Terahertz metamaterial liquid detector optimized by deep learning Authors: Xuyan Li, Zhibo Wu, Yujie Zhang, Bin Li Published: 2025 ABSTRACT This study proposes a label-free ethanol liquid detection method based on the absorption peak shift of a metamaterial terahertz detector. The designed metamaterial liquid detector adopts a five-layer structure comprising interdigitated copper metal wires, a polyimide dielectric layer, nested double-letter-shaped VO2 open-loop square rings, and a copper metal plate. Structural parameters were optimized using a deep neural network (DNN), achieving high absorption rates and displacement values under various operating conditions. The detector exhibits a sensitivity of 3.33 THz/RIU and a figure of merit (FOM) of 66 RIU⁻¹. CST simulations reveal that the absorption peak shifts toward the infrared end as droplet volume increases, toward the blue light end as ethanol concentration rises, and the peak spacing narrows under elevated temperatures. Analysis of magnetic field strength and surface current reveals the operating mechanism: At room temperature, VO₂ plays a negligible role in absorption wave regulation, but its enhanced conductivity at elevated temperatures provides dual assurance for ethanol detection accuracy. This study not only offers an efficient and precise new method for ethanol liquid detection but also provides a reference for future applications in food safety, environmental monitoring, and biomedicine. Introduction Metamaterials are a class of artificial electromagnetic materials with unique physical properties1. Since Veselago proposed the relevant theory and predicted special electromagnetic phenomena in the mid-20th century2, research on metamaterials has continued to this day and is still evolving. Today, metamaterials have evolved into various types, including left-handed materials3, double-negative (DNG) materials4, and materials used for light field manipulation5 and functional device construction6. The unique electromagnetic properties of metamaterials stem from the interaction between their structure and incident electromagnetic waves7, and this electromagnetic response is precisely controlled by the geometric parameters of the structure. In recent years, metamaterials have garnered significant attention due to their ability to generate electromagnetic induction reflection (EIR) phenomena8. Their high sensitivity to environmental changes endows them with immense application potential in the fields of biological9,10 and chemical sensing11,12. Concurrently, the rapid advancement of terahertz technology has positioned terahertz waves as a highly competitive candidate technology in the field of sensing and detection13. Picosecond-level terahertz pulse widths can effectively suppress far-infrared background noise, and their low-energy characteristics avoid damaging the detected materials14. Additionally, terahertz waves exhibit correlations with biological molecular vibrations, giving terahertz technology a unique advantage in biological molecular vibration analysis15, particularly in detecting liquid components such as ethanol16. Traditional sensors face numerous drawbacks when detecting ethanol and other liquids, such as large sample requirements, limited sensitivity, and cumbersome procedures involving the addition of detection reagents17. In contrast, metamaterial biosensors, as an emerging label-free sensing method, can cleverly convert changes in external refractive index into changes in optical signals, potentially breaking through the resolution bottleneck of traditional sensors. However, some metamaterial biosensors still face issues such as low sensitivity of SRR rings in ultra-thin silicon nitride substrate biosensors, insufficient sensing accuracy due to mismatches between the measurement environment and actual biological systems, and difficulties in simultaneously meeting sensitivity and FOM value requirements18. Against this backdrop, this study focuses on a label-free detection method for ethanol liquids using the absorption peak shift of metamaterial terahertz detectors, aiming to explore an efficient, precise, and convenient new approach for ethanol liquid