisition and data analysis are fully automated once samples are prepared. Feature extraction and classification are performed algorithmically, without subjective human intervention. As a result, potential operator-related cognitive bias does not influence the numerical descriptors or the classification outcome.
At the present stage, the study was conducted using a single EDS© system in order to establish a controlled and reproducible proof-of-concept. Inter-device comparison studies involving multiple EDS© units or alternative corona imaging platforms have not yet been conducted. However, such comparisons are explicitly planned as part of future work. To facilitate reproducibility, several device parameters must be strictly controlled. These include applied voltage amplitude and stability; excitation frequency; electrode geometry and spacing; optical acquisition parameters; and environmental conditions such as temperature, hygrometry, and atmospheric pressure. In this manuscript, we now emphasize these parameters as critical for reproducible acquisition. We agree that standardized protocols will be essential for inter-laboratory replication. These would include fixed acquisition settings, reference environmental conditions, standardized image preprocessing pipelines, and unified feature extraction procedures.
Finally, the establishment of reference standards, such as certified aqueous samples with controlled and documented treatment histories, represents an important and realistic avenue for method validation. Such standards would enable both device calibration and cross-platform comparison. We believe that addressing these aspects will significantly strengthen the methodological maturity of macroscopic corona effect imaging and facilitate its adoption by the analytical chemistry community.
Practical Perspectives and Applications
The proposed framework provides inferred parameters reflecting the global physico-chemical state of a liquid through its electrophotonic response. It is not intended to replace conventional analytical chemistry techniques, but rather to complement them as a rapid screening or early-warning tool capable of detecting reproducible differences under controlled conditions. Potential near-term applications include high-frequency monitoring of process waters, detection of process drifts, and qualitative comparison of liquid products in laboratory or industrial settings. The mention of applications such as counterfeit detection in high-value beverages should therefore be understood as exploratory and restricted to controlled contexts. The long-term stability of the extracted signatures under varying storage conditions, temperature, or light exposure remains to be systematically evaluated.
At present, macroscopic corona effect imaging should be regarded as a laboratory-based and professional analytical framework rather than a consumer-oriented technology. Although the LIMEC framework is not inherently restricted to laboratory-scale implementations, significant technical challenges currently limit miniaturization. These include high-voltage generation and stabilization, UV-sensitive imaging requirements, controlled electrode geometry and discharge conditions, sensitivity to environmental noise (temperature, humidity, electromagnetic interference), and overall system stability.
Moreover, the present methodology relies on controlled acquisition protocols and trained classification models optimized for reproducibility and interpretability rather than instantaneous consumer feedback. Each application domain would therefore require dedicated training datasets and validation procedures. Significant simplifications would be necessary to enable portable or consumer-grade implementations, inevitably involving trade-offs in sensitivity, spatial resolution, and interpretative depth. While laboratory and industrial applications prioritize accuracy, calibration, and environmental control, hypothetical consumer applications would necessarily favor robustness and qualitative trend detection over precise quantitative interpretation.
Importantly, the present work does not aim to promote the notion of “live water structure analysis” for individual health assessment. Instead, LIMEC is positioned as a phenomenological and integrative approach for detecting reproducible differences between well-defined experimental conditions. Continued advances in high-voltage electronics, optical miniaturization, and embedded computation may eventually enable simplified versions of corona-based sensing systems; however, any such developments would require extensive validation, standardized reference materials, and careful consideration of ethical and interpretative issues before broader deployment could be envisaged.
Perspectives Under Extreme Thermodynamic Conditions
The present study was deliberately restricted to ambient temperature and pressure conditions in order to establish a robust and reproducible proof-of-concept. At this stage, we have not experimentally investigated LIMEC under extreme thermodynamic conditions such as high temperature, high pressure, or supercritical water. However, these regimes indeed represent scientifically and industrially relevant extensions of the approach.
