All reagents used in this study were employed as received, without further purification. Ferric chloride hexahydrate (FeCl₃·6H₂O, AR), 1,3,5-benzenetricarboxylic acid (H₃BTC, AR), and terephthalic acid (H₂BDC, AR) were purchased from Aladdin Reagent Co., Ltd. N, N-dimethylformamide (DMF, AR), ammonia solution (NH₃·H₂O, AR), triethylamine (C₆H₁₅N, AR), methanol (CH₄O, AR), ethanol (C₂H₆O, AR), isopropanol (C₃H₈O, AR), n-butanol (C₄H₁₀O, AR), acetone (C₃H₆O, AR), acetonitrile (C₂H₃N, AR), and trimethylamine (C₃H₉N, AR) were obtained from Sinopharm Chemical Reagent Co., Ltd. Deionized water was prepared in the laboratory.

MIL-100(Fe) was synthesized using a hydrothermal method according to a reported procedure^[20]. Briefly, FeCl₃·6H₂O (4 mM) and H₃BTC (4 mM) were dissolved in 80 mL of DMF by ultrasonic agitation for 30 min. The solution was transferred to a 100 mL Teflon-lined autoclave and heated at 150°C for 12 h. After cooling to room temperature, the product was sequentially washed with DMF and anhydrous ethanol and then dried to obtain MIL-100(Fe). The synthesized MIL-100(Fe) was thermally treated for 12 h in a ceramic crucible under a nitrogen atmosphere at temperatures ranging from 200 to 600 °C. The resulting products were designated as Fe_x (x = 200, 300, 400, 500, and 600), where x represents the heat-treatment temperature. The synthesis process is illustrated in Supplementary Figure 1.

In the preparation of MIL-100(Fe), different amounts of terephthalic acid were added (5 mol.% H₂BDC, 10 mol.% H₂BDC, 15 mol.% H₂BDC, and 20 mol.% H₂BDC). The mixture was subjected to ultrasonic treatment followed by a hydrothermal reaction. After the reaction, the products were cooled to room temperature, sequentially washed with DMF and ethanol, and then dried. The synthesized powder was then heated at 400 °C for 12 h under a nitrogen atmosphere. The resulting samples were named x%-Fe400 (where x = 5, 10, 15, and 20). The synthesis flowchart is shown in Supplementary Figure 1.

Powder X-ray diffraction (XRD) patterns were recorded on a TD-3500 X-ray diffractometer (Dandong Tongda Science & Technology Co., Ltd., China) using Cu Kα radiation (λ = 1.54056 Å) over a 2θ range of 10-80° at a scanning rate of 6° min^-1 to investigate the phase composition and crystal structure of the samples. The surface morphology was observed by scanning electron microscopy (SEM, ZEISS Gemini Sigma 300, Germany). Transmission electron microscopy (TEM) and high-resolution transmission electron microscopy (HRTEM) analyses were performed on an FEI Talos F200S microscope to examine the microstructure and crystalline features of the materials. Elemental mapping was conducted simultaneously during TEM characterization to analyze the elemental distribution on the sample surface. X-ray photoelectron spectroscopy (XPS) was performed using an ESCALAB 250Xi+ spectrometer (Thermo Fisher Scientific, USA) to determine the elemental composition and chemical states of the samples. Nitrogen adsorption-desorption measurements were performed on an ASAP 2020 automatic physisorption analyzer (Micromeritics, USA) to determine the specific surface area and pore-size distribution. The specific surface area was calculated using the Brunauer-Emmett-Teller (BET) method, and the pore size distribution was obtained from the desorption branch based on the Barrett-Joyner-Halenda (BJH) method. Thermogravimetric analysis (TGA) was conducted on an SDT Q600 thermal analyzer (TA Instruments, USA) under a nitrogen atmosphere at a heating rate of 5 °C min^-1 to evaluate the thermal stability of MIL-100(Fe). Electron paramagnetic resonance (EPR) spectroscopy was performed on a Bruker EMXplus-6/1 spectrometer (Bruker, Germany) to analyze oxygen vacancy defects. Raman spectra were collected using a HORIBA LabRAM Odyssey spectrometer (Japan) with 532 nm excitation to probe molecular structure and chemical bonding. UV-vis absorption spectra were measured using a Cary 5,000 UV-vis spectrophotometer (Agilent, USA) in the wavelength range of 300-800 nm to estimate the band gap. Photoluminescence (PL) spectra were obtained using an F-7000 fluorescence spectrophotometer (Hitachi, Japan) to evaluate the defect-related emission characteristics of the materials.

