Sensor arrays, e-noses & machine-learning selectivity

Summary

No cheap H2 sensor covered elsewhere in this wiki — MOX chemiresistor, catalytic pellistor, palladium-resistive film, electrochemical cell — is truly specific to hydrogen; all of them respond, to varying degrees, to carbon monoxide, methane, hydrogen sulfide, humidity, and temperature. The “sensor array” or “electronic nose” approach works with this cross-sensitivity rather than trying to engineer it away: wire up several sensors whose blind spots differ, log everything simultaneously, and let statistics or machine learning do the discrimination no single sensor manages alone — the same approach biology uses, building a nose from hundreds of broadly-tuned, overlapping smell receptors rather than one receptor per odor [1][2].

Done well, this can turn a set of cheap, drift-prone parts into a useful screening tool, and recent work in chemometrics and deep learning is cutting the cost of keeping such an array calibrated over months of unattended field time — in the best documented case, by more than a hundredfold [13]. One limitation matters for exploration work: essentially nothing in this literature was built for, or tested on, natural-hydrogen soil gas. Every quantitative result here comes from indoor air quality, food science, or breath diagnostics, transplanted by analogy.

The underlying math also has a hard limit no software can cross: if every sensor in an array responds to CO and H2 in the same proportion, no classifier, however sophisticated, can separate the two from that array’s data alone. For Hestia, treat this as a legitimate but immature way to make a dense grid of cheap sensors more trustworthy, not a substitute for periodic ground-truth checks against a reference-grade instrument.

Mechanism & physics

Why arrays exist, and the response-space formalism

The foundational demonstration is Persaud and Dodd’s 1982 “model nose,” which used four semiconducting gas sensors with deliberately overlapping, non-specific selectivity — modeled on mammalian olfactory receptor coding — to show that a battery of broadly-tuned sensors plus pattern recognition could discriminate odorants without any single highly specific receptor [1]. Gardner and Bartlett coined “electronic nose” in 1994 as an instrument comprising “an array of electronic chemical sensors with partial specificity and an appropriate pattern-recognition system” [2].

This matters for Hestia because cheap, imperfect sensors, in sufficient number and diversity, plus the right computation, can in principle do a job no single cheap sensor can do alone.

Treat an array of $n$ sensors as a response vector $\mathbf{r}(t) = \mathbf{F}\big(\mathbf{c}(t), \mathbf{x}(t)\big) + \boldsymbol{\epsilon}(t)$, where $\mathbf{c} = [c_{H_2}, c_{CH_4}, c_{CO}, c_{H_2S}, c_{He}, \ldots]^T$ is the gas-concentration vector, $\mathbf{x} = [T, RH, P, \ldots]^T$ the environmental-covariate vector (see Environmental covariance and Helium & co-tracers), $\mathbf{F}$ the array’s nonlinear transfer function, and $\boldsymbol{\epsilon}$ measurement noise. The e-nose enterprise is the attempt to learn, from training data, an approximate inverse $\hat{\mathbf{c}} = \mathbf{G}(\mathbf{r})$. Linearizing $\mathbf{F}$ gives the sensitivity (Jacobian) matrix $S_{ij} = \partial r_i/\partial c_j$, so $\Delta\mathbf{r} \approx \mathbf{S}\,\Delta\mathbf{c}$, and if $\mathbf{S}$ is well-conditioned the pseudo-inverse recovers concentration changes:

$\Delta\hat{\mathbf{c}} = \mathbf{S}^{+}\Delta\mathbf{r} = (\mathbf{S}^T\mathbf{S})^{-1}\mathbf{S}^T\Delta\mathbf{r}$

Inversion quality is governed by the condition number $\kappa(\mathbf{S}) = \sigma_{\max}/\sigma_{\min}$, which makes the problem of correlated sensors precise: if every sensor responds to H2, CO, and CH4 in nearly the same relative proportions (a bank of undoped SnO2 MOX elements; see MOX chemiresistors), the rows of $\mathbf{S}$ are nearly parallel, $\kappa(\mathbf{S}) \to \infty$, and small sensor noise is amplified into large, unstable concentration estimates.

Adding more of the same kind of sensor adds redundant rows, not new information; what improves conditioning is genuine diversity — different dopants, operating temperatures, or entirely different transduction principles (chemiresistive + electrochemical + thermal-conductivity), each pointing $\mathbf{S}$‘s rows in a genuinely different direction in gas-space.