Investigating such conditions would require substantial technical adaptations of the EDS© system. In particular, operation at elevated temperatures (>80 °C) or under pressurized conditions would necessitate the development of sealed and thermally resistant experimental cells, capable of withstanding both high voltage and controlled temperature–pressure environments. Optical access would need to be ensured through high-temperature, UV-transparent windows, while maintaining electrical insulation and discharge stability.
For very high temperatures (150–200 °C) and supercritical water conditions (>374 °C, >221 bar), the classical droplet-based corona geometry used in the present work would no longer be applicable. Alternative configurations, such as confined discharge geometries or indirect electrophotonic measurements through windowed pressure vessels, would need to be designed. Moreover, the physicochemical properties of water in these regimes differ fundamentally from those at ambient conditions, implying that both discharge behavior and feature interpretation would require careful re-evaluation.
High-pressure compressed water relevant to deep ocean or geological contexts would raise similar challenges, particularly with respect to pressure-resistant housings, electromagnetic compatibility, and safe high-voltage operation.
Overall, while extending LIMEC to extreme temperature and pressure regimes is technically non-trivial, it represents an interesting prospect for future research rather than an immediate application of the current setup. The present work therefore focuses on establishing methodological validity under controlled ambient conditions, which constitutes a necessary prerequisite before considering such advanced extensions.
Cost and Usability
From an operational perspective, the proposed system is based on standard components, including a high-voltage generator, an optical imaging assembly, and a camera. Consumable costs per measurement are low, as no reagents or complex sample preparation are required. Sample handling is minimal, and image acquisition can be automated, which reduces operator intervention and facilitates repeatability.
While full industrial deployment will require further engineering developments to improve robustness, scalability, and long-term stability, these characteristics suggest a reasonable cost structure and practical potential for routine measurements. Exact commercial pricing depends on system configuration and manufacturer and is therefore not provided.
Future Developments
Ongoing developments toward automation and robotization of the acquisition system are expected to substantially increase throughput while reducing operator-related variability. In parallel, future studies will explore the use of closed and pressurized experimental cells, enabling investigations under non-ambient conditions. In this context, temperature and pressure would be introduced as controlled contextual variables, while maintaining a consistent feature extraction and classification framework.
Although the present study focuses on stabilized and time-integrated signatures, macroscopic corona effect imaging may potentially be extended toward temporal investigations of water dynamics. With adapted acquisition protocols and faster imaging systems, it could become possible to monitor the evolution of corona discharge patterns following controlled electromagnetic perturbations and to investigate relaxation kinetics as water returns toward baseline states.
Importantly, LIMEC should not be regarded as a substitute for established spectroscopic techniques such as NMR relaxometry, terahertz spectroscopy, or Raman spectroscopy. Rather, it provides a complementary macroscopic perspective, capturing collective electrophotonic responses arising from the interaction between the liquid, the applied electric field, and the surrounding gas.
Finally, future work may examine whether different types of electromagnetic perturbations (e.g., microwave, radiofrequency, or static magnetic fields) give rise to distinguishable corona signatures. Such investigations would extend the scope of LIMEC beyond quality monitoring toward exploratory studies of water behavior under electromagnetic fields, while remaining within a phenomenological and integrative framework.
Exploratory Extensions
At this time, the LIMEC + Machine Learning framework has not been applied to human plasma or serum samples. The current study is intentionally restricted to well-defined aqueous systems in order to establish a controlled and interpretable proof-of-concept. Extending the approach to biological fluids would require substantial methodological adaptations and careful consideration of additional sources of variability.
In contrast to bottled water, blood plasma contains high concentrations of proteins, lipids, electrolytes, and other biomolecules, as well as potential cellular components depending on sample preparation. These constituents strongly influence electrical conductivity, surface tension, viscosity, and interfacial behavior, all of which are expected to affect corona discharge patterns. As a result, additional descriptors and preprocessing steps would be required to account for protein content, macromolecular aggregation, and sample heterogeneity. While it is conceivable that macroscopic corona signatures could differ between physiological or pathological states, any such application would require extensive validation against established clinical markers, rigorous control of confounding factors (e.g., sample handling, anticoagulants, storage time), and large, well-characterized cohorts. Importantly, the present work does not make any claims regarding medical diagnostics, disease detection, or health assessment.
