2.1 Preparation of the HDIM-Based BSGS

Figure 2a presents the synthetic structure of the HDIM, where MXene is uniformly dispersed in the self-healing prepolymer as a conductive material. The synthesis procedure of the prepolymer is depicted in Figure S1 (Supporting Information), with hydroxyl-terminated polybutadiene (HTPB), 1,10-decanediol (DE), and isophorone diisocyanate (IPDI) as the primary constituents, and dibutyltin dilaurate (DBTDL) as the catalyst. Notably, within the HDIM, HTPB acts as the soft component for controlling flexibility, while DE and IPDI collectively serve as the hard segments. They contribute to hydrogen bonding by forming urea/urethane linkages, impacting the mechanical characteristics and self-healing properties of the material.^[35] The absence of significant peaks of the N═C═O stretching bond at 2264 cm⁻¹ in the FTIR spectrum shown in Figure 2b indicates the complete conversion of diisocyanate monomers into urethane bonds, an essential requirement for the formation of hydrogen bonds.^[15] After preparing the transparent prepolymer (Figure S2, Supporting Information), a solution of single-layered Ti₃C₂T_x MXene with a concentration of 5 mg mL⁻¹ is added proportionally to the prepolymer solution to create a thick mixture, ultimately leading to the production of HDIM (Figure S3, Supporting Information). The X-ray photoelectron spectroscopy (XPS) patterns and elemental mapping in Figure 2c,d revealed the incorporation of Ti and F elements into the HDIM through MXene.^{[36, 37]} All of these results demonstrated that MXene was dispersed uniformly within the HDIM. Furthermore, the water contact angle (CA) of ≈104° evidenced the HDIM's good hydrophilicity, visible in Figure 2e. The thermogravimetric analysis (TGA) of HDIM indicated its outstanding thermal stability, even at elevated temperatures of up to 240° C (Figure S4, Supporting Information).

It is remarkable that HDIM can be re-dissolved in chloroform to produce an ink-like solution upon request (Figure 2f; Figure S5, Supporting Information). Moreover, this ink-like solution is suitable for the development of printed sensors and displays a good level of electrical conductivity when applied to commonly used substrates such as PU, PTFE, and paper (Figure 2g; Figure S6, Supporting Information).^[38] To produce a BSGS with HDIM, the ink-like property was utilized in a screen-printing process (Figure 2h). For flexibility and breathability, PU electrospun film was chosen as the substrate, and further details can be found in Supporting Information (Figures S7 and S8, Supporting Information).^[39-41] The PU electrospun film exhibited an ideal hydrophobic quality (Figure S9, Supporting Information), meaning sweat or wound fluids would not affect the performance of the HDIM. Printing procedures used the HDIM ink to create an array with a width of ≈5 mm (Figure 2i; Figure S10, Supporting Information) which enabled cost savings and improved efficiency. Additionally, the BSGS had great air permeability which ensured skin comfort upon direct contact with the sensors (Figure S11, Supporting Information).

2.2 VOCs Sensing Performance of BSGS

The assessment of repeatability is an essential measurement for sensors in actual sensing scenarios. In order to analyze this index, the BSGS was continuously assessed through three cycles. Figure 3b demonstrates that the BSGS responded reliably to the same concentration. The response time, which is measured as the time it takes to reach 90% of the maximum response, was recorded as 160 s for 10 ppm of ethanol. The recovery time, which is the time needed to go back to 10% of the stabilized response in dry air, was measured to be 265 s (Figure S13, Supporting Information). Additionally, the BSGS maintained a stable response to 10 ppm of ethanol over 180 days, without any significant degradation in terms of conductivity and sensitivity (Figure S14, Supporting Information). This emphasizes its strength and reliability.

In addition to ethanol, the BSGS was also assessed for its real-time reaction toward three other distinct VOCs (Figure 3d), with excellent reproducibility being noted (Figure S15, Supporting Information). Notably, the response speed and amplitude of BSGS were highest for chloroform among these four VOCs (Figure 3e). This could be attributed to the more potent interaction between the HDIM material and the larger polarity of chloroform.^{[46, 47]} In addition, chloroform's capability of dissolving HDIM can result in structural and property changes in the material. These changes may appear in the form of discrepancies in surface topography, conductivity, and adsorption characteristics of the HDIM, thereby amplifying the effect of the VOCs detection procedure and the related response.

2.3 Self-Healing and VOCs Sensing Properties of BSGS after Treatment

When used in wearable applications, sensors are vulnerable to external forces which may cause them to be damaged or broken.^[48-50] Consequently, it is essential to include self-healing capabilities in order to improve the longevity and dependability of these gadgets. To evaluate the self-healing properties, scratches were deliberately inflicted on HDIM surfaces and then placed in various temperature conditions. As can be seen in Figure 4a, it took roughly 72 h for the scratches to recover completely at room temperature. In contrast, at 40° C and 60° C, the healing time drastically decreased to 60 min and only 10 min, respectively, which shows that HDIM possesses a remarkable self-healing ability that can be further boosted by heating. Dynamic mechanical analysis (DMA) on HDIM disclosed more knowledge about its mechanical and self-healing behavior (Figure 4b). At room temperature, the storage modulus (E') exceeded the loss modulus (E″), which implies that the material behaves like an elastic solid. With higher temperatures, E′ and E″ dropped, demonstrating a transition to a more viscous state in HDIM (Figure S16, Supporting Information). Therefore, the transfer of hydrogen bonds in the urea/PU bonds was sped up, thereby improving the efficacy of self-healing.

