Data-centric artificial olfactory system based on the eigengraph

The study addresses the stagnation in artificial olfaction caused by sensitivity-centric sensor and data-analysis paradigms. Traditional electronic-nose research emphasizes maximizing sensitivity and uses magnitude-based pattern recognition (e.g., PCA/LDA) for selectivity, which is unreliable under varying environments and for complex gas mixtures. The authors hypothesize that time-domain waveform information (the eigengraph) produced by redox reactions between gases and sensing materials contains robust, intrinsic odor attributes. They aim to formalize the eigengraph concept, engineer sensors that reproducibly generate refined response waveforms, and extract discriminative frequency-domain features (MFCC) for deep-learning-based gas identification, thereby enabling standardized, data-centric artificial olfactory systems.

Foundational work on artificial olfactory systems dates back to metal-oxide gas sensors (Seiyama et al., 1962) and the electronic nose concept (Persaud & Dodd, 1982). Subsequent advances in chemiresistive sensors and analysis frameworks, bolstered by nanotechnology and AI, have progressed the field. However, selectivity has largely been pursued through sensitivity optimization and magnitude-based features (e.g., PCA/LDA, fingerprint charts, contour maps), which are susceptible to environmental variability, diffusion effects, and mixture interactions. Prior studies established receptor/transducer functions and power-law responses in metal-oxide sensors, but typically neglected temporal waveform structures. The authors identify this gap and propose leveraging intrinsic waveform attributes (eigengraphs) via frequency-domain feature extraction and deep learning to overcome cross-sensitivity and improve reproducibility.

Sensor design and fabrication: The team fabricated an olfactory receptor-like sensor array (ORSA) inspired by human olfactory receptors. A wafer-scale, semiconductor-compatible process was used to produce 3×3 arrays of chemiresistive sensors on 4-inch SiO2/Si wafers with Au/Ti interdigitated electrodes. Vertically aligned porous SnO2 nanorods were deposited via electron-beam evaporation with glancing angle deposition (GLAD). Bimetallic nanolayers—transition metals (Co, Ni, Cu) and noble metals (Pt, Pd, Au)—were sequentially deposited with spatially addressable shadow masks to define channels and then annealed (550 °C, 2 h) to form p-type transition-metal oxides (Co3O4, NiO, CuO) and noble metal nanoparticles on SnO2 nanorods, producing catalytic nanojunctions (p–n heterojunctions and spillover-active noble metals). Optimization of catalytic layers: Deposition thicknesses of each metallic nanolayer were varied (0.5, 1, 2 nm). Time-series sensing waveforms and baseline resistances were analyzed across nine channel types (e.g., CuO/Pt, NiO/Pd, Co3O4/Au). Results showed unique reaction waveforms per catalyst combination; 2 nm layers often caused excessively high/unstable resistance (e.g., CuO-based channels), and in one case (2 nm Co3O4/Au) exhibited p-type-like behavior to reducing gases. 0.5 nm layers yielded stable signals but limited waveform diversification and performance. The optimal specification was 1 nm per metallic nanolayer. Baseline resistance order was Pt > Pd > Au across thicknesses. Morphological analysis indicated Au-driven agglomeration reduced contact area/interface length, affecting electron transfer; XPS confirmed electron transfer directions (Sn 3d5/2 shifts). Measurement system: A dedicated measurement/monitoring setup minimized humidity, temperature, and vibration variations. Sensors were biased at 1 V; gas concentrations were controlled with MFCs under a total flow of 1000 sccm. Sensors were aged at 300 °C in dry air; target gases were introduced after baseline stabilization. Signals were sampled via switching matrices and sourcemeters; resistance signals were normalized to sensitivity. Preliminary selectivity assessment: Sensitivity S was computed as [(R_air−R_gas)/R_gas]×100% (equivalently (R_air/R_gas−1)×100%). For 10 ppm VOCs (acetone, toluene, xylene, ethanol, H2S, NH3) at 300 °C, the ORSA outperformed bare SnO2 nanorods; e.g., SnO2–CuO/Au improved ethanol and NH3 sensitivities by 27.3× and 37.15× vs. bare SnO2. PCA on sensitivity showed ethanol, NH3, H2S separable, but acetone, toluene, xylene overlapped, motivating waveform-based analysis. Eigengraph formalization: Based on receptor and transducer functions and gas–material interactions (including power-law behavior), the authors define eigengraphs as intrinsic, reproducible time-series waveforms arising from specific gas–material redox interactions. They emphasize creating standardized sensors and stable measurement environments to generate unique, reproducible eigengraphs across gases and materials. Feature engineering (FFT/MFCC): Time-series signals were converted to frequency-domain via FFT, generating power spectra. To reduce dimensionality while preserving perceptually salient attributes, Mel-frequency cepstral coefficients (MFCC) were extracted using librosa. For VOC database creation: each original signal (6840 points) was segmented into 20 response windows of 342 points; sampling rates and FFT sizes were set per Nyquist and power-of-two constraints (e.g., 50/171/342 points samples with sampling rates 100/400/800 and FFT sizes 64/256/512, zero-padded as needed). Mel filter counts scaled with sample size (e.g., 32/64/128), and the lowest 20 MFCCs were used per segment. A total dataset (reported) of 9 channels × 13 gas types × 20 segments = 2370 samples was constructed with 117 labeled classes (gas species and mixing ratios). Deep learning: A DNN with three hidden layers (ReLU activations) and Softmax output (117 classes) ingested 20-D MFCC vectors. Training sample sizes were varied (full 342 points, top 171 points, and top 50 points capturing the pure response onset). Learning rates of 1e−4 and 1e−5 were tested. Training accuracy targets were 99.9% and 99.5%, with convergence epochs/times recorded. Euclidean distances quantified intra-/inter-channel reproducibility. For empirical exhaust-gas experiments, MFCC features were expanded (three 470-point sections per 1410-point signal) to 60-D (DNN) or transformed to 16×16×3 inputs (CNN, CNN-LSTM). Fourfold cross-validation was performed across four simultaneously fabricated ORSA chips. Empirical targets: Two standard exhaust gases (gasoline: CO, CO2, C3H8; diesel: CO, CO2, C3H8, NO, O2) and four components (CO, CO2, NO, NO2) were tested; mixtures (CO+CO2; CO+CO2+NO) were also evaluated. Signals displayed distinct eigengraphs per gas class. Cross-validation assessed DNN, CNN, and CNN-LSTM performance with learning rate 1e−5 and MFCC configurations as summarized in the study.