detection, and to provide innovative ideas and technical support for the field of liquid detection sensing. Research indicates that the absorption peak position of metamaterial terahertz detectors is highly sensitive to the refractive index of the surrounding environment. When the concentration or volume of ethanol liquid changes, its refractive index also changes, leading to a shift in the absorption peak position of the metamaterial terahertz detector. By precisely measuring and analyzing this shift, label-free detection of ethanol liquid concentration can be achieved. Design and method The schematic diagram illustrating the performance of the metamaterial terahertz liquid detector designed in this paper is shown in Fig. 1. As can be seen from the figure, as the volume of the liquid droplet gradually increases, the peak of the liquid detector undergoes a significant red shift under terahertz wave irradiation. CST 2023 was employed for model simulation using a regular tetrahedral mesh and a frequency-domain solver. Concurrently, the X and Y direction boundaries were defined as “unit cell,” while the Z direction boundary was set as “open.” Its specific structure is shown in Figs. 2a, c, d), and e. It consists of five layers: the top layer is a finger-shaped copper metal wire (conductivity σ = 5.96 × 10⁷ S/m); the dielectric layers between the two metal layers are made of polyimide (PI, relative permittivity 3.4, dielectric loss tangent 0.06)19; the middle layer consists of two cross-shaped VO₂ structures nested within an open square ring; and the bottom layer is a copper metal plate. The structural parameters are listed in Table 1: Fig. 1Rendering. Rendering. Table 1Detailed parameter table.Structure parameter (µm)PRdadbdcdedf100454040303576dghCu1hPI1hVO2hCu2hPI2hCu3120.4120.40.4120.4 Table 1Detailed parameter table.Structure parameter (µm)PRdadbdcdedf100454040303576dghCu1hPI1hVO2hCu2hPI2hCu3120.4120.40.4120.4 Detailed parameter table. To address the challenge of optimizing multiple structural parameters and target performance in terahertz (THz) devices19, this study introduces a fully data-driven deep learning framework20,21. This framework employs a deep neural network (DNN)22,23 for the inverse design and performance prediction of terahertz metamaterial liquid detectors. Its customized architecture is specifically designed for the regression task of mapping high-dimensional structural parameters to electromagnetic absorption responses24,25. The DNN comprises three fully connected hidden layers with 64, 48, and 24 neurons respectively. Each layer employs the rectified linear unit (ReLU) activation function to introduce nonlinearity while mitigating gradient vanishing. To enhance generalization and prevent overfitting, L2 regularization (weight decay = 1e-4) and dropout (probability = 0.2) are applied to all hidden layers. The input layer is designed to receive 14 geometric and material parameters, including periodicity, resonator dimensions, and layer thickness. All parameters are normalized using Z-scores to ensure a mean of zero and variance of one. The output layer predicts the absorption rate at a specific terahertz frequency, employing single-node regression with a linear activation function. The training dataset was generated using CST Microwave Studio, yielding 16,489 simulated samples covering a wide range of structural configurations and their corresponding absorption efficiencies. To enhance model robustness and convergence stability, the dataset was augmented with Gaussian noise (σ = 0.01) to simulate manufacturing tolerances and measurement uncertainties. The dataset was stratified into 70% training, 20% validation, and 10% test sets, ensuring consistent distribution across all absorption rate intervals. The model was trained using the Adam optimizer with an initial learning rate of 1e-3, employing a decay strategy: when the validation loss plateaued for 10 consecutive epochs, the learning rate was reduced by 50%. A hybrid loss function combining Mean Squared Error (MSE) and Mean Absolute Error (MAE) was adopted to balance sensitivity to outliers and overall accuracy. Training spanned 300 epochs with early