Feature extraction, dimensionality reduction, and classification

Raw output is rarely classified directly. Steady-state features ($\Delta R/R_0$, $\log(R/R_0)$, plateau value) and transient features (derivatives, rise/fall time constants, area under the transient, wavelet/Fourier coefficients) are the two standard families; transient features are often more discriminative but inherit more short-term noise [3][4].

PCA eigen-decomposes the feature covariance, $\boldsymbol{\Sigma}\mathbf{v}_k = \lambda_k \mathbf{v}_k$, and projects onto the top-$k$ eigenvectors [3] — but it maximizes total variance with no knowledge of which direction is gas identity versus drift or humidity. If humidity-driven variance exceeds H2-driven variance, as documented for MOX at Hestia-relevant concentrations, the first principal component preferentially tracks humidity, and a naive PCA-then-cluster pipeline “discriminates” humidity states, not gas identity.

LDA corrects this using labels, maximizing between-class over within-class scatter, $\mathbf{w}^{*} = \arg\max_{\mathbf{w}} \frac{\mathbf{w}^T \mathbf{S}_B \mathbf{w}}{\mathbf{w}^T \mathbf{S}_W \mathbf{w}}$ [3][4] — more discriminating, but it needs labeled “true anomaly vs. artifact” examples, often the very thing under investigation in a field program.

Four method families cover quantification and classification:

• For quantification, Partial Least Squares (PLS) regression decomposes $\mathbf{r} = \mathbf{TP}^T + \mathbf{E}$, $y = \mathbf{TQ}^T + f$, built for the collinear-predictor, small-sample regime typical of e-nose calibration [3][4].
• For classification, Support Vector Machines find a maximum-margin boundary, $f(\mathbf{x}) = \mathrm{sign}\left(\sum_i \alpha_i y_i K(\mathbf{x}_i, \mathbf{x}) + b\right)$, achieving “high classification accuracy in small sample sizes” but degrading once feature count exceeds sample count [3], the regime a temperature-cycled sensor can create if raw time points go unreduced.
• Random Forest (bootstrap-aggregated decision trees) is comparatively robust to noisy, correlated, nonlinearly-interacting features and gives feature-importance rankings; it is the model of choice in one relevant co-evolved-gas field study, where a low-cost array (XPOD) colocated with a reference-grade methane analyzer achieved $r = 0.719$, RMSD $= 0.3673$ ppm, outperforming multilinear regression and a neural network that could not resolve individual spikes [26].
• Neural networks, up to 1-D CNN/CNN-LSTM hybrids, learn feature representations automatically; one review reports a 1-D CNN outperforming SVM/MLP/KNN/Random Forest “by around 10% on average” [4]. This is a real gain, but it demands more data and carries a risk: spurious temporal structure in training data (batches of different gases recorded in different weeks) lets a flexible model learn timing instead of chemistry, inflating benchmark accuracy (Vergara/Dennler discussion below).

Temperature-modulated operation, resolving power, and dimensionality

A single MOX chemiresistor’s Arrhenius-activated surface chemistry is strongly temperature-dependent, and different gases have different optimal reaction temperatures. Temperature-modulated/cycled operation (TMO/TCO) drives the heater through a periodic program, e.g. $T(t) = T_0 + \Delta T \sin(2\pi t/\tau)$, so within one cycle the sensor passes through many effective operating points, each with a different selectivity profile; sampling $R(t)$ at $k$ points produces a $k$-dimensional feature vector from one physical sensor — a virtual array at zero added parts count, though not zero firmware complexity or heater thermal-cycling stress.

This underlies the on-chip “AI” claims of the Bosch BME688 and Renesas ZMOD families (Vendors, below).

A more rigorous alternative to eyeballing a PCA plot asks, in Shannon’s sense, how many distinguishable gas states an array can resolve given its noise floor. A 2018 method estimates resolving power (RP) as the ratio of the response-space hypervolume spanned by sensor signals to the hypervolume occupied by noise, validated on temperature-modulated MOX sensors exposed to CO–humidity mixtures [5] — sensors occupying the same subspace add overfitting risk, not resolving power.

A less formal version of this question drove the 1990s Caltech carbon-black/polymer-composite array program, an early proof of the concept outside the MOX materials family, which also produced some of the first array-specific drift-correction comparisons [6][7].