A homogeneous paste was prepared by uniformly dispersing 0.02 g of MIL-100(Fe)-derived γ-Fe₂O₃ sensing material in 15 μL of ethanol, followed by coating the paste onto ceramic tubes (1.2 mm OD (outer diameter), 0.8 mm ID (inner diameter), and 4 mm length). Each tube incorporated gold electrodes at both ends, which were connected to platinum wires. A Ni-Cr alloy heating wire inserted into an alumina tube enabled precise temperature control via regulated voltage application. To enhance stability, all sensors were aged at 240 °C for 48 h in ambient air.

Gas-sensing performance was evaluated using a WS-30B system (Weisheng Instruments Co., Ltd.). The operating temperature was controlled by adjusting the heating wire voltage, while relative humidity (RH) was regulated through controlled water evaporation prior to testing. All measurements were performed under standard conditions (25% ± 5% RH, 23 ± 2 °C) unless otherwise noted.

The sensor resistances in air and target gas are denoted as R_a and R_g, respectively. For n-type semiconductors in a reducing atmosphere, the response (R_es) is defined as R_a/R_g. The response time (T_res) and recovery time (T_rec) are defined as the time to reach 90% of the total resistance change upon gas exposure and air recovery, respectively. Sensitivity is defined as the reciprocal of the slope of the sensor’s calibration curve. The concentration of volatile organic compounds generated by thermal evaporation was calculated using^[21]:

(1) $$ mathrm{C}(\mathrm{ppm})=\frac{\mathrm{V}_{\mathrm{x}}(\mathrm{~mL}) \times \mathrm{P} \times \rho(\mathrm{g} / \mathrm{mL}) \times 22.4(\mathrm{~L} / \mathrm{mol})}{\mathrm{M}(\mathrm{~g} / \mathrm{mol}) \times \mathrm{V}(\mathrm{~L})} \times 10^{6} \\ $$

where C is the gas concentration (ppm), V_X is the liquid volume (mL), P is the purity (fraction), ρ is the density (g/mL), M is the molecular weight (g/mol), and V is the chamber volume (18 L). In practice, the liquid samples were precisely measured and introduced into the evaporation platform using a microsyringe.

The TGA of MIL-100(Fe) reveals a multistage thermal decomposition process, as shown in Supplementary Figure 2. In the first stage (25-138 °C), the material loses 35.76% of its weight due to the removal of adsorbed and crystalline water. This stage does not damage the framework but prepares it for further degradation^[22]. The second stage (138-426 °C) involves a 25.7% weight loss^[23], primarily due to the removal of terminal unbridged anionic ligands and partial degradation of the organic framework, marking the onset of framework breakdown. In the third stage (426-455 °C)^[24], a further weight loss of 13.68% occurs as the framework undergoes more significant decomposition. The fourth stage (455-560 °C) shows a 10.24% weight loss, indicating continued degradation and eventual collapse of the framework as the remaining organic components decompose. Based on the TGA results, further processing was carried out at heat treatment temperatures of 200, 300, 400, 500, and 600 °C.

Phase analysis was conducted using XRD patterns of all samples. As shown in Supplementary Figure 3A, Fe200 and Fe300 showed no characteristic diffraction peaks corresponding to Fe₂O₃. As the temperature increased, the diffraction peaks of Fe400 and Fe500 were indexed to γ-Fe₂O₃ (JCPDS No. 39-1346)^[25], whereas those of Fe600 corresponded to α-Fe₂O₃ (JCPDS No. 33-0664)^[26], indicating a phase transition from a metastable to a stable iron oxide phase. Maghemite (γ-Fe₂O₃) is a metastable polymorph of iron(III) oxide that gradually transforms into the thermodynamically stable hematite (α-Fe₂O₃) phase upon thermal treatment. Under ambient conditions, α- Fe₂O₃ is the most stable Fe₂O₃ polymorph, whereas γ-Fe₂O₃ can transform into α-Fe₂O₃ at elevated temperatures, typically within several hundred degrees Celsius, depending on the synthesis method and particle size^[27,28]. As shown in Supplementary Figure 3B, the sample still retains the γ-Fe₂O₃ phase after introducing H₂BDC, with sharper diffraction peaks. Combined with the data in Supplementary Table 1, this further confirms that the addition of H₂BDC significantly enhances phase purity and crystallinity. Scherrer equation analysis [Supplementary Table 2] shows that the crystallite size increases progressively with higher ligand concentrations. Overall, the XRD results highlight a strong synergistic effect between thermal treatment and chemical modulation: temperature governs phase evolution^[29], while an optimal concentration of H₂BDC enhances crystallinity and promotes crystal growth.