stopping based on validation loss convergence. The final model achieved MAE of 0.001, RMSE of 0.002, and a coefficient of determination (R²) exceeding 0.98 on the test set, demonstrating exceptional prediction fidelity. To further validate model reliability, sensitivity analysis was conducted using the Sobol index. Results revealed that the core parameters influencing absorption efficiency are the VO₂ layer thickness and outer ring radius, a conclusion highly consistent with electromagnetic theory and resonance behavior. This interpretable analysis not only verifies the model’s physical consistency but also provides actionable insights for future device optimization. Fig. 2Parameters of each part (a) Structural diagram (b) Deep learning model diagram (c) Interdigitated layer diagram (d) VO2 layer diagram (e) Layer thickness diagram. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SP_{S} = \left[ \begin{gathered} \begin{array}{*{20}c} P & R & {\begin{array}{*{20}c} {d_{a} } & {d_{b} } & {\begin{array}{*{20}c} {d_{c} } & {d_{e} } & {\begin{array}{*{20}c} {d_{f} } & {d_{g} } & {\begin{array}{*{20}c} {h_{{Cu1}} } & {\begin{array}{*{20}c} {h_{{PI1}} } & {h_{{VO2}} } & {\begin{array}{*{20}c} {h_{{Cu2}} } & {h_{{PI2}} } & {h_{{Cu3}} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} \hfill \\ \begin{array}{*{20}c} {\begin{array}{*{20}c} {P^{1} } & {R^{1} } & {\begin{array}{*{20}c} {d_{a}^{1} } & {d_{b}^{1} } & {\begin{array}{*{20}c} {d_{c}^{1} } & {d_{e}^{1} } & {\begin{array}{*{20}c} {d_{f}^{1} } & {d_{g}^{1} } & {\begin{array}{*{20}c} {h_{{Cu1}}^{1} } & {h_{{PI1}}^{1} } & {\begin{array}{*{20}c} {h_{{VO2}}^{1} } & {h_{{Cu}}^{1} } & {\begin{array}{*{20}c} {h_{{Cu2}}^{1} } & {h_{{Cu3}}^{1} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \begin{gathered} \begin{array}{*{20}c} {p^{{n - 1}} } & {R^{{n - 1}} } & {\begin{array}{*{20}c} {d_{a}^{{n - 1}} } & {d_{b}^{{n - 1}} } & {\begin{array}{*{20}c} {d_{c}^{{n - 1}} } & {d_{e}^{{n - 1}} } & {\begin{array}{*{20}c} {d_{f}^{{n - 1}} } & {f_{g}^{{n - 1}} } & {\begin{array}{*{20}c} {h_{{Cu1}}^{{n - 1}} } & {h_{{PI1}}^{{n - 1}} } & {\begin{array}{*{20}c} {h_{{VO2}}^{{n - 1}} } & {h_{{Cu2}}^{{n - 1}} } & {\begin{array}{*{20}c} {h_{{PI2}}^{{n - 1}} } & {h_{{Cu3}}^{{n - 1}} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} \hfill \\ \begin{array}{*{20}c} {P^{n} } & {R^{n} } & {\begin{array}{*{20}c} {d_{a}^{n} } & {d_{b}^{n} } & {\begin{array}{*{20}c} {\begin{array}{*{20}c} {d_{c}^{n} } & {d_{e}^{n} } \\ \end{array} } & {d_{f}^{n} } & {\begin{array}{*{20}c} {d_{g}^{n} } & {h_{{Cu1}}^{n} } & {\begin{array}{*{20}c} {h_{{PI1}}^{n} } & {h_{{VO2}}^{n} } & {\begin{array}{*{20}c} {h_{{Cu2}}^{n} } & {\begin{array}{*{20}c} {h_{{PI2}}^{n} } & {h_{{Cu3}}^{n} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} } \\ \end{array} \hfill \\ \end{gathered} \\ \end{array} \hfill \\ \end{gathered} \right]$$\end{document} Parameters of each part (a) Structural diagram (b) Deep learning model diagram (c) Interdigitated layer diagram (d) VO2 layer diagram (e) Layer thickness diagram. For ease of analysis, three datasets with structurally optimal parameters were manually selected from the predicted output data, with the structural parameters detailed in Table 2. In these three sets of predicted results, the detectors achieved high absorption rates and high shifts, thereby effectively validating the accuracy of the proposed deep neural network (DNN) model. It is worth noting that the prediction errors of Predicted_1 and Predicted_2 are slightly higher than those of Predicted_3. After comprehensively comparing performance and absorption peak shift, Predicted_3 was ultimately selected as the final choice, with its performance characteristics shown in Fig. 3. Table 2Parameter set.ModelNumberStructure parameter (µm)PRdadbdcdedfdghCu1hPI1hVO2hCu2hPI2hCu3DNNPredicted_1100403840353070100.4120.40.4120.4Predicted_210042403832336880.4140.40.4140.4Predicted_3100454040303576120.4120.40.4120.4 Table 2Parameter set.ModelNumberStructure parameter (µm)PRdadbdcdedfdghCu1hPI1hVO2hCu2hPI2hCu3DNNPredicted_1100403840353070100.4120.40.4120.4Predicted_210042403832336880.4140.40.4140.4Predicted_3100454040303576120.4120.40.4120.4 