More generally: $n$ sensors × $k$ TMO time points × several derived statistics per channel easily yields $p = 50$–$500$ feature dimensions from a handful of physical sensors, while field calibration campaigns are typically data-poor (a few hundred exposures over days to weeks). This curse of dimensionality is a consistent methodological thread across the reviews surveyed here [3][4], and motivates aggressive dimensionality reduction, regularized/ensemble methods over unconstrained deep nets, and the calibration-transfer program below.

Drift as a stochastic process, and calibration transfer

Drift is best modeled as a slowly time-varying process superimposed on true signal, $r(t) = r_0(\mathbf{c}) + d(t) + \eta(t)$, where $d(t)$ evolves over weeks–months (aging, poisoning, thermal annealing — see MOX chemiresistors) and $\eta(t)$ over seconds. Short-term repeatability can look excellent (CV 0.38–1.21% in one one-year, 62-sensor field dataset) while the same sensors’ baseline wanders 25.7–40.8% CV over the full deployment [15]; a spot-check and a full deployment therefore measure different things.

A model trained in “batch 1” must generalize to “batch k” months later, whose feature distribution has shifted: domain shift, which more batch-1 training data cannot fix. The chemometric responses, in increasing sophistication:

• Component/Orthogonal Signal Correction (CC/OSC) — estimate a common drift direction $\hat{\mathbf{d}}$ and project it out, $\mathbf{r}_{\text{corrected}} = \mathbf{r} - (\mathbf{r}\cdot\hat{\mathbf{d}})\hat{\mathbf{d}}$, OSC additionally constraining the removed direction to be uncorrelated with target $y$ [8][9]. Its key assumption — drift affects every analyte class the same way — is often false in practice.
• Direct/Piecewise Standardization (DS/PDS) — learn a linear transfer $\mathbf{r}_{\text{master}} \approx \mathbf{F}\,\mathbf{r}_{\text{slave}} + \mathbf{b}$ from a small set of paired transfer-standard measurements; a five-unit twin-array comparison found DS cut transfer samples needed relative to full recalibration [12].
• Deep transfer learning — fine-tune a network pretrained on a large source dataset with a small target sample. A controlled comparison against classical DS/PDS/OSC for MOX sensors found transfer learning cut new calibration samples required by up to 99.3% (~10 days of exposure down to ~2 hours) while holding RMSE around 18 ppb acetone [13]. This is one of the more Hestia-relevant numbers here: the maintenance burden of a field array, not just its accuracy, is a controllable engineering variable.
• Unsupervised domain adaptation — align source and target feature distributions (e.g., Maximum Mean Discrepancy) without new target labels. A 2024 attention-based multi-source framework reported 83.20% accuracy on a 36-month benchmark and 93.96% on a self-collected 30-month dataset [14] — valuable where “the true H2 concentration right now” is exactly the unknown.