As shown in Figure 1A, the MIL-100 sample thermally treated at 400 °C under an N₂ atmosphere exhibits a composite morphology consisting of stacked octahedra and irregular nanoparticles. This is attributed to the synergistic effect of the “morphological inheritance” and “decomposition-recrystallization” mechanisms during the heat treatment process^[30,31]. As further shown in Supplementary Figure 4A and B, with increasing heat treatment temperature, the framework gradually decomposes at 500 and 600 °C, accompanied by localized structural collapse and significant sintering-induced agglomeration. This structural evolution is consistent with the TGA results. SEM images in Figure 1B and Supplementary Figure 4C-E reveal that as the H₂BDC content increases, the Fe₂O₃ particle size progressively increases, indicating a pronounced size-tuning effect consistent with the XRD results. Meanwhile, the morphology changes accordingly, likely due to additional framework decomposition induced by the higher ligand concentration.

The microstructure and elemental composition of the Fe400 and 15%-Fe400 samples were systematically characterized using TEM. The corresponding magnified TEM images and HRTEM insets [Figure 1C and D] reveal well-defined lattice fringes, with measured interplanar spacings closely matching the characteristic planes of γ-Fe₂O₃, thereby confirming the phase composition of the samples. This result is consistent with the XRD analysis discussed earlier. The selected-area electron diffraction (SAED) patterns [Figure 1E and F] display sharp polycrystalline rings, indicating that both samples possess a typical polycrystalline structure^[32]. The energy-dispersive X-ray spectroscopy (EDS) elemental mapping results [Figure 1G and H] show a uniform spatial distribution of Fe and O in both samples. Notably, the oxygen signal is markedly stronger in the 15%-Fe400 sample.

Figure 1I and J presents the nitrogen adsorption-desorption isotherms and corresponding pore-size distributions of the Fe400 and 15%-Fe400 samples. According to IUPAC classification^[33], both samples exhibit type IV isotherms with H3-type hysteresis loops, confirming the successful formation of a porous structure through thermal treatment of the MIL-100(Fe) precursor. Although such porous structures can provide transport pathways for gas adsorption and diffusion, the BET specific surface areas of Fe400 and 15%-Fe400 are 18.81 and 19.87 m² g^-1, respectively, showing only a slight difference.

According to the EDS results [Figure 2A and B], the Fe:O atomic ratio changes from 1.8:1 for Fe400 to 1.8:2.16 for 15%-Fe400, indicating an increase in oxygen content after modification. This oxygen enrichment is mainly attributed to structural and chemical changes induced by the addition of 15% H₂BDC to the precursor. Specifically, the incorporation of H₂BDC optimizes the coordination environment of the metal centers and promotes the formation of structural defects within the framework. During subsequent heat treatment, these defects evolve into oxygen vacancies, which enhance oxygen adsorption and retention. As a result, the H₂BDC-modified sample shows a significantly higher oxygen content than the unmodified Fe400, highlighting the important role of organic ligand modulation in controlling oxygen distribution during thermal processing.

Figure 2C-G presents the XPS survey spectra as well as the Fe 2p and O 1s spectra of Fe400 and 15%-Fe400, while the corresponding spectra of the other samples are shown in Supplementary Figure 5. The survey spectra [Figure 2C] show clear signals from Fe, O, and C, with no detectable impurities, confirming the high phase purity of γ-Fe₂O₃ and consistent with the XRD results. Gaussian fitting of the Fe 2p spectra [Figure 2D and E] reveals peaks at 710.4 eV and 724.0 eV, assigned to Fe 2p_3/2 and Fe 2p_1/2, respectively, which are characteristic of Fe³⁺ species^[34]. The O 1s spectra [Figure 2F-I] can be deconvoluted into three components corresponding to lattice oxygen (O_L), oxygen vacancies (O_V), and chemisorbed oxygen (O_C). Among all samples, Fe400 and 15%-Fe400 show the highest oxygen-vacancy concentrations, while EDS confirms that 15%-Fe400 has the highest overall oxygen content. These results indicate that introducing H₂BDC into the precursor promotes the formation and stabilization of oxygen vacancies during heat treatment. The increased concentration of oxygen vacancies enhances gas adsorption, facilitates charge transfer, and improves surface reactivity, thereby contributing to the superior gas-sensing performance of 15%-Fe400.