• Optimized sensing stack: 1 nm bimetallic nanolayers (transition metal underlayer and noble metal overlayer) on GLAD SnO2 nanorods yielded stable, diversified, and reproducible eigengraphs. Thicker (2 nm) layers produced unstable/high-resistance signals; thinner (0.5 nm) layers reduced diversification/performance. Baseline resistance followed Pt > Pd > Au; XPS confirmed electron transfer consistent with morphology-dependent nanojunctions.
• Sensitivity gains: Compared to bare SnO2, ORSA channels exhibited markedly higher sensitivities. Notably, SnO2–CuO/Au sensitivity increased by 27.3× (ethanol, 10 ppm) and 37.15× (NH3, 10 ppm) over bare SnO2. PCA on sensitivity separated ethanol, NH3, H2S but could not distinguish acetone/toluene/xylene, underscoring the need for waveform features.
• Eigengraph-based deep learning (117 classes): Using MFCC features from eigengraphs, a DNN achieved complete convergence (training accuracy 99.9% or 99.5%, loss ~0) across training sample sizes (50, 171, 342 points) and learning rates (1e−4, 1e−5). Reducing the training signal to the 50-point pure-response window decreased training time by 3–4× while maintaining 99.9% accuracy, contrary to typical performance degradation with smaller samples, attributed to standardized, invariant waveform quality.
• Reproducibility: Euclidean-distance analyses showed minimal distances along intra-channel comparisons vs. inter-channel, demonstrating stable, reproducible eigengraph generation for given gas–material pairs.
• Empirical exhaust-gas classification: For 6 classes (2 exhausts + 4 components), fourfold cross-validation achieved average accuracies: DNN 96.1%, CNN 99.8%, CNN-LSTM 100%. For 4 mixture classes (CO+CO2; CO+CO2+NO; gasoline; diesel), average accuracies were: DNN 99.3%, CNN 99.7%, CNN-LSTM 99.9%. Training times were markedly shorter for CNN/CNN-LSTM vs. DNN. Distinct eigengraph waveforms were visually separable, even for complex mixtures.

The findings validate a data-centric paradigm for artificial olfaction that prioritizes intrinsic waveform attributes over magnitude-only metrics. By engineering standardized, reproducible eigengraphs through controlled nanostructures and measurement environments, and by encoding these waveforms via MFCC, the approach overcomes cross-sensitivity and mixture-induced distortions that confound sensitivity-based methods. Deep learning models trained on MFCC features achieve near-perfect classification, even with drastically reduced sample lengths, highlighting that robust feature engineering can reduce computational burden without sacrificing accuracy. The empirical success on complex exhaust mixtures demonstrates real-world applicability and suggests that eigengraph-based analysis can generalize to other electrochemical sensing domains.

This work introduces and formalizes the eigengraph concept in electrochemistry and implements a standardized, data-centric artificial olfactory system. A wafer-scale GLAD-fabricated ORSA with combinatorial noble-metal/transition-metal-oxide nanojunctions generates stable, unique waveforms per gas–material interaction. MFCC features distilled from these waveforms enable highly accurate deep learning classification across single gases, mixtures, and real exhaust gases, with reduced training times. The approach provides a blueprint for standardized artificial olfaction and may extend to broader electrochemical analyses and applications (e.g., appliances, robotics, environmental monitoring, defense, forensics). Future work could establish international standards for eigengraph acquisition, expand gas libraries under varied environmental conditions, explore transfer learning across sensor batches, and integrate low-power on-sensor signal processing for field deployment.

The study emphasizes optimized fabrication and controlled measurement conditions (dry air, fixed temperature ~300 °C, minimized humidity/vibration), which may differ from unconstrained real-world environments; standardized hardware and protocols are required to maintain eigengraph stability. Gas sets, concentrations, and mixtures, while diverse (VOCs and exhaust gases), cover a limited chemical space; broader validation across more gases, humidity/temperature variations, and sensor aging would strengthen generalizability. The approach currently relies on high-temperature operation and dedicated instrumentation, which may impact portability and energy use. The dataset construction and class labeling are specific to the fabricated ORSA and experimental design, potentially limiting direct transfer to other sensor types without recalibration or retraining.