Variants & subtypes

• Homogeneous chemiresistor arrays. Multiple MOX or conducting-polymer elements differing in dopant, catalyst, or base oxide, co-packaged and read together (including the classic carbon-black/polymer-composite family [6][7]). Diversity comes from materials chemistry, not computation.
• Heterogeneous multi-modality arrays. Genuinely different transduction principles combined (MOX + electrochemical + thermal-conductivity + optical) — the structurally strongest way to improve $\kappa(\mathbf{S})$, because different physics rarely shares the same cross-sensitivity pattern. This is the architecture implicitly recommended for Hestia field stations below.
• Covariate-augmented arrays. A gas-array plus dedicated T/RH/pressure/O2 sensors, where the covariate channels feed the same regression model as regressors to explain away environmental confounds — the computational backbone of every practical field e-nose deployment and of the environmental-covariance framework generally.
• Virtual (single-sensor, temporally diverse) arrays. TMO/TCO exploitation of one physical chemiresistor (above) — the cheapest way to buy array-like selectivity with zero added parts count.
• Photonic/biochemical cross-reactive arrays. A structurally different implementation of the same principle: Aryballe’s NeOse platform grafts peptide receptors of differing affinity onto a silicon-photonic Mach-Zehnder network, converting exposure into a spatial interference “olfactive fingerprint” matched by machine learning [24][25]. Not currently marketed or characterized for H2, but architecturally the same array-plus-ML paradigm — see also Exotic optical.
• Colorimetric/chemo-optical arrays. Dye/pigment spots whose color change is imaged and pattern-matched — covered more fully under UV/EUV & fluorescence and Plasmonic & metamaterial sensors; noted here only for taxonomic completeness.
• TinyML / on-device (edge) inference. A cross-cutting deployment pattern, not a distinct architecture: once trained, a model can be quantized and run directly on a microcontroller co-located with the array instead of streaming raw data off-site. A lightweight CNN built for this constraint, applied to spatiotemporal “pseudo-image” representations of e-nose transients, targets handheld/microcontroller-scale odor recognition [31]; commercial MLOps tooling (Edge Impulse) has been used to deploy gas-classification models to microcontroller hardware for industrial gas/VOC identification [32], and a dedicated TinyML gas-leakage detector demonstrates the same pattern end-to-end for leak detection [33]. For a solar/battery-powered, intermittently-connected Hestia field node, this trades cloud compute for local power and firmware complexity, and may fit poorly with methods like unsupervised domain adaptation that assume a continuous stream of unlabeled target-domain data (Open questions, below).
• Task-type taxonomy. Orthogonal to architecture: unsupervised (PCA, clustering) for drift monitoring without labels; supervised classification (LDA, SVM, RF, CNN) for gas-identity or artifact-vs-anomaly decisions; supervised regression (PLS, SVR, ANN/CNN) for concentration quantification; and domain adaptation/transfer learning as a cross-cutting correction layer for any of the above.

Performance envelope

Every figure below comes from a different sensor family, gas target, and deployment condition, and the numbers are not directly comparable to one another; they establish the range of outcomes reported in this literature, not a single expected number for a Hestia H2 array.

Metric / claim	Value	Context	Ref
Short- vs. long-term drift, 62-sensor MOX array, 12-month field dataset	CV 0.38–1.21% (short-term) vs. 25.7–40.8% (long-term)	Continuously logged with co-located T/RH (mean 40.4 °C / 76.0% RH)	[15]
Calibration-transfer sample reduction (deep transfer learning vs. classical recalibration)	Up to 99.3% fewer samples (≈10 days → ≈2 h)	MOX gas sensors, 6-sensor pooled initial model; RMSE ≈18 ppb acetone	[13]
1-D CNN vs. SVM/MLP/KNN/Random Forest, raw e-nose time series	~10% average accuracy improvement	Review synthesis across multiple e-nose studies	[4]
Unsupervised multi-source domain adaptation (AMDS-PFFA) gas-recognition accuracy	83.20% (36-month benchmark); 93.96% (30-month self-built dataset)	No new target-domain labels required	[14]
Low-cost sensor array (XPOD) vs. reference-grade CH4 analyzer, mobile survey, RF calibration	$r = 0.719$; RMSD $= 0.3673$ ppm	ANN/multilinear alternatives worse or failed on spikes	[26]
Humidity-compensation system (random forest)	Resolution improved 87.7% → 99.3%	Electronic-nose humidity-compensation study	(unverified)
MOX H2:CO discrimination, SnO2–Co3O4 nanocomposite + Extreme Learning Machine	Response 22.75 at 300 ppm H2; H2:CO ratio ≈11.7:1	ML step required even for a purpose-engineered “ultra-selective” material	[28]
Sensirion SGP41 field VOC-index accuracy vs. reference PID	$R^2 = 0.27$ absolute; RMSE 64 ppb (unchanged by adding humidity); unit-to-unit $R^2 = 0.85$	Precision vs. accuracy gap for a modern MOX-derived commercial product	[23]
ZMOD4410 response to hydrogen	”Impressively, but not selectively,” full ppm range	Manufacturer’s own datasheet; TVOC/IAQ-optimized, not H2-calibrated	[20]
Resolving-power methodology validation	Validated on TMO MOX sensors, CO–humidity mixtures	Feature-selection gains over naive all-channel designs	[5]

Field verdict

Reading these together. The domain-adaptation and calibration-transfer results ([13][14]) are the most encouraging for Hestia: the cost of keeping a field array’s model current can drop by one to two orders of magnitude relative to full recalibration, without necessarily sacrificing much accuracy. The XPOD methane result [26] and SGP41 field-accuracy result [23] illustrate the opposite lesson: even a well-executed ML calibration on a real field platform tops out around $r \approx 0.7$–$0.85$ against a reference instrument for the best cases here, and a commercial, actively-marketed “AI” sensor achieves only $R^2 = 0.27$ against an absolute reference in real-world testing.