As shown in Figure 2J, the Raman spectra of Fe400 and 15%-Fe400 exhibit distinct differences in vibrational behavior. The 15%-Fe400 sample shows broader and weaker peaks than Fe400, indicating a higher degree of structural disorder and a higher defect density. Such peak broadening and intensity reduction are typically associated with oxygen vacancies and lattice distortions in γ-Fe₂O₃ structures^[35]. The addition of H₂BDC likely modifies the local coordination environment of Fe atoms, promoting oxygen vacancy formation during thermal treatment.

To further confirm these findings, EPR spectroscopy was used to probe unpaired electrons associated with oxygen-deficient sites [Figure 2K and L]. As shown in Figure 2K, the magnetic field-swept EPR spectra of 15%-Fe400 exhibit a significantly stronger signal than Fe400, indicating a higher concentration of paramagnetic centers. In the corresponding g-factor plots [Figure 2L], both samples exhibit resonance peaks at g 2.003, characteristic of oxygen-vacancy-related centers^[36]. The much stronger signal at this g value for 15%-Fe400 further confirms its higher oxygen-vacancy concentration. Together, the EPR and Raman results provide strong evidence that H₂BDC modification promotes the formation and stabilization of oxygen vacancies. This increased vacancy concentration enhances surface reactivity and charge-transfer capability, thereby improving the gas-sensing performance of 15%-Fe400.

To evaluate the gas-sensing performance of MIL-100(Fe)-derived materials, their fundamental sensing properties were systematically studied. As shown in Supplementary Figure 6A, the responses of samples treated at different temperatures toward 100 ppm n-butanol were compared. Fe200 and Fe300 showed negligible responses and were therefore excluded from further analysis. Based on the structural characterization, this weak response is mainly attributed to the incomplete conversion of the MIL-100(Fe) precursor into crystalline Fe₂O₃ at relatively low treatment temperatures. Such insufficient phase transformation leads to a poorly crystallized or partially decomposed structure, which is unfavorable for forming continuous electron-transport pathways within the sensing layer^[37]. In addition, the limited formation of crystalline γ-Fe₂O₃ reduces the number of effective surface sites for oxygen adsorption and activation, thereby weakening the interaction between adsorbed oxygen species and n-butanol molecules^[38]. Consequently, the charge-transfer process is inhibited, leading to an almost undetectable response of Fe200 and Fe300 to n-butanol.

Figure 3A shows that the responses of all sensors to 100 ppm n-butanol follow a characteristic volcano-shaped trend with increasing operating temperature. The operating temperature influences sensing performance by controlling the balance between adsorption, activation, and desorption of gas molecules. At lower temperatures, molecular activation is insufficient, whereas at higher temperatures, desorption dominates^[39]. The results show that the optimal operating temperature for the 15%-Fe400 sensor is 200 °C, which is significantly lower than the 250 °C required for the other sensors. At this temperature, the 15%-Fe400 sensor achieves a response of 142.3, which is more than 2.13 times higher than that of the other sensors. This improved performance is attributed to the higher oxygen-vacancy concentration, well-developed porous structure, and larger specific surface area of the 15%-Fe400 material.

Supplementary Figure 6B and C shows the dynamic response-recovery behavior of the sensors toward 1-100 ppm n-butanol under the same conditions. The sensor response increases with increasing gas concentration and quickly returns to baseline upon exposure to air. As shown in Figure 3B and C, the responses exhibit a strong linear relationship with concentration across the tested range, indicating suitability for quantitative detection over a wide concentration range. Using the theoretical detection limit equation (3×RMSnoise/slope)^[40], the Fe400 sensor shows a detection limit of 325 ppb for n-butanol, which is significantly lower than that of the other samples. Although the 15%-Fe400 sensor shows slightly reduced linear fitting accuracy (R² = 0.981) at higher concentrations due to its stronger response, it achieves a theoretical detection limit of 130 ppb, the lowest among all samples, demonstrating excellent potential for trace-level detection. Table 1 summarizes the linear-fitting equations and corresponding limits of detection (LOD) for different sensors at their optimal operating temperatures. Notably, at 200 °C, the 15%-Fe400 sensor exhibits the lowest detection limit, highlighting its superior sensitivity and quantitative capability.