Moreover, translation toward medical applications would raise substantial ethical and regulatory considerations, including patient consent, data protection, and compliance with medical device regulations. Addressing these aspects would require dedicated clinical studies conducted under appropriate ethical approvals and regulatory frameworks. For these reasons, we view potential biomedical applications as a long-term and exploratory perspective rather than a near-term objective.
Conclusion
In this study, we employed the EDS© system to acquire ultraviolet corona discharge images around water droplets and to extract quantitative descriptors used as input variables for supervised classification models.
The results demonstrate that macroscopic corona effect imaging, combined with multivariate feature extraction and supervised learning, enables reproducible discrimination between untreated bottled water samples and the same water subjected to short microwave exposure under controlled experimental conditions.
Importantly, this work does not claim to demonstrate permanent or well-defined microscopic structural modifications of water. The extracted descriptors reflect global and integrated electro-optical responses arising from liquid–field–gas interactions, and the observed differences should therefore be interpreted as phenomenological macroscopic signatures rather than as direct evidence of specific molecular rearrangements.
The main contribution of this study is to establish the feasibility of a sensitive, low-sample-volume, and non-destructive imaging-based framework capable of detecting subtle and systematic differences in water behavior associated with distinct electromagnetic histories.
Such an approach may complement conventional analytical techniques as a rapid screening or early-warning tool for monitoring process waters, assessing the consistency of liquid products, or detecting deviations in closed industrial circuits. Future developments will focus on improving automation, throughput, and inter-laboratory reproducibility, as well as on extending the methodology to other classes of aqueous systems.
References
Andersson MP (2023). The shape of water – How cluster formation provides a unifying explanation of water’s anomalous properties. Journal of Molecular Liquids 383, 122169.
Arani R, Bono I, Giudice E, Preparata G (1995). QED coherence and the thermodynamics of water. International Journal of Modern Physics B 9(15), 1813–1841.
Bishop CM (2006). Pattern recognition and machine learning. Springer.
Bohr H, Bohr J (2000). Microwave-enhanced folding and denaturation of globular proteins. Physical Review E 61(4), 4310–4314.
Bono I, Del Giudice E, Gamberale L, Henry M (2012). Emergence of the coherent structure of liquid water. Water 4(3), 510–532.
Chaplin M (2006). Do we underestimate the importance of water in cell biology? Nature Reviews Molecular Cell Biology 7, 861–866.
Chen J, Davidson JH (2002). Electron density and energy distributions in the positive DC corona: Interpretation for corona-enhanced chemical reactions. Plasma Chemistry and Plasma Processing 22(2), 199–224.
Coudert FX (2007). L’eau et l’électron hydraté en milieu confiné (Doctoral dissertation, Université Paris-Sud).
De Grotthuss CJT (1806). Sur la décomposition de l’eau à l’aide de l’électricité galvanique. Annales de Chimie 58, 54.
Einstein A (1905a). Zur Elektrodynamik bewegter Körper. Annalen der Physik 322(10), 891–921.
Einstein A (1905b). Ist die Trägheit eines Körpers von seinem Energieinhalt abhängig? Annalen der Physik 323(13), 639–641.
Ellison WJ, Lamkaouchi K, Moreau JM (1996). Water: A dielectric reference. Journal of Molecular Liquids 68(2–3), 171–279.
Foster KR, Glaser R (2007). Thermal mechanisms of interaction of radiofrequency energy with biological systems with relevance to exposure guidelines. Health Physics 92(6), 609–620.
Fresnel A (1826). Mémoire sur la diffraction de la lumière. Mémoires de l’Académie des Sciences 5, 339–475.
Fridman A (2008). Plasma chemistry. Cambridge University Press.
Fridman G, Friedman G, Gutsol A, Shekhter AB, Vasilets VN, Fridman A (2008). Applied plasma medicine. Plasma Processes and Polymers 5(6), 503–533.