Neither is a metrologically traceable H2 measurement in the sense a GC or calibrated electrochemical cell provides — see Gas chromatography, Electrochemical cells, and Calibration & QA/QC — they are useful, cost-effective screening signals. Apply the discipline in Reading the specs: treat any array/ML accuracy figure as a ceiling on field performance, not an expectation, and check whether the output is a calibrated concentration or a unitless index before it enters a decision threshold.

Figures of merit

The single-device figures of merit — sensitivity, LOD, selectivity, drift — are defined on the Figures of merit page; an array plus a classification/regression layer is evaluated as a pipeline (hardware + features + model), so several of the rows below have no single-sensor analogue. Concentration values follow Concentration units.

Figures-of-merit table for this modality (9 rows)

Figure of merit	Symbol / formula	What it means here	Typical value for this technology	How to read it
Classification accuracy / F1 (gas ID)	$\mathrm{Acc}=(TP+TN)/N$; $F_1=2PR/(P+R)$	Fraction of labeled test exposures correctly assigned to the right gas class; F1 balances false positives against false negatives, more informative than raw accuracy once one class dominates a calibration set	1-D CNN outperforms SVM/MLP/KNN/Random Forest “by around 10% on average” on raw e-nose time series [4]; unsupervised multi-source domain adaptation (AMDS-PFFA) reaches 83.20% (36-month benchmark) and 93.96% (30-month self-built dataset) with no new target-domain labels [14]; SVM reaches “high classification accuracy in small sample sizes” but degrades once feature count exceeds sample count [3]	A single headline accuracy number hides whether errors cluster on one gas pair (e.g., H2 vs. CO); at least 17 published accuracy claims on the field’s most-used public benchmark are likely inflated by a temporal-clustering artifact rather than genuine gas discrimination [10][11]
Quantification error (RMSE / R² / r)	$\mathrm{RMSE}=\sqrt{\tfrac{1}{N}\sum(\hat c_i-c_i)^2}$; $R^2$, Pearson $r$	How close a regression model’s predicted concentration comes to a reference-instrument value, in absolute units — the array/ML analogue of a single device’s sensitivity-plus-LOD pairing	Random-forest-calibrated low-cost array (XPOD) vs. reference CH4 analyzer: $r=0.719$, RMSD $=0.3673$ ppm [26]; deep transfer learning on MOX gas sensors: RMSE ≈18 ppb acetone [13]; Sensirion SGP41 vs. reference PID: $R^2=0.27$ absolute, RMSE 64 ppb (unchanged by adding humidity), unit-to-unit $R^2=0.85$ [23]	Precision and accuracy are different claims — SGP41’s $R^2=0.85$ unit-to-unit says the sensor is repeatable, not that it is right; treat any array RMSE/R² figure as valid only for that paper’s reference instrument and gas matrix, not portable to H2-in-soil-headspace without re-validation
Array size & sensor diversity (orthogonality)	$n$ sensors; $\kappa(\mathbf{S})=\sigma_{\max}/\sigma_{\min}$; resolving power (RP)	Whether the array’s channels encode genuinely independent information about gas identity, or are redundant rows pointing in nearly the same direction of response space (Mechanism & physics, above)	No source reviewed reports a numeric $\kappa(\mathbf{S})$ or RP value for any H2-relevant array (unverified); the resolving-power hypervolume-ratio method is validated only on temperature-modulated MOX sensors in CO–humidity mixtures [5]; heterogeneous (multi-transduction-principle) arrays are consistently reported to improve conditioning more than adding same-family sensors [3][4]	Sensor count is not the FoM — a larger $n$ with a large $\kappa(\mathbf{S})$ still cannot be inverted reliably; before trusting any separability number below, check whether the array is homogeneous (MOX-only) or heterogeneous (MOX+EC+TCD)
Separability (PCA variance explained; LDA / Mahalanobis distance)	$\lambda_k/\sum_j\lambda_j$ (PCA); $D_M^2=(\mathbf{x}-\boldsymbol{\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x}-\boldsymbol{\mu})$	Whether gas classes (or “anomaly vs. background”) form distinct clusters in reduced-dimension space — PCA maximizes total variance with no label information; LDA/Mahalanobis distance use labels to separate classes directly	No source reviewed reports a numeric PCA variance-explained or Mahalanobis-distance figure for an H2-relevant array (unverified); documented instead is the qualitative failure mode — where humidity-driven variance exceeds H2-driven variance (established for MOX at Hestia-relevant concentrations), the first principal component tracks humidity, not gas identity	A clean PCA cluster plot proves nothing about gas separability unless the axes are labeled with % variance explained and cross-checked against known humidity/temperature states from the same run; LDA needs labeled ground truth a field program often lacks