Figure 3D shows that all sensors rapidly return to baseline after gas switching, with minimal fluctuations and excellent repeatability, indicating stable operation. Selectivity tests [Figure 3E] show that all sensors exhibit good selectivity toward n-butanol, with the 15%-Fe400 sensor showing the best performance. Its response to n-butanol is approximately 12.1, 6.4, 5.6, and 3.6 times higher than those to methanol, ethanol, isopropanol, and triethylamine, respectively, confirming excellent selectivity. In addition, responses to alcohols increase with increasing carbon chain length, attributed to additional -CH₂- groups that promote surface dehydrogenation reactions and enhance sensing performance^[41].

Long-term stability tests [Figure 3F] show that the sensor response variation remains within ±10% over 30 days, confirming excellent durability. All data are presented as mean ± standard error of the mean from 35 independent measurements, and error bars represent the standard error of the mean. Figure 3G compares the response, response time, and recovery time of unmodified heat-treated samples Fe400, Fe500, and Fe600. Among them, Fe400 shows the best overall sensing performance, with fast response and recovery times (T_res = 53 s, T_rec = 48 s) and a high response (R_es = 66.96).

Figure 3H presents the response times and recovery times of sensors prepared with different H₂BDC loadings. Although the 15%-Fe400 sensor achieves the highest response (R_es = 142.3), it exhibits longer response and recovery times (T_res = 92 s, T_rec = 102 s). This is attributed to its higher response magnitude, which requires more extensive gas desorption, and to the relatively low operating temperature, which slows the desorption process^[42]. Detailed response-recovery profiles are shown in Supplementary Figure 6D and E.

To evaluate environmental applicability, the responses of Fe400 and 15%-Fe400 to n-butanol were measured under different RH levels [Figure 3I]. With increasing humidity, both sensors show an approximately linear decrease in response. This is mainly due to competitive adsorption of water molecules on active sites, which occupy oxygen adsorption sites, inhibit oxygen activation, and weaken the charge-transfer reaction between n-butanol chemisorbed oxygen species, thereby reducing the effective sensitivity of the sensing layer^[43]. Table 2 compares the sensing performance of the sensors developed in this work with previously reported materials. The Fe400 sensor shows fast response and recovery, while the 15%-Fe400 sensor exhibits a higher response and lower detection limit, confirming its enhanced sensitivity and improved sensing performance.

Gas sensors operating for extended periods in complex environments often experience performance degradation due to fluctuations in ambient temperature and humidity. Prolonged exposure of the sensing layer to such varying conditions can lead to baseline drift and signal distortion, thereby reducing detection accuracy and long-term stability^[52]. Among these factors, humidity has the greatest impact because adsorbed water molecules compete for active sites and alter charge-transfer pathways, leading to systematic measurement errors.

As shown in Figure 3I, the 15%-Fe400 sensor exhibits clear humidity-dependent deviations, highlighting the need for compensation. To visualize this effect, a ring-stacked chart was used to compare the daily measured concentration (Conc) with the reference value (Ref) obtained from the linear fitting equation [Supplementary Figure 7A]. In this chart, the pink and blue regions represent Conc and Ref, respectively. The degree of overlap between these regions intuitively reflects measurement consistency - greater overlap indicates closer agreement between sensor readings and reference values, implying higher reliability. This visualization effectively captures day-to-day variations in measurement deviation and provides a straightforward way to assess long-term stability.

To correct humidity-induced signal drift, a machine learning-based calibration framework was developed using 35 days of continuous monitoring data collected under varying humidity conditions. The raw response data, along with temperature and relative humidity, were used as model inputs. A total of 35 datasets were randomly split into training and test sets at an 8:2 ratio [Supplementary Table 2]. The 15%-Fe400 sensor, selected for its strong response amplitude, exhibited distinct signal variation, resulting in more reliable calibration results.

This study first applied linear regression (LR) for single-variable calibration^[53]. As shown in Supplementary Figure 7B, LR slightly improved accuracy; however, its correction capability remained limited under complex conditions because it cannot capture the coupled effects of temperature and humidity. To address this limitation, a multidimensional calibration framework was developed by incorporating temperature and humidity as additional input features. Four machine learning algorithms - neural network (NN), decision tree (DT), random forest (RF), and extra trees (ET) - were used for nonlinear modeling. The NN model consisted of two hidden layers with 64 neurons each using rectified linear unit (ReLU) activation, followed by a linear output layer. For tree-based models, the DT model was implemented without a predefined maximum depth; the RF model used 100 trees with unrestricted depth; and the ET model used 50 trees with no maximum depth, a minimum split size of 6 samples, a minimum leaf size of 2 samples, and all features considered at each split.

