Open Access
Add to BibSonomy
Abstract
The whispering gallery mode (WGM) optical microresonator sensors are emerging as a promising platform for precise temperature measurements, driven by their excellent sensitivity, resolution and integration. Nevertheless, challenges endure regarding stability, single resonant mode tracking, and real-time monitoring. Here, we demonstrate a temperature measurement approach based on convolutional neural network (CNN), leveraging the recognition of multimode barcode images acquired from a WGM microbottle resonator (MBR) sensor with robust packaged microresonator-taper coupling structure (packaged-MTCS). Our work ensures not only a high sensitivity of −14.28 pm/℃ and remarkable resolution of 3.5 × 10−4 ℃ across a broad dynamic range of 96 ℃ but also fulfills the demands for real-time temperature measurement with an average detection accuracy of 96.85% and a speed of 0.68s per image. These results highlight the potential of high-performance WGM MBR sensors in various fields and lay the groundwork for stable soliton microcomb excitation through thermal tuning.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
The increasing demand for temperature monitoring can be attributed to the quick advancement in fields like biomedicine [1], microelectronics [2], and optofluidics [3]. Precise and non-destructive temperature measurement [4] and compensation [5] are crucial in special application scenarios [6], where exploration of complex dynamics in biochemical systems [7], transition dynamics of phase-change materials [8], and thermo-optic effects in nonlinear optical processes [9] is required. Presently, modern temperature sensing technologies and devices, including nanothermometers [10], microwave temperature sensors [11], thermal imaging cameras [12], and deep learning-based temperature prediction models [13], have made significant progress in measurement scale, response time, visualization, and large-scale applications. Nevertheless, persistent issues encompass elevated expenses, reduced resistance to interference, intricate calibration processes, and limited generalization capabilities. Consequently, researchers are focusing on enhancing sensitivity, resolution and integration while simultaneously addressing these obstacles.
Over the past decade, micro/nano integrated sensor devices employing optical technology, particularly whispering gallery mode (WGM) optical microresonators [14], have emerged as a promising platform for precise temperature measurements [15]. WGM microresonators significantly enhance the interaction between light and matter by confining low-power light at the micro/nano scale for a prolonged duration, allowing even subtle temperature changes to induce noticeable shifts in resonance wavelength [16]. The high quality (Q) factor of WGM microresonator corresponds to an ultra-narrow linewidth of the resonant wavelength, facilitating the precise observation of minute changes in resonance [17]. In addition to their high sensitivity and resolution, WGM microresonator temperature sensors also exhibit exceptional performance with respect to measurement range [18], portability [19], and environmental adaptability [20]. Numerous reports have reported on WGM microresonator temperature sensors using different materials (Silica (SiO2) [18], silicon nitride (Si3N4) [5], and polymers [21], etc.) and various shapes (Microspheres [22], microrings [5], and microbottles [16], etc.). However, considerable challenges endure, spanning concerns related to stability, single resonant mode tracking, and real-timing monitoring thereof.
The stability of the microresonator-taper coupling structure (MTCS) [23] can be compromised by inherent defects, such as tapered fiber with a waist diameter of less than 2 μm [24], and external factors, including dust, airflow, and mechanical vibrations [25]. To address these issues and optimize cost-efficiency, commonly employed strategies involve wholly-packaged [26] and spot-packaged [27] techniques, aiming to prevent deterioration and fracture of the tapered fiber while eliminating environmental noise [28]. Despite the stabilization achieved by packaged microresonator-taper coupling structure (packaged-MTCS), there are limitations to the single resonance mode tracking, including the inability to distinguish measurements of multiple physical quantities [15], difficulties in locking the resonance mode [25], and a temperature measurement range limited by laser scanning [22]. An optical WGM barcode technique has been proposed, demonstrating precise temperature measurements across a broader range by monitoring multimode spectra [18]. However, obstacles persist regarding low measurement accuracy and speed. Currently, the field of image recognition, with a particular emphasis on convolutional neural network-based (CNN-based) deep-learning models, is advancing rapidly. This is attributed to their remarkable potential in optical sensing applications [29], offering both high detection accuracy and speed.