Gabriel C, Gabriel S, Grant EH, Halstead BSJ, Mingos DMP (1998). Dielectric parameters relevant to microwave dielectric heating. Chemical Society Reviews 27(3), 213–224.
Goldman M, Goldman A, Sigmond RS (1985). The corona discharge, its properties and specific uses. Pure and Applied Chemistry 57(9), 1353–1362.
Guyon I, Elisseeff A (2003). An introduction to variable and feature selection. Journal of Machine Learning Research 3, 1157–1182.
Henry M (2016). L’eau et la physique quantique. Dangles.
Janadri CS, Sheeparamatti BG, Kagawade V (2017). Multiclassification of Kirlian images using SVM technique. Proceedings of ICACCI, 2246–2250.
Kaatze U (1989). Complex permittivity of water as a function of frequency and temperature. Journal of Chemical & Engineering Data 34(4), 371–374.
Korotkov K (2013). Energy fields electrophotonic analysis in humans and nature. eBookIt.
Laroussi M (2009). Low-temperature plasmas for medicine? IEEE Transactions on Plasma Science 37(6), 714–725.
Loeb LB (1965). Electrical coronas: Their basic physical mechanisms. University of California Press.
Malus EL (1811). Mémoire sur de nouveaux phénomènes d’optique. Journal de Physique 72, 393–398.
Maxwell JC (1873). A Treatise on Electricity and Magnetism. Clarendon Press, Oxford.
Meredith RJ (1998). Engineers’ handbook of industrial microwave heating. Institution of Engineering and Technology.
Metaxas AC, Meredith RJ (1983). Industrial microwave heating. Peter Peregrinus Ltd.
Newton I (1952). Optics. Dover Publications.
Pakhomov AG, Murphy MR (2000). A comprehensive review of the research on biological effects of pulsed radiofrequency radiation in Russia and the former Soviet Union. In: Lin JC (ed.), Advances in Electromagnetic Fields in Living Systems, vol. 3, 265–290. Springer.
Peek FW (1929). Dielectric phenomena in high voltage engineering. McGraw-Hill.
Perakis F, Amann-Winkel K, Lehmkühler F, Sprung M, Mariedahl D, Sellberg JA, Nilsson A (2017). Diffusive dynamics during the high-to-low density transition in amorphous ice. Proceedings of the National Academy of Sciences 114(31), 8193–8198.
Planck M (1900). On the theory of the energy distribution law of the normal spectrum. Verhandlungen der Deutschen Physikalischen Gesellschaft 2, 237–245.
Pollack GH (2013). The fourth phase of water. Ebner & Sons, Chapter 3
Raizer YP (1991). Gas discharge physics. Springer.
Raveloson F, Roussel J, Vandanjon L, de la Bardonnie H (2024). Multidimensional statistical patterns observed on water drops corona discharge pictures. Water 13.
Soper AK (2013). The radial distribution functions of water as derived from radiation total scattering experiments: Is there anything we can say for sure? ISRN Physical Chemistry 2013, Article 279463.
Sun Q (2009). Local statistical interpretation for water structure. Chemical Physics Letters 480, 1–8.
Vandanjon L (2021). L’interprétation quantique du rôle de l’eau dans le vivant. Matrix. ISSN 2266-9949.
Voeikov V, Korotkov KG (2017). The emerging science of water. CreateSpace.
Watterson JG (1987). A role for water in cell structure. Biochemical Journal 248(2), 615–617.
Wiggins P (2008). Life depends upon two kinds of water. PLoS ONE 3(1), e1406.
Young T (1802). On the theory of light and colours. Philosophical Transactions of the Royal Society of London 92, 12–48.
Appendices
Discussion with Reviewers
Reviewer 1: Add more detail on environmental electromagnetic shielding during preparation and experimentation.
Authors: Graduated tubes containing the solutions were placed in an aluminum enclosure to limit their exposure to electromagnetic waves during the preparation and experimental phases. In addition, excessive exposure to light was avoided.