Drift rate: short-term repeatability vs. long-term wander	CV(short), CV(long) — coefficient of variation	The gap between what a spot-check calibration verifies and what an unattended deployment actually experiences over its full lifetime	CV 0.38–1.21% short-term vs. 25.7–40.8% long-term, 62-sensor MOX array, 12-month field dataset, co-logged T/RH [15]	This ~20–60× gap is why a factory or bench calibration cannot stand in for field validation; any accuracy claim not qualified by deployment duration should be assumed to be the short-term number
Post-drift-correction accuracy	$\Delta N_{\text{samples}}$; RMSE/accuracy after correction	Whether a drift-correction method preserves accuracy while cutting the labeled recalibration effort needed to maintain it — the FoM that determines whether an unattended, multi-month field array is operationally viable	Deep transfer learning cuts new calibration samples needed by up to 99.3% (≈10 days of exposure down to ≈2 h) while holding RMSE ≈18 ppb acetone [13]; Direct Standardization cut transfer samples needed relative to full recalibration in a five-unit twin-array comparison [12]; unsupervised domain adaptation needs zero new target-domain labels at 83.20–93.96% accuracy [14]; a random-forest humidity-compensation system reportedly raised resolution 87.7% → 99.3% (unverified — no citation identified for this figure)	A model that is more accurate on day one but needs a full recalibration campaign every month is worse for an unattended station than one that is less accurate but self-corrects from a two-hour spot-check; the 99.3% figure is demonstrated only under continuously-powered lab/indoor conditions, not yet intermittent solar/battery field deployment [13][14]
Calibration-transfer error across time-batches or units	RMSE/accuracy on held-out later batch or twin unit	Whether a model trained on one instrument, or one point in time, still performs when applied to a different physical unit or a later “batch” months on — domain shift, not just noise	AMDS-PFFA domain adaptation: 83.20% accuracy on a 36-month benchmark, 93.96% on a 30-month self-collected dataset [14]; the field’s most-used public drift benchmark (16 MOX sensors, 6 gases, 36 months) was shown in 2021 to contain a temporal-clustering confound that likely inflates classification accuracy in at least 17 downstream publications [10][11]	Any single accuracy number quoted from this benchmark family should be treated skeptically until it is confirmed that gas-exposure and background-sampling order were randomized, not clustered by batch, during data collection
Model size / on-device (TinyML) inference cost	Model size (KB), inference latency, power draw	Whether a trained classifier/regressor can run directly on a field-node microcontroller instead of streaming raw data off-site — a deployment-cost FoM with no single-sensor analogue	No source reviewed reports a specific model-size, latency, or power figure for an H2-relevant array (unverified); demonstrated only qualitatively — a lightweight CNN on spatiotemporal “pseudo-image” e-nose transients targets handheld/microcontroller-scale odor recognition [31]; commercial MLOps tooling (Edge Impulse) has deployed gas-classification models to microcontroller hardware for industrial gas/VOC identification [32]; a dedicated TinyML gas-leakage detector demonstrates the same pattern end-to-end [33]	TinyML trades cloud compute for local power and firmware complexity, and may fit poorly with domain-adaptation methods that assume a continuous stream of unlabeled target-domain data (Open questions, below) — treat on-device deployment and continuous domain adaptation as two separate engineering commitments
Covariate-logging dependency	— (not a single number)	Nearly every row above — separability, drift-corrected accuracy, calibration-transfer error — is conditional on temperature/RH/pressure being logged as explicit regressors, not an optional add-on	A regression-based correction on commercial MEMS MOX parts reduced raw baseline standard deviation by 46–96% depending on the part once ambient T/RH were included as explicit regressors (unverified — no citation identified for this figure)	An array/ML accuracy or drift-correction number reported without stating whether T/RH/P were co-logged and used as regressors should be treated as incomplete — per Environmental covariance below, humidity is usually the largest-variance confound in raw response data