Parameter set. The metamaterial terahertz liquid detector designed in this paper complies with transmission line theory26. The fully reflective backplane of the detector can be equivalent to a segmented line input impedance Zin obtained by connecting the super surface layer impedance and the medium impedance in parallel. The complex equivalent impedance Zeff of the super surface layer can be expressed as27:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{eff}=Re\left({Z}_{eff}\right)+Im\left({Z}_{eff}\right)=M+jX$$\end{document} In Eq. (1): M is the real part of the equivalent impedance; X is the imaginary part of the equivalent impedance; j is the imaginary unit. The relationship between the input impedance Zin and the real part M and imaginary part X of the metamaterial equivalent impedance Zeff is given by28:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{in}=\frac{j(M-X){Z}_{0}\mathrm{s}\mathrm{i}\mathrm{n}\left(\beta\:d\right)}{\sqrt{{\epsilon\:}_{r}}Mcos\left(\beta\:d\right)+j\left[\sqrt{{\epsilon\:}_{r}}Mcos\left(\beta\:d\right)+{Z}_{0}\mathrm{s}\mathrm{i}\mathrm{n}\left(\beta\:d\right)\right]}$$\end{document} In Eq. (2): Z0 = 377 Ω, which is the characteristic impedance of free space; εr is the relative permittivity of the dielectric layer; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:\:=\:2\pi\:f\:\sqrt{{\epsilon\:}_{r}}\:/c$$\end{document}, For the propagation constant of the incident electromagnetic wave (f is the frequency of the incident electromagnetic wave, c is the speed of light in a vacuum); d is the thickness of the medium layer. The absorption rate A of the metamaterial terahertz liquid detector can be expressed as the relationship between S11 and S21 as shown in29:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:A=1-{\left|{S}_{11}\right|}^{2}-{\left|{S}_{21}\right|}^{2}\approx\:1-{\left|{S}_{11}\right|}^{2}$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{11}=\frac{-\sqrt{{\epsilon\:}_{r}}\left[Mcos\left(\beta\:d\right)+Xsin\left(\beta\:d\right)\right]+j[\left(M-{Z}_{0}\right)\mathrm{sin}\left(\beta\:d\right)-\sqrt{{\epsilon\:}_{r}}Xcos(\beta\:d\left)\right]}{\sqrt{{\epsilon\:}_{r}}\left[Mcos\left(\beta\:d\right)-Xsin\left(\beta\:d\right)\right]+j[\left(M+{Z}_{0}\right)\mathrm{sin}\left(\beta\:d\right)-\sqrt{{\epsilon\:}_{r}}Xcos(\beta\:d\left)\right]}$$\end{document} From Eqs. (5) and (6), it can be seen that as Zin approaches Z0, the reflection coefficient S11 decreases, and the absorption rate A of the liquid detector increases. When Zin = Z0, S11 = 0, and the absorption rate A of the liquid detector reaches its maximum value of 1. Next, we introduce important parameters such as sensitivity and quality factor to evaluate sensor performance30,31:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S=\frac{\varDelta\:f}{\varDelta\:n}$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:FOM=\frac{S}{FWHM}$$\end{document} Sensitivity (S) is defined as the displacement rate of the resonance frequency with respect to changes in the analyte’s refractive index, while the full width at half maximum (FWHM) characterizes the sharpness of the absorption peak32–34. The figure of merit (FOM) serves as a comprehensive evaluation metric, simultaneously considering both the sensitivity characteristics and spectral line shape of the sensing unit. A higher FOM value indicates that the device exhibits lower noise levels and superior precision when detecting minute refractive index changes35,36. The metamaterial terahertz detector consists of periodically arranged metamaterial structures and corresponding substrates. Its unique electromagnetic properties enable it to exhibit excellent absorption performance in the terahertz band. The design and optimization of the metamaterial structure are crucial for achieving efficient ethanol liquid detection. In this study, by carefully designing the geometric shape, dimensions, and arrangement of the metamaterial, the detector’s response to changes in the refractive index of ethanol liquid can be effectively enhanced, thereby improving detection sensitivity and accuracy. Additionally, this study also thoroughly explores the feasibility of the