In this work, as such, we propose a temperature measurement technique that combines multimode spectra from WGM microbottle resonator (MBR) sensors, formed by a robust packaged-MTCS, with a CNN-based deep-learning model for the recognition of multimode barcode images. This approach exhibits remarkable efficacy in both the inherent temperature sensing parameters (sensitivity, resolution, and dynamic range) of the sensor and real-time temperature measurement capabilities (measurement accuracy and speed) provided by the CNN-based deep-learning model. Employing an unconventional approach that deviates from wholly-package and spot-package structures, specifically by omitting ultraviolet (UV) glue injection at the coupling point, we showcase packaged-MTCS with highly robustness. Leveraging CNN-based deep-learning model and barcode images derived from the mapping of the temperature-dependent overall multimode spectra of the WGM MBR sensor, we achieve real-time temperature measurements. Furthermore, there is currently a gap in existing research regarding the integration of a WGM MBR temperature sensor with an image recognition intelligent system utilizing a CNN-based deep-learning model. This represents a significant and promising future trend for achieving real-time temperature measurement, as well as effective temperature compensation.
2. Principle and simulation of temperature sensing
Light entering a WGM microresonator with radius R and refractive index n at an incident angle near ${\pi / 2}$ is confined through total internal reflection. A WGM resonant mode occurs when the optical path in the resonator is an integer multiple of the wavelength:
Where m is the WGM angular mode number and ${\lambda _m}$ is the resonance wavelength at the initial temperature. The SiO2 MBRs confine the WGM to the resonant axis between two turning points, resulting in the generation of two resonant modes, namely azimuthal mode and axial mode. These modes have different free spectral ranges (FSR), which can be expressed as the azimuthal mode FSR, $FS{R_{\lambda ,m}}$, and axial mode FSR, $FS{R_{\lambda ,q}}$, respectively:The sensing mechanism of WGM microresonators relies on the resonance shift, as shown in Fig. 1(a), which results from the temperature-dependent refractive index and size of microresonator [30]:
Fig. 1. Principle and simulation of temperature sensing in SiO2 MBRs. (a) Sensing mechanism of WGM microresonators based on resonance shift. (b) The transmission spectra evolute with temperature, ranging from 20 ℃ to 70 ℃, the inset shows the enlarged view of the resonant dips. (c)Simulated temperature sensing sensitivity.
When the resonance wavelength of the WGM MBR sensor is shifted by one $FS{R_{\lambda ,m}}$, the temperature changes by one temperature period, $TP$:
The temperature period determines the dynamic range of temperature measurement. Both the sensitivity, S, and the wavelength resolution of the spectral measurement system, $\Delta {\lambda _{\min }}$, determine the temperature measurement resolution of the WGM MBR sensor, $\Delta {T_{\min }}$:
The finite difference time domain (FDTD) solutions in the Lumerical software was used to simulate the sensing performance of the SiO2 MBR ($n = 1.4442$, $\alpha = 1.19 \times {10^{ - 5}}$, $\beta = 5.5 \times {10^{ - 7}}$). By changing the refractive index and size of microresonator to simulate the temperature change from 20 ℃ to 70 ℃ in 10 ℃ increments, the resonant wavelength at 1105.09 nm will be redshifted with the change of temperature, as shown in Fig. 1(b). The simulation sensitivity (13.71 pm/℃) was determined through linear fitting of the five datasets, as shown in Fig. 1(c), exhibiting close agreement with the theoretical sensitivity (13.76 pm/℃).
3. Preparation of SiO2 MBR sensors with packaged-MTCS
The high-Q SiO2 MBRs with varying geometric parameters, depicted in Fig. 2(a), were prepared by using arc discharge of a fiber fusion splicer (Comcore, PFS-500S) to melt single-mode fiber (Corning, SMF-28e+). Here, the MBR with uniform morphology was successfully fabricated by setting the appropriate discharge intensity (150bit), discharge duration (1s), discharge position (∼3 mm), and discharges times (3 times). The successful fabrication of a tapered fiber with 2.13 μm waist diameter that matches the phase of the MBR depended on precise control of heating range (∼2 mm), hydrogen gas flow rate (180 mL/min), and tapering speed (300 μm/s) using the hydrogen-oxygen flame tapering technique. The Polydimethylsiloxane (PDMS, Dow Corning, Sylgard 184) covers with cross-shaped grooves (2 × 2 mm2 cross section) were created by reverse molding technique, followed by precision shaping of the inner surfaces of these cross-shaped grooves using a femtosecond laser system (Light Conversion, UAB, PH2-SP).