The damaged HDIM reestablished its electronic conductivity when the two pieces were brought together (Figure 4c). Moreover, HDIM's conductivity stability during the stretching-relaxing process was noteworthy (Figure S17, Supporting Information), attesting to its potential in daily use. To investigate the BSGS's dependability in practical applications, we compared its VOC sensing performance before and after varied treatments. At 10 ppm of ethanol gas, there was little difference in the BSGS's response that was repaired from scratch compared to the original response (Figure 4d). Additionally, Figure 4e presents the results of the fatigue test, which included stretching and relaxing the BSGS multiple times. Despite the heightened noise level, the response value at a similar response time was little altered. These results affirm the BSGS's capability to remain resilient in testing circumstances.

2.4 VOCs Identification by MISSA

Drawing on the benefits of the gas filtration effect and stackable design exhibited in masks, we have crafted a unique self-healing sensor array, called MISSA, to obtain more information and simplify the detection of different VOCs. Figure 5a illustrates that MISSA was simply produced by aligning three equal BSGS units (with the same shape and size) in a vertical line and tightly enclosing them with tape, thus providing a layered effect. In addition, PU electrospun film, much like the filtration system found in masks, is capable of changing the diffusion rate of gas molecules arriving at layer two and layer three. Figure 5b presents the comprehensive set of various response values and their accompanying time patterns shown by each layer of MISSA in terms of ethanol gas sensing. This can be attributed to the absorption properties and obstruction effect caused by the PU film, leading to a sequential decrease in the response values and rates across layer one, layer two, and layer three. The outcomes suggest that MISSA can be proficiently used as a sensor array to create diverse dynamic reactions to the same VOC. Figure 5c portrays the response traits of MISSA to four VOCs at a concentration of 10 ppm. This unique set of code bars enables an easy examination of certain VOC types, thus permitting efficient classification within the sensor array.

To enable mobile gas surveillance and compact display, we merged the MISSA with a FPCB equipped with Bluetooth technology to develop a portable electronic system, as shown in Figures S18 and S19 (Supporting Information). Moreover, a tailor-made application program was designed to obtain and process the real-time VOC sensing data, which was subsequently exhibited on a smartphone. As illustrated in Figure 5d, when the MISSA-based electronic system was placed in the ethanol atmosphere, the smartphone screen showed three real-time response graphs matching the MISSA three-layer construction. Thus, the smartphone-compatible VOC sensing system, blending MISSA with portable circuits, features superior sensing performance and holds potential for application as a wearable covering in VOC gas monitoring.

2.5 VOCs Analysis and Identification with MISSA

Acquiring dynamic gas response curves for four gas species with concentrations of 0.2, 5, and 10 ppm through MISSA, 12 distinct features that accurately reflected the respective response curves were extracted by utilizing fitting parameters labeled according to the gas type and concentration. This process is essential to protect human health in an environment with VOCs frequently encountered in a mixture. These parameters include the maximum response value ($${\mathrm{S}}_{{\mathrm{max}}}^{\mathrm{i}}$$), response time ($${\mathrm{t}}_{{\mathrm{res}}}^{\mathrm{i}}$$), recovery time ($${\mathrm{t}}_{{\mathrm{rec}}}^{\mathrm{i}}$$), and the offset ($${\mathrm{y}}_{{\mathrm{off}}}^{\mathrm{i}}$$), where i represents the layer number of MISSA (Figure 6a). Each line connects the values of the parameters along their respective axes, representing a single point within the 12D parameter space. This visualization offers valuable insights into the clustering patterns of gas responses at specific parameters, highlighting the convergence of measurements.

To visualize data more clearly, we employed PCA dimensionality reduction to extract the primary features and transformed the 12D parameters into a 2D space represented by PC1 and PC2 (Figure 6b). From the left panel of the figure, it can be seen that PC1 (accounting for 66.6% of the variance) and PC2 (accounting for 26.1% of the variance) have the highest variances and the combination of the two amounts to a variance of ≈92%. This indicates that PCA retains almost all of the original data information and achieves effective dimension reduction. Moreover, the right panel of Figure 6b displays the loading scores of the extracted parameters onto the main principal component. The combination of PC1 and PC2 forms vectors that demonstrate their respective influences on the principal components, as suggested by the gray arrows in Figure 6c. This dimensionality reduction approach enabled the separation of four gas species, each with three different concentrations, into 12 clusters, thus making it applicable for subsequent gas identification and concentration prediction (Figure 6c). K-Nearest Neighbors (kNN) is both a straightforward and resilient method for multi-category problems with small sample sizes. To create kNN model, the PCA results were utilized as the training data due to its noteworthy efficiency and precision. As seen in Figure 6d, the PCA-assisted kNN produces a clear, discernible boundary decision map for the four VOC gas groups. We used the five-fold cross-validation and illustrated the confusion matrix in Figure 6e which shows the results of 446 test samples on the identification of the four VOC gases. This displays an impressive identification accuracy of 99.77%, further testifying to the usefulness of PCA-assisted kNN in precisely recognizing the four VOCs.

To predict the concentration of gas, a linear regression model was created using data from 0.2 and 10 ppm of ethanol. The 5 ppm data was only used to validate the model's prediction accuracy. The red dashed line in Figure 6f displays the fitted line. The model yielded a high prediction accuracy of 98.3% for ethanol (calculation details in Supporting Information). The regression surface was then projected into a 2D PC space to display the prediction regions (Figure 6g). The x-axis and y-axis represent PC1, PC2, and the z-axis corresponds to the ethanol concentration. Notably, the red block within the pink decision region (representing 5 ppm ethanol) matched the data that was not used to train the regression model. This proves the effectiveness of the PCA-assisted ML technique for identifying VOCs and accurately predicting their concentrations.