metamaterial terahertz liquid detector in practical applications and the feasibility of the manufacturing process. In summary, the method of using the absorption peak shift of the metamaterial terahertz detector for label-free detection of ethanol liquids not only provides a new technical approach for liquid detection but also holds broad application prospects in fields such as food safety37,38, environmental monitoring39,40, and biomedicine41,42. Through further research and optimized design of metamaterial terahertz detectors, it is anticipated that their detection performance can be enhanced, thereby advancing the practical application of terahertz technology in liquid detection. Results and discussion The following results were obtained through CST simulation. Figure 3 shows the simulated effects of droplet size and ethanol concentration on the absorption peak shift of the metamaterial terahertz liquid sensor. As depicted in Fig. 3a and b, the absorption peak shifts significantly toward the red end of the spectrum as droplet volume increases, regardless of whether at room temperature or elevated temperatures (The redshift mentioned in the subsequent text refers to the shift of absorption peaks toward lower frequencies, while blueshift is the opposite). This result demonstrates that the proposed detector can effectively detect minute liquid droplets. Additionally, Figs. 3c and d show that as the ethanol volume fraction in the ethanol solution increases, the absorption peak exhibits a noticeable blue shift. However, at room temperature, when the ethanol volume fraction reaches 60%, red shift interference occurs, thereby increasing the complexity of the detection process. Notably, a secondary absorption peak emerges under these conditions, which is only present in ethanol solutions with a volume fraction exceeding 60% and is not observed in other ethanol solutions with different volume fractions but similar absorption rates. This unique feature enables the metamaterial terahertz liquid sensor designed in this study to achieve precise liquid detection. Additionally, Fig. 3e shows that the performance of this liquid detector exhibits a specific pattern with temperature changes: as temperature increases, the distance between the two absorption peaks gradually decreases and shifts toward the intermediate position. Fig. 3Performance characteristic parameter results (a,b) Droplet size detection results (c,d) Ethanol content detection (e) Temperature rise detector performance change. Performance characteristic parameter results (a,b) Droplet size detection results (c,d) Ethanol content detection (e) Temperature rise detector performance change. Next, the operating principle of the terahertz metamaterial liquid detector will be elucidated through parameter inversion, magnetic field strength, and surface current perspectives, as illustrated in Figs. 4 and 5, and 6. At room temperature, magnetic resonance phenomena occur at the 0.48 and 0.76 THz frequency bands43,44, while electromagnetic resonance45,46 and impedance matching47,48 effects are simultaneously present. As the temperature rises, similar phenomena occur at 0.57 and 0.74 THz, thereby enabling the detector’s functionality. Fig. 4(a) Equivalent dielectric constant and magnetic permeability at room temperature (b) Equivalent dielectric constant and magnetic permeability after temperature increase (c) Equivalent impedance at room temperature (d) Impedance matching after temperature increase. (a) Equivalent dielectric constant and magnetic permeability at room temperature (b) Equivalent dielectric constant and magnetic permeability after temperature increase (c) Equivalent impedance at room temperature (d) Impedance matching after temperature increase. Fig. 5Magnetic field strength and surface current of each layer of the terahertz metamaterial liquid detector at room temperature. Magnetic field strength and surface current of each layer of the terahertz metamaterial liquid detector at room temperature. Fig. 6Magnetic field strength and surface current diagrams of each layer of the metamaterial terahertz metamaterial liquid detector after temperature increase. Magnetic field strength