Fig. 2. WGM MBR Sensor based on packaged-MTCS. (a) The SiO2 MBR with varying bottleneck distances ${L_b}$, bottle radiuses ${R_b}$, stem radiuses ${R_s}$ and axial curvatures $\Delta k$. (b) Schematics of packaged-MTCS. (c) The setup for locating the optimal coupling point between the SiO2 MBR and the tapered fiber at different MBR diameters, tapered fiber diameters, and coupling gaps. (d), (e) Photograph of MTCS at various microscope magnification levels. AAM: Aluminum alloy mold, FRFH: Flexible resistive foil heater.
Figure 2(b) shows the schematics of packaged-MTCS. The tapered fiber and MBR were affixed to the glass film to create the MTCS using a precise amount of UV glue. Subsequently, the MTCS was securely bonded with a PDMS cover and wrapped by an aluminum alloy mold (AAM) to form a robust packaged-MTCS. The packaged-MTCS was then adhered to a flexible resistive foil heater for conducting temperature sensing experiments. The fabrication and packaging steps of the packaged-MTCS are detailed in Supplement 1, Figs. S1-S4. Before packaging, the MBR underwent controlled movement along the x, y, and z directions using a five-axis motorized translation stage with 30 nm resolution (Thorlabs, 8081-M) was employed to move, as shown in Fig. 2(c)-(e) and Supplement 1, Fig. S5(b). This adjustment involved identifying the optimal coupling point, considering variations in MBR diameters, tapered fiber diameters, and coupling gaps, as shown in Supplement 1, Figs. S6-S8, respectively, with the purpose of the optical resonance characteristics.
4. Experimental results and discussion
4.1 Optical characterization of SiO2 MBR sensors
The pump light from a continuous wave (CW) tunable laser (Toptica, CTL 1550, 10 kHz linewidth) was coupled to the high-Q SiO2 MBRs for performance characterization, following its passage through a polarization controller (PC), a fiber optical isolator (FOI), and a variable optical attenuator (VOA). To accurately measure the Q factor and the relative frequency of the resonance dips, a 50/50 optical coupler (OC) was used to split the light into the MTCS /packaged-MTCS and the fiber Mach-Zehnder interferometer (MZI) as a reference, respectively. The transmission spectra were recorded by a high-speed real-time oscilloscope (OSC, Keysight, UXR0404A) after being converted into electrical signals by photodetectors (PD, Koheron, PD100). The schematic of the experimental setup for characterizing the performance of MTCSs or packaged-MTCSs is shown in Supplement 1, Fig. S5(a). The pump light of ∼15μW at 1550.96 nm was injected into the tapered fiber with a wide scan range greater than the $n \times FSR$ and a scan speed of 1 nm/s. After obtaining the multimode transmission spectra quickly, the high-resolution resonant dips were obtained by periodic fine scanning in the range of 150 pm (Corresponding to the scanning amplitude Vpp = 75 V) with the scanning mode of triangular wave and the scanning speed of 25 Hz.
Compared to other microresonators such as microspheres, microrings, or microdisks, the SiO2 MBRs can excite a denser transmission spectrum by combining azimuthal and axial modes, which makes it more advantageous in multimode temperature sensing. Obviously, the transmission spectra excited by wide scanning (Fig. 3(a) and (b)) and fine scanning (Fig. 4(a) and (b)) are relatively dense before and after packaging, but the extinction ratios have changed due to small disturbances during the injection of UV glue. The selected resonance dips, with typical loaded Q factors of $2.27 \times {10^7}$ and $1.90 \times {10^7}$, are presented in Fig. 4(c) and (d), along with the reference 9.2 MHz fiber MZI trace. It is observed that the resonant dip broadens after the packaging. The number of prominent resonant dips in a single $FS{R_{\lambda ,m}}$ is 201 and 194, respectively, and the corresponding load Q factor distribution are shown in Fig. 3(c) and (d). Although the average Q factor slightly reduces after packaging, it remains on the order of ${10^7}$, which is much higher than the highest Q factor of the Si3N4 microring resonator [5]. The reason for this is that the designed packaged-MTCS deviates from wholly-package and spot-package, with no UV glue injected at the coupling point. This not only ensures the robustness of packaged-MTCS but also does not compromise the related Q factor associated with measurement resolution.