Reviewer 1: Discuss plans for inter-device/inter-laboratory validation (availability, experimental possibilities).
Authors: For now, there are three prototypes of this EDS technology, which have enabled us to validate our model. Each comparative experiment between two samples requires approximately 8 hours of manual work. This effort is necessary to generate datasets of sufficient size (around 300 acquisitions and several thousand images when considering each frame).
To facilitate experimentation as well as inter-device and inter-laboratory comparisons, we plan to automate (robotize) technology. This would offer several advantages, including reducing human influence and improving control over atmospheric parameters (pressure, temperature, hygrometry, electromagnetic interference, light conditions, etc.).
There are other devices capable of producing corona discharge images from water droplets, particularly the Bio-Well system. However, the parameters currently used (voltage, frequency, and captured wavelengths) do not allow for direct comparison. To our knowledge, other available devices also do not enable the acquisition of large volumes of images.
Reviewer 2: Is there any particular reason for using “microwave” treatment?
Authors: Microwave treatment was selected as a simple and reproducible perturbation applied to chemically identical water samples. To our knowledge, no standard analytical method currently allows rapid, non-destructive discrimination between water samples solely based on prior microwave exposure.
The objective was not to investigate microwave effects per se, but to assess the sensitivity of the proposed approach to subtle differences induced by distinct electromagnetic histories.
Reviewer 2: How easy is it to perform this measurement if it is proposed for commercial use?
Authors: In its current laboratory configuration, approximately two operators are required to perform 150–200 measurements in two hours. Ongoing developments toward full automation are expected to substantially increase throughput and reduce human intervention, making routine commercial use feasible.
Reviewer 3: Why do the extracted variables differ between the two waters?
Authors: The extracted variables are not strictly identical because the samples under investigation are not identical systems. Although they originate from the same commercial water source, one subset has undergone microwave exposure, which constitutes a distinct physical pathway. Our objective was not to impose an identical variable set a priori, but rather to evaluate the intrinsic discriminative capacity of the LIMEC–machine learning coupling.
To ensure methodological coherence between the two studies, we therefore adopted a comparative strategy based on grouping the microwave-treated samples (from both water brands) and contrasting them with the non-treated samples. This approach allows us to assess whether samples exposed to microwave radiation share common statistical patterns that differentiate them from untreated waters, independently of brand-specific characteristics.
This strategy emphasizes the robustness of the classification framework and its ability to detect subtle but consistent differences induced by electromagnetic exposure. In ongoing and future work, we plan to extend this approach to additional water types. One long-term objective is to investigate whether such analyses may reveal reproducible features that could be interpreted as a specific microwave-related signature in water, while remaining cautious about over-interpretation at the current stage.
Reviewer 3: Would it not have been preferable to have a double-blind set up to avoid potential operator biases?
Authors: To minimize potential operator-related biases, the experimental procedure was designed to limit the influence of individual intention or expectation, even though a formal double-blind protocol was not implemented at this stage. In practice, the workflow naturally included regular alternation of personnel across tasks, without explicit instruction or emphasis, as well as rotation between different working positions (e.g., standing and seated tasks) over approximately four-hour periods. In addition, the handling and processing of sample batches were alternated throughout the experiments.
Within our working hypothesis, the parameter space extracted from corona discharge images is high-dimensional. If operator-related effects (such as mood, attention, or intention) were to measurably influence the acquired signals, we assume that such effects could, in principle, be identified and separated statistically from microwave-induced effects within the same experimental series. In other words, the richness of the parameter set may allow the coexistence and differentiation of multiples sources of variability, including both sample-related and operator-related contributions.
Nevertheless, we fully agree that future experiments should incorporate stricter controls. In particular, the use of robotic or fully automated acquisition systems appears essential. Automation will not only allow a substantial increase in the number of acquisitions, samples, and batches, but will also provide a stable reference condition free from human intervention. Such a reference is necessary to quantitatively assess and, if needed, isolate any operator-related influence.