and surface current diagrams of each layer of the metamaterial terahertz metamaterial liquid detector after temperature increase. Next, this paper presents detailed observations of the magnetic field intensity and equivalent current patterns of the metamaterial sensor at room temperature and after heating. Figures 5a and g show the magnetic field strength distribution of the metamaterial liquid detector’s interdigitated layer at a frequency of 0.48 THz under room temperature conditions. It can be clearly seen that the regions with stronger magnetic field strength are primarily concentrated near the center of the interdigitated structure. Figures 6a and g illustrate the magnetic field strength distribution characteristics of the interdigitated layer of the sensor at a frequency of 0.74 THz. At this frequency, the regions with stronger magnetic field strength exhibit a pattern of diffusion from the center toward the periphery compared to the magnetic field strength regions at 0.57 THz. By comparing Figs. 5 and 6 and a significant difference in surface currents on the interdigital layer can be observed. In regions with relatively high current, the direction of current flow is almost entirely reversed. Comparing Figs. 5d, g and 6d, g, it can be seen that after the temperature rises, the magnetic field strength distribution of the interdigitated layer is basically consistent with the magnitude and direction of the surface current, while the field strength distribution of the intermediate phase-change resonant layer undergoes significant changes, with the field strength shifting from the outer ring toward the center, and the areas of higher field strength notably expanding. Meanwhile, the magnitude of the surface current significantly increases, but its direction remains largely unchanged. This phenomenon indicates that at room temperature, VO₂ contributes little to absorptive control; however, after temperature increases, due to enhanced conductivity, VO2 significantly enhances its contribution to absorptive efficiency. After an induced current is generated on the surface of the terahertz detector, according to the power loss equation Ploss = I2R34 (where Ploss represents electromagnetic power loss, I represents surface current, and R represents surface resistance), the induced current on the detector surface dissipates electromagnetic waves in the form of ohmic loss. Subsequently, this paper discusses the fabrication process of terahertz metamaterial liquid detectors, as shown in Fig. 7. First, as depicted in Figure (a), the silicon substrate was sequentially cleaned with acetone, anhydrous ethanol, and deionized water to remove contaminants. Subsequently, it was heated to 120 °C using a constant-temperature heater and baked for 15 min to thoroughly evaporate surface moisture. Next, a polyimide solution was spin-coated onto the silicon substrate using a spin coater. After baking for curing, a polyimide film was formed. Subsequently, photoresist was spin-coated onto the surface of the polyimide film using the spin coater again. Following this, a pre-bake was performed to completely decompose and evaporate the solvents within the photoresist. The photoresist-coated sample is placed in a lithography machine, exposed through a mask, and immersed in developer solution. This dissolves the photoresist in areas shielded by the mask pattern, forming the complete pattern. Residual developer is removed with deionized water, and the sample is dried using a nitrogen gun. Post-development baking is performed: samples are placed on a 120 °C heating platform for three minutes to remove residual solvents and enhance adhesion between the surface and substrate. Subsequently, copper is deposited onto the photoresist surface via magnetron sputtering. After deposition, samples are immersed in acetone to remove the photoresist, followed by rinsing with deionized water. Finally, the sample is immersed in hydrofluoric acid (HF) solution to strip the silicon substrate, yielding a copper layer coated with polyimide. Additionally, a VO₂ film was prepared (as shown in Figure (b)). First, a precursor solution was prepared: polyethyleneimine (PEI), ethylenediaminetetraacetic acid (EDTA), and ammonium metavanadate were sequentially added to deionized water and stirred to dissolve, forming a yellow-green vanadium metal ion-complexed polymer solution. Ultrafiltration using filter paper then removes excess aqueous solution and free small molecules, yielding the PAD-method precursor solution. The metal ion concentration in the precursor solution is subsequently measured using an inductively coupled plasma emission spectrometer. If ion doping is required, the dopant ions are added to the precursor solution at an appropriate ratio and stirred uniformly. The precursor solution is then spin-coated onto a clean substrate at 6000 rpm. Finally, the sample coated with the precursor solution colloidal film is placed in a high-temperature tube furnace. Under a reducing atmosphere of 98.5% nitrogen and 1.5% hydrogen, the reducing gas first passes through a scrubber bottle filled with deionized water, annealing the precursor film in a humid environment. Subsequently, the film was held at 450 °C for 2 h to decompose and volatilize the PEI and EDTA. The temperature was then raised to 500 °C and maintained for 2 h to crystallize the film. Finally, the film was slowly cooled to room temperature before peeling off the VO₂ film. However, radial thickness striations readily appear in the VO₂ layer during spin coating at 6000 rpm. Moreover, fluctuations in water vapor partial pressure of ± 2% during the two-step annealing at 450–500 °C cause the V4+/V3+ ratio to drift by > 5%, leading to deviation from the designed phase transition temperature. Additionally, 24.8 μm thick PI-copper-VO₂ stacks develop 1–3 μm microcracks during HF etching of silicon wafers due to interfacial stress differences, causing THz peak splitting and approximately 30% FWHM increase. Therefore, the plan involves first introducing staged variable-speed spin coating + closed-loop PID water bubbling to suppress VO₂ thickness non-uniformity to < 2 nm. Subsequently, a 200 nm Parylene C transition layer will be inserted between PI and VO₂, with stress-release grooves etched at 500 μm pitch to reduce crack density to < 0.2 mm⁻¹. For copper patterning, switch to LOR-10 A/AZ5214 dual-layer photoresist supplemented with O2plasma + low-power ultrasonic stripping to reduce side etch width below 0.2 μm; Finally, a 5 nm Al2O2 diffusion barrier was deposited via ALD between Cu and VO2. The annealing atmosphere was changed to 0.5% O2/99.5% N2 to suppress Cu diffusion, meeting the requirements for high-sensitivity liquid detection. Fig. 7Process flow chart. Process flow chart. Finally, this study compiles comparative statistical data between recent similar studies and the present research, as detailed in Table 3. Table 3Key metrics comparison Table.ReferencesApplicationsSensitivity (THz/RIU)Year published 49 Absorber0.082023 50 Bio Sensor0.682024 30 Multi-disease diagnosis sensor0.952026 51 Absorber0.058/0.0712025This workAbsorber3.33– Table 3Key metrics comparison Table.ReferencesApplicationsSensitivity (THz/RIU)Year published 49 Absorber0.082023 50 Bio Sensor0.682024 30 Multi-disease diagnosis sensor0.952026 51 Absorber0.058/0.0712025This workAbsorber3.33– Key metrics comparison Table. Conclusion In this study, this study successfully established a label-free ethanol detection method based on terahertz absorption peak drift in metamaterials. A five-layer coupled structure achieved high sensitivity of 3.33 THz/RIU and an outstanding figure of merit (FOM) of 66 RIU⁻¹. DNN optimization and CST simulation jointly revealed that droplet volume, ethanol concentration, and ambient temperature induce absorption peak red shift, blue shift, and peak spacing compression, respectively. Post-heating induces a secondary safeguard due to the abrupt increase in VO₂ conductivity, significantly enhancing detection reliability and repeatability. This strategy, which is label-free, responsive, and compatible with various preparation methods, offers a novel approach for trace liquid analysis in food safety, environmental monitoring, and biomedicine. It provides a viable paradigm for the design and industrialization of next-generation intelligent terahertz sensors.