Fig. 3. Characterization of the performance of MTCS and packaged-MTCS by wide scanning. The typical transmission spectra of an MBR (${L_b} = 345.54{\ \mathrm{\mu} \mathrm{m}}$, ${R_b} = 193.04{\ \mathrm{\mu} \mathrm{m}}$, ${R_s} = 61.98{\ \mathrm{\mu} \mathrm{m}}$, $\Delta k = 0.0026\mathrm{\ \mu }{\textrm{m}^{\textrm{ - 1}}}$) coupled with tapered fiber (2.13 μm waist diameter) in the range of 1550 to 1552 nm both in (a) MTCS and (b) packaged-MTCS. Histogram of the load Q factors for (c) 201 and (d) 194 prominent resonant dips from (a) and (b), respectively.
Fig. 4. Characterization of the performance of MTCS and packaged-MTCS by fine scanning. The typical transmission spectra both in (a) MTCS and (b) packaged-MTCS. The measured single resonance dip (red dot) and its Lorentzian fit (red line) used to demonstrate load Q factors, and the measured fiber MZI reference (blue dot) and its Sinusoidal fit (blue line) used for calibration both in (c) MTCS and (d) packaged-MTCS.
4.2 Robustness characterization of packaged-MTCS
The Q factors of both MTCS and packaged-MTCS were continuously monitored in standard laboratory conditions and under extreme conditions simulated by smoke flow through a glass dome, aiming to assess their stability. Notably, the average Q factors of packaged-MTCS remain stable at ${10^7}$ in both conditions, as depicted by blue dotted line in Fig. 5(a) and (b), although exhibiting ultra-small fluctuations. However, the average Q factor of MTCS (depicted by red dotted line in Fig. 5(a)) experiences a significant decrease of approximately one order of magnitude over 30 days, attributed to the presence of dust and airflow in the regular experimental environment. Similarly, the average Q factor of MTCS (depicted by red dotted line in Fig. 5(b)) drops sharply by two orders of magnitude within 6 minutes of the smoke exposure and then levels off after pollution reached its peak.
Fig. 5. Robustness characterization of packaged-MTCS by Q-stability test and vibration test. Changes in average load Q factors for MTCS (red dotted line) and packaged-MTCS (blue dotted line) in (a) normal and (b) extreme environments. The green dashed line represents the introduction of smoke flow inside the glass dome to simulate the extreme environment. The shifts of the resonance wavelength (green dots) both in (c) MTCS and (d) packaged-MTCS over 30 trials, and the typical resonant dips before (blue line) and after (red line) vibration.
The robustness of the proposed packaged-MTCS was further demonstrated by real-time monitoring of the WGM resonant dips of MTCS and packaged-MTCS at a vibration frequency of 150 Hz in a constant temperature and humidity environment. As shown in Fig. 5(c), the typical resonant dip of MTCS is shifted by a maximum of 10.21 pm after vibration in 30 trials, and its shape changes greatly. However, the typical resonant dip of packaged-MTCS is shifted by only a maximum of 0.0036 pm, as shown in Fig. 5(d). Although the extinction ratio is slightly decreased, its Q factor remains around $6 \times {10^7}$, indicating that packaged-MTCS can resist small mechanical disturbances and exhibit good robustness. Through the Q-stability test and vibration test, it is apparent that the proposed packaged-MTCS can not only sustain a high Q factor, but also provide a promising and dependable avenue for applications like temperature sensing, where the requirements for portability, flexibility, and robustness are particularly exacting.
4.3 Measurement of temperature sensing parameters
In packaged-MTCS with good robustness, temperature is the only factor that varies the transmission spectra. To achieve temperature sensing parameters, including sensitivity, resolution, and dynamic range, the packaged-MTCS was mounted to a flexible resistive foil heater (FRFH, Thorlabs, TLK-H), with temperature control managed by a heater temperature controller (HTC, Thorlabs, TC300) offering a resolution of 0.1 ℃. When the temperature was controlled at a stable value (32.0 ℃, 32.1 ℃, 32.2 ℃, etc.), the spectral shift was observed with the OSC and the data was recorded with a data acquisition card (DAQ, IN, USB-6009) triggered by the laser at the same frequency. The experimental setup is depicted in Fig. 6(a).
