A *Drosophila*-inspired intelligent olfactory biomimetic sensing system for gas recognition in complex environments

A bionic gas sensing system mimics biological olfaction to generate distinctive fingerprints for various gases and has advanced significantly since early work by Wilkens and Hatman. Such systems are applied in indoor air quality monitoring, medical care, hazardous gas detection, environmental monitoring, and food quality control. Semiconductor gas sensors are widely adopted owing to high sensitivity, fast response, low cost, and compatibility with micro-electronic processes. Typical sensors comprise a microheater to optimize operating temperature and a gas-sensitive material, often metal oxide semiconductors like SnO2, ZnO, WO3, and TiO2. Despite growing importance, integrating sensor arrays with circuits and algorithms increases size and power consumption, making coordination between modules crucial. Integration complexity and limited fault tolerance remain key challenges. Monolithic and multichip approaches each have trade-offs, with multichip designs often preferred for flexibility and cost. Combining multiple sensors with multivariate data processing enhances selectivity and performance. Artificial neural networks have become popular due to improved computing power and data availability; however, many require substantial training data, which is time-consuming to collect in gas testing. Efficient, accurate recognition with limited data is a challenge. Drosophila olfactory circuitry rapidly generates specific labels for odors, motivating a locality-sensitive hashing-based approach for gas type identification and concentration prediction. This study reports an 18-channel MEMS sensor array integrated with CMOS readout circuits and a Drosophila-inspired algorithm to emulate biological olfactory sensing from acquisition to processing, enabling accurate detection of environmental temperature, humidity, gas type, and concentration to support healthcare, environmental monitoring, and intelligent alarm systems.

The paper situates its work within biomimetic gas sensing and electronic noses, tracing from early e-nose development to modern applications in indoor air quality, medical diagnostics, hazardous gas detection, environmental and food quality monitoring. It reviews semiconductor (MOS) gas sensors for their sensitivity, speed, low cost, and microelectronic compatibility, highlighting microheater roles and common MOS materials (SnO2, ZnO, WO3, TiO2). Integration challenges are discussed: increases in size/power with arrays, circuits, and algorithms; complexity of integration; and limited fault tolerance in designs. Prior system integration strategies include monolithic integrated microheater sensor systems and multichip designs to improve flexibility and reduce cost. On algorithms, the paper notes multivariate processing and ANNs (e.g., multilayer perceptrons, learning vector quantization) for classification, but emphasizes their data-hungry nature and the time-consuming gas testing process, motivating approaches effective with limited data. Inspired by Drosophila’s rapid, label-based olfactory coding, the study adopts locality-sensitive hashing for type identification and concentration prediction as a lightweight, fast alternative.

System overview: An 18-channel MEMS multisensor array (16 gas sensors, 1 temperature sensor, 1 humidity sensor) is integrated with CMOS readout circuits via flip-chip bonding to form a compact device (~5 × 8 mm2, <1 mm thick). The design biomimetically mirrors biological olfactory receptors (sensors), olfactory bulb (CMOS preprocessing), and cortex (neural network algorithm). MEMS gas sensor array: Each gas sensor has a five-layer structure comprising a platinum heating resistor, insulated/isolation layers, interdigital electrodes (Ta/Pt), substrate, and a deep silicon cavity for thermal isolation. Fabrication steps include: (i) SiO2 deposition and dry etching, (ii) sputter/lift-off formation of 30/300 nm Ta/Pt interdigital electrodes, (iii) PECVD of SiO2/SiNx and pad opening, (iv) sputter/lift-off Ta/Pt heater patterning, (v) deep backside Si etch, (vi) TMAH release of suspended membrane, (vii) Au bumping (~50 µm) and inkjet printing of MOS sensing materials, and (viii) flip-chip integration with CMOS. Mechanical robustness is improved via enhanced angle compensation at beam-to-frame connections to mitigate thermal-expansion stress. The 2.5 × 2.5 mm2 array hosts 18 units with a shared ground, with microplate membranes ~150 × 150 µm2 for gas sensors and ~160 × 160 µm2 for the temperature sensor. Sensing materials: Eight MOS nanomaterials are inkjet-printed (two units each): SnO2, ZnO, Au–SnO2, WO3, Pt–SnO2, Fe2O3, TiO2, Pd–SnO2. Nanostructures (spheres/polyhedra) have typical sizes: SnO2 ~20 nm, ZnO 20–30 nm, Au–SnO2 10–20 nm, WO3 30–50 nm, Pt–SnO2 10–20 nm, Fe2O3 20–50 nm, TiO2 150–200 nm, Pd–SnO2 10–20 nm. Sensors operate at 250–400 °C depending on material, with excellent thermal isolation minimizing inter-unit heat disturbance (membrane peak ~367 °C with minimal crosstalk). Environmental sensors: A platinum resistance temperature detector (serpentine, 8 µm line width) on a suspended membrane measures ambient temperature with TCR ~0.00222/°C, linear from ~8 to 100 °C. A humidity sensor with MoO3 coating on a 150 × 150 µm2 membrane measures RH in the 30–70% range. CMOS readout circuitry: The readout comprises a heating circuit to bias microheaters and a resistance-to-voltage conversion chain (reference current source, programmable current source, unity-gain buffer) to acquire sensor responses. Vertical 3D integration is achieved by inverting the MEMS chip and flip-chip bonding to the CMOS chip, avoiding TSVs. Algorithmic pipeline (Drosophila-inspired): - Signal acquisition: 18-channel array collects multi-gas responses along with environmental temperature and humidity. - Preprocessing and reconstruction: A lightweight combination of a shallow neural network and a residual neural network is used to process and reconstruct olfactory data, compensating for environmental perturbations (e.g., high humidity) and potential partial sensor failures. - Odor labeling and hashing: Inspired by Drosophila, distinct labels are assigned to odors. A sparse, binary random matrix increases data dimensionality (expansion), enabling more neurons to represent objects. - Winner-takes-all (WTA): Dimensionality reduction through WTA supports rapid identification and recognition. - Tasks: (1) Qualitative identification of gas types; (2) Quantitative prediction of concentration across 3–5 gradient levels per gas. Experimental characterization: Gas responses are measured under standard conditions (e.g., 24–25 °C, ~40% RH) and under varying conditions (e.g., 12 °C, 65% RH) to assess robustness. Single-channel performance (e.g., SnO2 to 10 ppm acetone) is evaluated for response time and stability. Thermal simulations (COMSOL) assess heating dynamics and power consumption. Long-term stability is monitored over weeks to months.