All four models significantly improved prediction accuracy. Among them, the ET-based model achieved the best performance, with predicted concentrations closely matching reference values [Figure 4A and Supplementary Figure 7C-E]. Cumulative distribution function (CDF) analysis further confirmed that the ET model reached 100% cumulative probability at an error threshold of 0.28, outperforming conventional methods [Figure 4B]. In the target plot [Figure 4C], the ET model point lies closest to the origin, indicating minimal deviation and superior accuracy^[54]. The quantitative comparison [Table 3] further shows that the ET model achieves the lowest mean squared error (MSE) and the highest coefficient of determination (R²), confirming its strong accuracy and generalization for humidity correction.

When applied to other sensors [Supplementary Figure 7F-H], the correction model produced outputs that were highly consistent with the reference values, confirming its strong generalizability and practical applicability. The corresponding raw data are listed in Supplementary Tables 3-5. After correction using the ET model, the R² for the 5%-Fe400, 10%-Fe400, 15%-Fe400, and 20%-Fe400 sensors were 0.9628, 0.9677, 0.9793, and 0.8734, respectively. Among them, the 15%-Fe400 sensor exhibited the highest fitting accuracy, mainly due to its higher oxygen vacancy concentration (as shown in Supplementary Figure 6A). The increased oxygen vacancies provide more active adsorption sites, leading to a more balanced competition between gas molecules and water vapor. Consequently, this sensor shows a more stable response and the smallest calibration error under varying humidity conditions.

Building on the drift correction approach discussed above, additional experiments were conducted to evaluate the sensors’ multi-gas recognition capability. Four sensors - 5%-Fe400, 10%-Fe400, 15%-Fe400, and 20%-Fe400 - were tested for the detection of ten single-component gases and one mixed gas, each at a concentration of 100 ppm [Figure 4D and Supplementary Figure 7I-K]. Each gas was sampled continuously for 200 s at 1 Hz, yielding a total of 2,200 feature datasets. The corrected signals were used as input features, and the dataset was split into training and testing sets in an 8:2 ratio. To further evaluate model robustness, stratified five-fold cross-validation was performed for the best-performing classification model. Four machine learning algorithms, including Backpropagation Neural Network (BPNN), Support Vector Machine (SVM), Random Forest (RF), and CatBoost, were employed for classification and identification. The BPNN contained one hidden layer with 100 neurons and was trained with a maximum of 1,000 iterations. The RF classifier used 100 trees without a predefined maximum depth. The CatBoost classifier was trained using 700 boosting iterations, a learning rate of 0.1, a tree depth of 10, an L2 leaf regularization coefficient of 5, a border count of 64, a bagging temperature of 1, and a random strength of 5. The results showed that the CatBoost model achieved the highest recognition accuracy on the test set (99.55%, Figure 4E) due to its gradient-boosting mechanism and ability to handle complex nonlinear gas-response data^[55].To further evaluate model robustness and check for overfitting, leave-one-out cross-validation (LOOCV) was performed on the optimized Extra Trees model. As shown in Supplementary Figure 8A, the LOOCV root mean square error (RMSE) is 1.89 times that of the training set, indicating that the validation error remains within an acceptable range. Supplementary Figure 8B presents the five-fold cross-validation outcomes of the CatBoost classifier. The results show stable performance across all folds, further confirming that this classification model is not overfitted.

Feature analysis plot [Figure 4F] revealed that the 15%-Fe400 sensor contributed most significantly to gas discrimination. This is attributed to its optimized iron-doping ratio and oxygen vacancy concentration, which not only maintain a high carrier concentration but also create abundant active adsorption sites. As a result, response differences among different gas molecules are amplified, enhancing both sensitivity and discrimination. In summary, machine learning not only improved humidity-drift correction of the gas sensors but also provided high-quality input features for multi-gas selectivity analysis. By combining ET-based calibration with CatBoost-based recognition, this study introduces a “correct-then-identify” strategy that offers a promising approach for developing intelligent, environmentally adaptive gas-sensing systems.

Conventional MOS gas sensors detect target gases by monitoring changes in the sensing material’s resistance during adsorption and desorption. When the sensor is exposed to air, oxygen molecules adsorb on the material surface and capture electrons from the conduction band, forming chemisorbed oxygen species such as O₂^-, O^-, and O^2-(as shown in the equations below)^[56]. The sensors developed in this work exhibit an optimal operating temperature range of 100-300 °C, within which O^- is the dominant adsorbed oxygen species.