Fig. 6. Temperature measurement system and performance characterization. (a) Schematic of the experimental setup for temperature measurement. Evolution of transmission spectra of MBR with (b) increasing or (c) decreasing temperature in a single experiment. For illustration purposes and clarity, only the 10 pm transmission spectrum around the traced resonant dip (red line) is shown. Inset shows the enlarged view of the 10 pm transmission spectrum at 35 ℃. Data averaged over 10 experiments (blue dots) of resonance dip wavelength shift with (d) increasing or (e) decreasing temperature and their Linear fittings (blue line), confidence bands (dark blue areas), and prediction bands (light blue areas). PC: Polarization controller; FOI: Fiber optical isolator; VOA: Variable optical attenuator; OC: Optical coupler; PD: Photodetector; DAQ: Data acquisition card; HTC: Heater temperature controller; OSC: Oscilloscope.
The temperature range, from 30°C to 40.0°C, was chosen for temperature measurements based on its practical relevance in daily life and industry, aligning with common literature practices [18,31]. During a controlled temperature increase from 30°C to 39°C in 1°C increments, the resonant dip exhibits a blueshift of 125.56 pm, as illustrated in Fig. 6(b), representing the outcome of a single experiment. The average sensitivity, determined through Linear fitting from 10 experiments, is $- 14.38$ pm/°C (Fig. 6(d)). Similarly, during a controlled temperature reduction from 40°C to 31°C in 1°C decrements, the resonant dip experiences redshift of 128.11 pm, as shown in Fig. 6(c). The average sensitivity from Linear fitting is $- 14.17$ pm/°C (Fig. 6(e)). Consequently, the overall average sensitivity of the WGM MBR sensor is $- 14.28$ pm/℃, a value influenced by the presence of UV glue. The thermo-optical coefficient of UV glue is an order of magnitude larger and negative compared to that of silica glass, leading to both a blue shift in the spectrum with rising temperature and an enhancement in temperature sensing capability.
The temperature measurement dynamic range of the WGM MBR sensor is determined by the $FS{R_{\lambda ,m}}$. Excluding the inherent limitation of CW laser, which is a maximum fine scanning range of 280 pm determined by a maximum modulation voltage of 140 V, the dynamic range can be calculated as 96°C, according to Eq. (6). Hence, reducing the MBR radius has the potential to broaden the temperature measurement dynamic range, provided that the sensor material remains resilient to irreversible damage within the specified range. The temperature measurement resolution of the WGM MBR sensor is determined by both the sensitivity and the wavelength resolution of the spectral measurement system. The latter, in turn, is determined by a narrow-linewidth CW tunable laser and a packaged-MTCS with a high Q factor corresponding to a narrow-linewidth of 0.05 pm. Assuming that the minimum resolvable resolution of resonance is ${1 / {10}}$ of the linewidth, the WGM MBR sensor has the resolution of $3.5 \times {10^{ - 4}}$ ℃, according to Eq. (7). Compared to the outcomes presented in [18], our work demonstrates a sensitivity approximately three times higher and achieves a resolution surpassing it by an order of magnitude. Additionally, in comparison to the findings in [32,33], our work provides enhanced flexibility for tailoring the temperature measurement dynamic range to MBRs radii and specific application requirements.
4.4 CNN-based temperature measurement with multimode barcode images
From Fig. 6(b) and (c), it is clear that once the single-mode spectrum shifts beyond the designated fine scanning range of 150 pm, its capability is limited to measuring temperature changes within a range of ∼10.5°C. Moreover, the maximum fine scanning range of CW laser is 280 pm, corresponding to a measurement range of only 19.6°C. Evidently, single-mode tracking cannot harness the full potential of the theoretically wide dynamic range for temperature measurement provided by a broad FSR. Furthermore, Fig. 7(a) shows that the MBR retains multiple WGM modes within the 10 pm scanning range. Even if a certain WGM mode moves out of the 10 pm scanning range due to temperature variation, it does not impact the acquisition of key information within the overall transmission spectrum influenced by thermal-optical and thermal expansion effects, until the mode reappears following a $1 \times FS{R_{\lambda ,m}}$shift.
Fig. 7. Temperature-dependent multimode sensing mechanisms. (a) Multimode transmission spectra of MBR and (b) their corresponding multimode barcode images within the 10 pm range. (c)Schematic of CNN-based temperature measurement with multimode barcode images.
