Table 1 Summary of AI-enhanced gas sensors

PSO-SVM

CO and CO₂

Highest predictive accuracy for single/dual gas inversion; suppressed cavity-mode noise; reduced LoD to 8.2 ppmv (CO) and 13.2/4.7 ppmv (CO₂/CO)

Ref. ¹⁸²

Machine Learning Classifier (MLC)

Common acyclic hydrocarbons

Enabled handheld FADA microspectrometer to classify hydrocarbons down to 75 ppm; single-gas LoD ~32 ppm; fast, label-free gas identification

Ref. ²⁶

1D Convolutional Neural Network

H₂O, CO₂, O₂, N₂O, CO, CH₂, NO, SO₂, NO₂, NH₃

Automated feature extraction from IR spectra; speciation accuracy 82–97% on multi-component mixtures; replaces manual spectral interpretation

Ref. ²⁹⁰

Convolutional Neural Network

Benzene, Toluene, Ethylbenzene, Xylene

Precise simultaneous concentration predictions; R² > 0.99 for toluene/ethylbenzene/o-xylene and >0.96 for benzene; robust against low-PEL species

Ref. ²⁹¹

TSMC-Net (Deep Convolutional Neural Network)

Eight volatile organic compounds (VOCs)

High precision, recall, and accuracy for multigas classification from THz absorption; interpretable via class activation maps; portable THz sensor ready

Ref. ²⁹²

Support Vector Machine with Selectivity Factor Analysis

Mixed gases via metal-oxide sensor arrays

Established a direct proportionality between a sensor’s selectivity factor and concentration-prediction accuracy; showed that combining sensors with complementary selectivity profiles can significantly boost prediction performance

Ref. ²⁹³

Deep Neural Network

Hydrogen (H₂) at concentrations below the conventional limit of detection

Extracted “hidden” sensing signals buried in noise, enhancing detection of H₂ below the traditional LOD; demonstrated universality across different sensor materials without modifying the sensors themselves

Ref. ²⁹⁴

Random Forest, ANN, k-Nearest Neighbors

Natural gas mixtures (CH₄, two NG simulants, CH₄ + NH₃)

Achieved >98% identification accuracy across four gas classes; optimized model complexity to avoid overfitting; demonstrated sub-10 ms training and <0.1 ms inference on single-board hardware, enabling real-time IoT deployment

Ref. ²⁹⁵

Grey-Box SVM (physical + ML hybrid)

NO_x and NH₃ in combustion exhaust

Predicted sensor outputs (NO_x cell current) over wide operating conditions and ammonia cross-sensitivities with high accuracy; leverages physical insight to choose only nine key features, reducing overfitting and computational load

Ref. ²⁹⁶

Boosted Regression Trees, Boosted Linear Regression, Gaussian Process

Ambient NO₂ and O_x (O₃ + NO₂) using clustered low-cost electrochemical sensors

Combined cluster-median signals from six sensors with ML to suppress inter-sensor drift and environmental variability; ML models outperformed linear regression, producing NO₂/O_x estimates comparable in RMSE to reference monitors while consuming <200 mW

Ref. ²⁹⁷

Unsupervised Neural Network with Physics-Informed Augmentation

Five gas species over 2900–3100 cm^-1

Overcame scarce data and baseline-drift issues; achieved simultaneous identification, concentration retrieval, and pressure prediction for 31 mixtures with sub-ppb sensitivity

Ref. ²⁹⁸

Machine-Learning-Driven Wavelength Selection + Dual ANN Design

Various gas mixtures via micro-resonator arrays

Automatically selected optimal probe wavelengths based on gas absorption fingerprints; two ANNs then map wavelengths to resonator geometries, enabling compact, tunable sensor modules with precise multi-gas detection

Ref. ²⁸

Stepwise Multilayer Perceptron (SMLP)

CO₂ and CH₄ in NDIR multipass optical measurement

Iteratively performs mixture classification then selective regression, automating feature extraction; yields 98.21% classification accuracy and normalized RMSE of 0.42%/0.45% with only 0.5 s of data and fewer training samples

Ref. ²⁹⁹

Deep Belief Network & Convolutional Neural Network

Toxic gas dispersion from point source emissions

Modeled complex dispersion fields with DBN and CNN to predict concentration contours faster and more accurately than Gaussian plume, CFD, and traditional ML models; demonstrated improved prediction accuracy for emergency response planning

Ref. ³⁰⁰

3D Convolutional Neural Network (LCNet)

O₃ and Cl₂ in gas mixtures via liquid-crystal film optical responses

Analyzed spatiotemporal color patterns of LC anchoring transitions to simultaneously identify O₃ and Cl₂ and quantify their concentrations; revealed that O₃ is detected via transition timing and Cl₂ via late-stage color fluctuations

Ref. ³⁰¹

Deep-learning neural separator

CO and CH₄

Resolves ultra-high spectral overlap via simulated training; achieves R² of 0.9996 (CH₄) and 0.9930 (CO); real-time LoD of 120.9 ppm (CH₄) and 0.5 ppm (CO); low system complexity and simultaneous detection

Ref. ³⁰²

VOC-Net (1D CNN)

Volatile organic compounds (VOCs)

Automates classification with >99% accuracy on simulations and 97% on noisy experimental data; provides interpretable Grad-CAM explanations

Ref. ³⁰³

1D CNN & deep MLP

Methane and acetylene

End-to-end concentration retrieval from direct absorption spectra; surpasses wavelength modulation spectroscopy in precision; robust to noise, laser aging, and circuit variations

Ref. ³⁰⁴

PCA, neural networks, regression

Humidity (water vapor)

Learns leaky-mode ATR reflectance dips to predict relative humidity with 0.3% accuracy using limited data; ML matches or exceeds physical-model fitting

Ref. ³⁰⁵

ML framework (multi-task + CNN)

Mixed-gas (ethylene, CO, CH₄)

Combines tailored pre-processing, multi-task learning, and CNN architecture to significantly boost mixed-gas concentration prediction on UCI dataset compared to prior methods

Ref. ³⁰⁶

PCA, ANN, DNN, 1D CNN

Indoor VOCs: benzene, xylene, toluene, formaldehyde, ethanol

Discriminates five indoor pollutants under varying humidity and temperature; DNNs on full transients offer best performance; can reduce array to two sensors without loss of accuracy

Ref. ³⁰⁷

Temporal-based SVM with moving-window decision logic

CO, O₃, NO₂

Achieves 100% accuracy in both training and testing for multi-pollutant mixtures; allows user-tunable confidence thresholds to control false alarms and detection confidence

Ref. ³⁰⁸

Convolutional Neural Network

CO, NH₃, NO₂, CH₄, acetone

Enables real-time gas identification with response times 1–19 s and 98% accuracy using a batch-uniform SMO sensor array; overcomes array non-uniformity

Ref. ³⁰⁹

Early-fusion multimodal AI (LSTM + CNN)

Four gas classes (via semiconductor array + thermal imagery)

Fuses gas-sensor time series (LSTM) and thermal images (CNN) to achieve 96% identification accuracy, outperforming sensor-only (82%) and vision-only (93%) models

Ref. ³¹⁰

Recurrent Neural Network (RNN) + neuromorphic hardware synapses

NO₂ and H₂S

Processes sequential FET-sensor transients with 1.94% error in high-level tests; hardware RNN system runs at 0.412 mW with 0.3% overall error, ensuring low-power, reliable gas classification

Ref. ³¹¹