(2) $$ \mathrm{O}_{2}(\mathrm{gas}) \rightarrow \mathrm{O}_{2}(\mathrm{ads}) $$

(3) $$ \mathrm{O}_{2}(\mathrm{ads})+\mathrm{e}^{-} \rightarrow \mathrm{O}_{2}^{-}(\mathrm{ads})\left(\mathrm{T} \leq 100^{\circ} \mathrm{C}\right) $$

(4) $$ \mathrm{O}_{2}^{-}(\mathrm{ads})+\mathrm{e}^{-} \rightarrow 2 \mathrm{O}^{-}(\mathrm{ads})\left(100^{\circ} \mathrm{C}<\mathrm{T}<300^{\circ} \mathrm{C}\right) $$

(5) $$ \mathrm{O}^{-}(\mathrm{ads})+\mathrm{e}^{-} \rightarrow \mathrm{O}^{2-}(\mathrm{ads})\left(\mathrm{T} \geq 300^{\circ} \mathrm{C}\right) $$

Unlike conventional MOS sensing materials, whose resistance modulation mainly depends on changes in the surface depletion layer width, the resistance variation in γ-Fe₂O₃ is primarily attributed to a reversible phase transition between a “defective spinel” (metastable state) and a “perfect spinel” (stable state), driven by the Fe³⁺/Fe²⁺ valence change in the lattice. This process enables chemiresistive gas sensing. γ-Fe₂O₃ has a typical defective spinel structure with octahedral Fe vacancies, whereas Fe₃O₄ possesses an inverse spinel structure containing mixed-valence Fe²⁺/Fe³⁺ species. Previous studies have shown that iron oxides can undergo reversible phase evolution under reducing and oxidizing atmospheres. Specifically, γ-Fe₂O₃ can be partially reduced to Fe₃O₄ through the conversion of Fe³⁺ to Fe²⁺, while Fe₃O₄ can be reoxidized to γ-Fe₂O₃ in oxygen-containing atmospheres. For example, Zhang et al.^[57] reported the structural evolution and reversible transformation among α-Fe₂O₃, Fe₃O₄, and γ-Fe₂O₃ nanoparticles under reducing and oxidizing conditions. In addition, Jung and Schimpf^[58] demonstrated that nanocrystalline γ-Fe₂O₃ can be reduced to Fe₃O₄ via partial reduction of Fe³⁺ to Fe²⁺. These studies provide literature support for the reversible γ-Fe₂O₃/Fe₃O₄ transformation proposed in this work.

When n-butanol is introduced, it reacts with surface oxygen species, releasing electrons that reduce part of the Fe³⁺ to Fe²⁺ in the vacancies. This transforms the material into Fe₃O₄ with a perfect spinel structure, forming a mixed-valence chain (-Fe²⁺-Fe³⁺-Fe²⁺-Fe³⁺-) and decreasing the resistance^[11]. Upon re-exposure to air, Fe₃O₄ is reoxidized to γ-Fe₂O₃, increasing resistance (as shown in the equations below). As shown in Figure 5A, XPS analysis of the 15%-Fe400 sample before and after exposure to n-butanol reveals a reversible redox process on the material surface. The shift of the Fe 2p peaks toward lower binding energies indicates partial reduction of Fe³⁺ to Fe²⁺, corresponding to a structural transition from γ-Fe₂O₃ to Fe₃O₄^[59]. The deconvoluted spectra in Figure 5B confirm that Fe³⁺ is the dominant species in the pristine sample, whereas additional peaks attributed to Fe²⁺ appear after n-butanol exposure, as shown in Figure 5C^[23]. More importantly, the emergence of Fe²⁺ components after n-butanol exposure provides direct evidence of mixed-valence Fe²⁺/Fe³⁺ states, a characteristic feature of Fe₃O₄. Therefore, the Fe 2p XPS evolution in Figure 5 confirms the surface redox reaction between n-butanol and γ-Fe₂O₃ and supports the local γ-Fe₂O₃ Fe₃O₄ transformation during sensing. After re-exposure to air, Fe²⁺ is reoxidized to Fe³⁺, corresponding to the reverse Fe₃O₄ γ-Fe₂O₃ process and recovery of sensor resistance.