Hence, the transmission spectra are mapped into the barcode images (Fig. 7(b)), with each line denoting an individual WGM mode, where the position of the line indicated the resonant frequency of the mode, the width represented the linewidth, the color denoted the extinction ratio, and the gap between two lines reflected the mode spacing. The CNN-based deep-learning model is used to extract the features in these images and achieve temperature classification, instead of relying on transmission spectral information from SiO2 WGM sensors with packaged-MTCS (Fig. 7(c)). The key to temperature measurement of CNN-based multimode barcode images is that barcodes obtained at the same temperature should have good repeatability. This was confirmed by subjecting the system to 100 cycles of temperature variation, oscillating between 30°C and 40°C, as illustrated in Supplement 1, Fig. S9 and Dataset 1 (Ref. [34]).
Creating a temperature-dependent multimode barcode image dataset was a prerequisite for real-time temperature measurement using CNN-based image recognition technology. The colormap was utilized to transform 10,000 sets of spectral data, recorded 100 times at 0.1°C intervals between 30.1°C and 40.0°C, into barcode images for training. The training set, as shown in Fig. 8(a), was initially processed through a feature extraction network that consisted of standard convolutions and inverted residual blocks comprising depthwise convolutions and pointwise convolutions. Then, it was fed into the classifier, which included a global average pooling (GAP) layer, a fully connected (FC) layer, and a softmax function. Each temperature-dependent multimode spectrum was measured 20 times and converted into barcode images as a test set, as shown in Fig. 8(b). This test set was then input into the pre-trained model for real-time temperature measurement. The architecture of CNN-based deep-learning model is depicted in Fig. 8(c), with more detailed description available in Supplement 1.
Fig. 8. Schematic of CNN-based deep-learning model for training and testing. (a) A training set containing 10,000 multimode barcode images. (b) A testing set containing 2000 multimode barcode images. (c) The CNN-based deep-learning model. Conv2d: Two-dimensional convolution; GAP: Global average pooling; BN: Bottleneck; FC: Fully connected; PW conv: Pointwise convolution; DW conv: Depthwise convolution.
The lightweight deep-learning model results in a significant reduction in computational load and parameter count, resulting in a substantial enhancement in measurement speed while maintaining accuracy (Fig. 9(a)). This makes it suitable for integration into many mobile applications and embedded systems, including smartphones, autonomous vehicles, and facial recognition. Figure 6(d) illustrates the model performance assessed on the test set after completing 500 training steps in the training set. The training accuracy exceeds 96% after 70 training steps and the validation accuracy surpasses 98% after 90 validation steps. The training and validation losses, representing the disparity between predicted and actual outcomes, exhibit values below 0.07 and 0.04, respectively, with asymptotic stability achieved after 60 training steps and 80 validation steps. Figure 9(b) illustrates the model's precise classification of multimode barcode images into 10 temperature classes ranging from 30.1 ℃ to 31.0 ℃, with only 4 images misclassified, and the extent of misclassification is minor. The confusion matrices for other temperatures are shown in Supplement 1, Fig. S10. In general, the model realizes the temperature measurement of multimode barcodes based on CNN with an average detection accuracy of 96.85% and an average detection speed of 0.68 ms per image.
Fig. 9. Characterization of CNN-based temperature measurement with multimode barcode images. (a) The accuracy and loss of the model during training. (b) The confusion matrix obtained by testing 200 images at temperature ranging from 30.1 ℃ to 31.0 ℃.
5. Conclusion
In summary, the temperature measurement using WGM MBR sensor with packaged-MTCS and a CNN-based deep-learning model for multimode barcode images was demonstrated to improve the performance in both temperature sensing parameters and real-time temperature measurement capabilities. The fabricated SiO2 WGM MBR sensors demonstrate commendable temperature sensing parameters, including exceptional sensitivity, elevated resolution, and an extensive dynamic range, owing to their material composition, structural properties, high Q factors, and large-sized MBRs with rich modes. The packaged-MTCSs are validated for its excellent repeatability and robustness through Q-stability test conducted in various conditions, as well as vibration test carried out in a constant temperature and humidity environment. The lightweight deep-learning model significantly reduces computational load and parameter count, leading to a substantial increase in measurement speed without compromising accuracy. It is only necessary to calibrate the sensing parameters of the WGM MBR sensors and employ the multimode barcode images for model training. This allows for the direct extraction of the real-time temperature using temperature-dependent multimode spectra and the trained model, eliminating the requirement to acquire details such as resonant frequency, linewidth, mode spacing, and other information from the transmission spectrum. This advancement paves the way for the intelligence and commercial viability of WGM MBR sensors.