- High-accuracy recognition: Qualitative identification of 7 gas types achieved 98.5% accuracy. - Quantitative prediction: Concentration level prediction (3–5 gradients per gas) achieved 93.2% accuracy. - Fast response: For 10 ppm acetone, T99 ≈ 11 s to reach steady response and return toward baseline. - Thermal performance: Infrared imaging shows membrane peak temperature ~367 °C with minimal inter-unit heat disturbance. Simulation indicates that with 30 mW power, a sensor reaches ~345 °C within 10 ms. - Environmental sensing: Temperature sensor exhibits TCR ≈ 0.00222/°C, linear from ~8–100 °C. Humidity sensor measures 30–70% RH accurately. - Stability and durability: Heater resistance drifted by ~3% over 100 days. Channel 1 responses to 50 ppm acetone remained stable (0.50 at 30 days; 0.48 at 60 days). Cyclic tests at 25 °C, 40% RH showed consistent multi-channel responses (e.g., to 200 ppm ethanol). - Compact integration: The MEMS array (2.5 × 2.5 mm2) and CMOS readout are integrated into a compact device (~5 × 8 mm2, <1 mm thick). - Robustness to environment: Slight response drifts observed under abnormal conditions (e.g., 12 °C, 65% RH) relative to standard (24 °C, 40% RH), with temperature and humidity channels enabling compensation. - Materials-level detail: Eight MOS materials printed via inkjet yielded nanoscale features (10–200 nm), supporting sensitivity and selectivity across 16 gas channels.

The study addresses the core challenge of achieving accurate, rapid, and low-power multi-gas recognition in complex environments with limited training data. By integrating an 18-channel MOS-based MEMS array with CMOS readout and a Drosophila-inspired algorithmic pipeline (dimensionality expansion via sparse random projections, WTA selection, and lightweight shallow + residual neural networks), the system emulates biological olfaction from sensing through processing to decision. This design enhances selectivity and fault tolerance, enabling reliable classification (98.5% for 7 gases) and quantification (93.2% across 3–5 concentration levels) despite environmental variability (humidity/temperature fluctuations) and potential partial sensor degradation. The inclusion of dedicated temperature and humidity sensors allows environmental compensation, further stabilizing gas responses. The compact integration (flip-chip 3D stacking without TSVs) demonstrates a practical path toward deployable intelligent olfactory systems. Collectively, the results support the feasibility of electronic noses for applications in emergency rescue alarms, healthcare, and environmental monitoring, where fast and accurate detection under non-ideal conditions is critical.

The work presents a compact, integrated electronic olfactory system comprising an 18-channel MEMS sensor array, CMOS readout circuits, and a Drosophila-inspired lightweight machine-learning algorithm. The system delivers high-accuracy qualitative (98.5%) and quantitative (93.2%) gas analysis, fast response, environmental sensing/compensation, and long-term stability, demonstrating suitability for real-world applications such as emergency alarms, healthcare, and environmental monitoring. Future research could expand the set of detectable gases, extend environmental operating ranges (e.g., wider temperature/humidity), enhance fault tolerance to sensor degradation, reduce power further, and validate performance in more complex gas mixtures and real-world field deployments.

- Scope of evaluation: Demonstrations focused on 7 gas types and concentration gradients (3–5 per gas); generalization to a broader spectrum of gases and complex mixtures remains to be validated. - Environmental range: Experimental validations emphasize moderate conditions (e.g., ~24–25 °C, ~40% RH) with limited tests at 12 °C and 65% RH; performance under wider temperature/humidity extremes is not fully characterized. - Humidity sensing range: Reported RH measurement spans 30–70%, potentially limiting compensation accuracy outside this range. - Data constraints: While the algorithm targets limited data scenarios, gas datasets are inherently time-consuming to collect; robustness to dataset shifts and long-term drift beyond 100 days requires further study. - Material/system dependencies: Performance depends on specific MOS materials and operating temperatures (250–400 °C); cross-sensitivity and long-term catalyst poisoning in harsh environments were not exhaustively assessed. - Partial damage tolerance: The algorithm is designed to handle partially damaged units, but systematic fault-injection and recovery evaluations are not detailed in the provided text.