(6) $$ \mathrm{C}_{4} \mathrm{H}_{8} \mathrm{OH}+12 \mathrm{O}^{-}(\mathrm{ads}) \rightarrow 4 \mathrm{CO}_{2}+5 \mathrm{H}_{2} \mathrm{O}+12 \mathrm{e}^{-} $$

(7) $$ \mathrm{Fe}^{3+}+\mathrm{e}^{-} \rightarrow \mathrm{Fe}^{2+} $$

(8) $$ \mathrm{Fe}^{2+}-\mathrm{e}^{-} \rightarrow \mathrm{Fe}^{3+} $$

(9) $$ H_{2}=2 \gamma-\mathrm{Fe}_{2}^{3+} \mathrm{O}_{3}^{2-} \leftrightarrows \mathrm{Fe}^{2+} \mathrm{O}^{2-} \cdot \mathrm{Fe}_{2}^{3+} \mathrm{O}_{3}^{2-}\left(\mathrm{Fe}_{3} \mathrm{O}_{4}, x=1\right) $$

From a mechanistic perspective, the gas-sensing behavior of MOF-derived γ-Fe₂O₃ arises from the synergistic interactions among oxygen adsorption-desorption, reversible iron valence transitions, and cation-vacancy migration. In γ-Fe₂O₃, cation vacancies at octahedral sites compensate for the positive charge contributed by Fe²⁺ and part of Fe^3+[60]. Under reducing conditions, Fe³⁺ ions are partially reduced to Fe²⁺, promoting structural conversion toward Fe₃O₄. When re-exposed to air, adsorbed oxygen species capture electrons and oxidize Fe²⁺ back to Fe³⁺, enabling the reversible γ-Fe₂O₃ Fe₃O₄ transition. Oxygen vacancies play a dual role in this process, acting as both active sites and regulatory centers. They facilitate the formation of reactive oxygen species and dynamically regulate vacancy distribution throughout the adsorption-reaction cycle. These vacancies originate from coordinatively unsaturated metal sites. An increase in their concentration lowers the phase-transition energy barrier and induces local lattice distortion, thereby accelerating structural transformation and resistance modulation while maintaining overall crystal stability^[61].

To further enhance oxygen-vacancy concentration, the coordination ratio between H₂BDC and H₃BTC was precisely adjusted during the synthesis of the MIL-100(Fe) precursor. Because H₂BDC exhibits a stronger coordination affinity toward Fe³⁺, some carboxyl groups in H₃BTC remain uncoordinated, leading to the formation of localized coordination defects^[62]. During subsequent heat treatment, these defects were converted into oxygen vacancies, significantly increasing their density and surface reactivity^[63]. XPS, Raman, and EPR analyses consistently confirm that the 15%-Fe400 sample exhibits the highest oxygen-vacancy concentration. Therefore, as illustrated in Figure 6, the increased oxygen-vacancy concentration synergistically enhances n-butanol sensing through the coupling of surface adsorption-reaction processes with the reversible γ-Fe₂O₃/Fe₃O₄ redox transition. Specifically, the oxygen-vacancy-rich structure introduces local lattice distortion and perturbs the Fe-O coordination environment, as evidenced by broadened Raman signals and modified O 1s XPS components. Such defect-induced distortion weakens local Fe-O bonding and generates electronically unsaturated Fe sites, making Fe³⁺ more susceptible to electron capture during reaction with n-butanol. Consequently, Fe³⁺ Fe²⁺ reduction is facilitated, promoting mixed-valence Fe²⁺/Fe³⁺ formation and lowering the energy barrier for the local γ-Fe₂O₃ Fe₃O₄ transformation. This transition induces a pronounced change in resistance during gas-solid interaction, enabling efficient conversion of chemical signals into electrical responses^[64].

In addition to facilitating the phase transition, oxygen vacancies also act as highly active adsorption-reaction sites^[65]. They promote oxygen adsorption and activation, accelerate n-butanol oxidation, and enhance electron release back to the sensing layer. These released electrons further drive the Fe³⁺/Fe²⁺ redox process and reinforce the reversible γ-Fe₂O₃ Fe₃O₄ transformation. Therefore, oxygen vacancies enhance sensing performance through two coupled roles: lowering the phase-transition barrier via local lattice distortion and increasing surface reactivity through abundant adsorption-reaction sites. This clear relationship between oxygen-vacancy-induced lattice distortion, facilitated Fe³⁺/Fe²⁺ redox conversion, and enhanced surface kinetics further supports the mechanistic model presented in Figure 6.