Funding
National Key Research and Development Program of China (2022YFF0705700); National Natural Science Foundation of China (52305595, U22A20205); Young Elite Scientists Sponsorship Program by CAST (2022QNRC001).
Acknowledgments
The authors express gratitude to Xueying Jin (Hefei University of Technology, Hefei, China) for her assistance with the fabrication of the structures and the design of the characterization setup. Ziye Wang (Jiangsu University of Science and Technology, Zhenjiang, China) contributed valuable insights during discussions on the measurement setup and conducted the initial analysis of transmission spectra.
Disclosures
The authors declare no conflicts of interest.
Data availability
The multimode barcode images can be found in Dataset 1 [34] and the rest of the data may be obtained from the authors upon reasonable request.
Supplemental document
See Supplement 1 for supporting content.
References
1. Y. L. Zhu, Z. Wang, R. X. Zhao, et al., “Pt decorated Ti3C2Tx MXene with NIR-II light amplified nanozyme catalytic activity for efficient phototheranostics,” Acs Nano 16(2), 3105–3118 (2022). [CrossRef]
2. C. D. S. Brites, S. Balabhadra, and L. D. Carlos, “Lanthanide-based thermometers: at the cutting-edge of luminescence thermometry,” Adv Opt Mater 7(5), 1801239 (2019). [CrossRef]
3. Y. X. Ren, G. G. K. Yip, L. M. Zhou, et al., “Hysteresis and balance of backaction force on dielectric particles photothermally mediated by photonic nanojet,” Nanophotonics 11(18), 4231–4244 (2022). [CrossRef]
4. P. Wang, W. Yu, G. X. Li, et al., “Printable, flexible, breathable and sweatproof bifunctional sensors based on an all-nanofiber platform for fully decoupled pressure-temperature sensing application,” Chem Eng J 452, 139174 (2023). [CrossRef]
5. C. Zhang, G. G. Kang, J. Wang, et al., “Photonic thermometer by silicon nitride microring resonator with milli-kelvin self-heating effect,” Measurement 188, 110494 (2022). [CrossRef]
6. J. S. Zhong, D. Q. Chen, Y. Z. Peng, et al., “A review on nanostructured glass ceramics for promising application in optical thermometry,” J Alloy Compd 763, 34–48 (2018). [CrossRef]
7. H. D. Liu, C. F. Du, L. L. Liao, et al., “Approaching intrinsic dynamics of MXenes hybrid hydrogel for 3D printed multimodal intelligent devices with ultrahigh superelasticity and temperature sensitivity,” Nat. Commun. 13(1), 3420 (2022). [CrossRef]
8. K. P. Kelley, S. V. Kalinin, E. Eliseev, et al., “Probing temperature-induced phase transitions at individual ferroelectric domain walls,” Adv Electron Mater 9(1), 2200552 (2023). [CrossRef]
9. T. Tan, Z. Yuan, H. Zhang, et al., “Multispecies and individual gas molecule detection using Stokes solitons in a graphene over-modal microresonator,” Nat. Commun. 12(1), 6716 (2021). [CrossRef]
10. M. Quintanilla and L. M. Liz-Marzan, “Guiding rules for selecting a nanothermometer,” Nano Today 19, 126–145 (2018). [CrossRef]
11. C. R. Lavers, J. D. Fisk, and B. J. T. Lavers, “Environmental temperature and material characterisation with planar evanescent microwave sensors for environmental analysis,” Proc. SPIE 11525, 638–646 (2020). [CrossRef]
12. W. Minkina, “Problems of remote temperature measurement of small objects of electricity power systems-on the example of lashing clamps of bridge connections on high voltage poles,” Energies 14(16), 5041 (2021). [CrossRef]
13. Y. Gao, S. Miyata, and Y. Akashi, “Multi-step solar irradiation prediction based on weather forecast and generative deep learning model,” Renew Energ 188, 637–650 (2022). [CrossRef]
14. D. Yu, M. Humar, K. Meserve, et al., “Whispering-gallery-mode sensors for biological and physical sensing,” Nature Reviews Methods Primers 1(1), 83 (2021). [CrossRef]
15. D. Q. Yang, J. H. Chen, Q. T. Cao, et al., “Operando monitoring transition dynamics of responsive polymer using optofluidic microcavities,” Light Sci Appl 10(1), 128 (2021). [CrossRef]
16. Y. H. Yin, T. X. Nie, and M. Ding, “Temperature sensor based on microbottle resonator immersed in isopropanol,” IEEE Photonics J. 13(3), 1–7 (2021). [CrossRef]
17. I. Brice, K. Grundsteins, K. Draguns, et al., “Whispering gallery mode resonator temperature compensation and refractive index sensing in glucose droplets,” Sensors 21(21), 7184 (2021). [CrossRef]
18. J. Liao and L. Yang, “Optical whispering-gallery mode barcodes for high-precision and wide-range temperature measurements,” Light: Sci. Appl. 10(1), 32 (2021). [CrossRef]
19. C. Li, M. Zhu, P. Ji, et al., “In-fiber BaTiO3 microsphere resonator for high-sensitivity temperature measurement,” Micromachines 12(3), 318 (2021). [CrossRef]
20. Q. J. Lu, X. G. Chen, L. Fu, et al., “On-chip real-time chemical sensors based on water-immersion-objective pumped whispering-gallery-mode microdisk laser,” Nanomaterials 9(3), 479 (2019). [CrossRef]
21. Z. Y. Li, C. R. Liao, J. Wang, et al., “Femtosecond laser microprinting of a fiber whispering gallery mode resonator for highly-sensitive temperature measurements,” J. Lightwave Technol. 37(4), 1241–1245 (2019). [CrossRef]
22. J. Yu, E. Lewis, G. Brambilla, et al., “Temperature sensing performance of microsphere resonators,” Sensors 18(8), 2515 (2018). [CrossRef]
23. Z. L. Newman, V. Maurice, T. Drake, et al., “Architecture for the photonic integration of an optical atomic clock,” Optica 6(5), 680–685 (2019). [CrossRef]
24. L. Cai, J. Y. Pan, and S. Hu, “Overview of the coupling methods used in whispering gallery mode resonator systems for sensing,” Opt Laser Eng 127, 105968 (2020). [CrossRef]
25. X. Jiang, A. J. Qavi, S. H. Huang, et al., “Whispering-gallery sensors,” Matter 3(2), 371–392 (2020). [CrossRef]
26. K. Wang, H. Wang, X. Y. Wu, et al., “Ultrasound sensing using packaged microsphere cavity in the underwater environment,” Sensors 22(11), 4190 (2022). [CrossRef]
27. Y. Z. Yan, C. L. Zou, S. B. Yan, et al., “Robust spot-packaged microsphere-taper coupling structure for in-line optical sensors,” IEEE Photonics Technol. Lett. 23(22), 1736–1738 (2011). [CrossRef]
28. Y. N. Zhang, T. Zhou, B. Han, et al., “Optical bio-chemical sensors based on whispering gallery mode resonators,” Nanoscale 10(29), 13832–13856 (2018). [CrossRef]
29. Y. Nishizaki, M. Valdivia, R. Horisaki, et al., “Deep learning wavefront sensing,” Opt Express 27(1), 240–251 (2019). [CrossRef]
30. M. P. Marisova, A. V. Andrianov, and E. A. Anashkina, “Comprehensive numerical analysis of temperature sensitivity of spherical microresonators based on silica and soft glasses,” Sensors 23(2), 717 (2023). [CrossRef]
31. X. G. Li, F. Li, X. Zhou, et al., “Optical fiber DNA biosensor with temperature monitoring based on double microcavities Fabry-Perot interference and vernier combined effect,” IEEE Trans. Instrum. Meas. 72, 1–8 (2023). [CrossRef]
32. C. Zhang, J. Wang, G. G. Kang, et al., “Soliton microcomb-assisted microring photonic thermometer with ultra-high resolution and broad range,” Photonics Res. 11(10), A44–A53 (2023). [CrossRef]
33. Y. Z. Liu, J. F. Jiang, K. Liu, et al., “Ultrahigh-resolution optical fiber thermometer based on microcavity opto-mechanical oscillation,” Adv Photon Res 3(9), 2200052 (2022). [CrossRef]
34. H. Li, “Multimode barcode images,” figshare (2023), https://doi.org/10.6084/m9.figshare.24821004.v3.