Reading the numbers

What dominates. Every other row here is downstream of one question: does accuracy survive months of unattended drift, and how much labeled effort does it take to keep it there? A lab-fresh classification/regression number (accuracy, F1, RMSE, R²) is a ceiling, not an expectation — the same 62-sensor MOX array that repeats to 0.38–1.21% CV over minutes wanders 25.7–40.8% CV over a year [15], and no array/ML accuracy figure in this literature has been validated past a few weeks under field-realistic, simultaneously-varying T/RH conditions.

Calibration-transfer and domain-adaptation methods that cut the labeled-recalibration burden by one to two orders of magnitude [13][14] therefore matter more here than raw accuracy: a field array that stays accurate with a two-hour quarterly check-in beats one that starts more accurate but needs a ten-day full recalibration to stay that way. All of it is conditional on temperature, RH, and pressure actually being logged as regressors from day one; skip that, and every separability and drift-correction number above reverts to whatever a raw, humidity-confounded PCA plot happens to show.

Environmental covariance & cross-sensitivity

The underlying device physics — humidity’s dissociative chemisorption on MOX surfaces, oxygen dependence, CO/CH4/H2S cross-response, pressure effects — is covered exhaustively in MOX chemiresistors and generalized across families in Environmental covariance. This section covers only the computational half: what an array/ML layer can and cannot do about it.

Humidity is the dominant confound the array/ML layer has to regress out. Water vapor’s effect on MOX baseline resistance is large enough, and common enough across co-located sensors, that it is frequently the largest-variance direction in raw response data — precisely why naive unsupervised PCA can track humidity, not gas identity (Mechanism, above).

The demonstrated fix is explicit covariate regression: one random-forest-based humidity-compensation system improved electronic-nose resolution from 87.7% to 99.3%, a 12.4-point gain (unverified — no citation identified for this figure). A companion regression-based correction on commercial MEMS MOX parts (mechanistic depth in MOX chemiresistors) reduced raw baseline standard deviation by 46–96% depending on the part once ambient temperature and humidity were included as explicit regressors (unverified — no citation identified for this figure) — concrete evidence the approach works, provided the covariates are actually logged.

Lab accuracy rarely survives field humidity/temperature swings. The Sensirion SGP41 field-validation result is a useful calibration-of-expectations number here: a commercial, ML-processed, MOX-derived VOC sensor, tested against a reference photoionization detector under real building conditions, achieved $R^2 = 0.27$ for absolute concentration — and adding humidity to the calibration did not reduce the RMSE [23]. The same device shows excellent unit-to-unit precision ($R^2 = 0.85$): repeatable and internally consistent, but not accurate against an absolute external reference.

Laboratory-chamber claims (single gas, single humidity, single temperature, fresh unaged sensor) are routinely an order of magnitude better than field performance against a reference instrument under realistic, simultaneously-varying conditions — apply this caveat to every number in the table above.

Cross-sensitivity & array/ML pitfalls

Cross-sensitivity is not solved by ML alone. No downstream regression can recover information the array never encoded: if every physical channel responds to CO and H2 in the same proportion (a poorly conditioned $\mathbf{S}$), no classifier or regression model — however sophisticated — can separate them from that array’s data alone. Even frontier materials-science work confirms this: a purpose-engineered SnO2–Co3O4 nanocomposite still required an Extreme Learning Machine post-processing step to reach only ~11.7:1 H2:CO discrimination [28]. Covariate and ML sophistication improve a well-conditioned hardware array; they cannot substitute for one.

Artifacts specific to the array/ML layer itself.

1. Batch/temporal confounding in training data. If calibration exposures are scheduled in clustered batches (all-H2 days, then all-CH4 days), a flexible classifier — especially a CNN — can learn to exploit the resulting temporal/drift correlation instead of the intended chemistry, inflating validation accuracy. A rigorous 2021 re-analysis of the most widely used public MOX drift-classification benchmark found exactly this — gases recorded in clustered batches, residual drift predicting gas identity from timing alone — concluding “previous studies conducted with this dataset very likely overestimate the accuracy of gas classification results,” implicating at least 17 downstream publications [10][11]. A direct warning against benchmarking a Hestia-built classifier purely against public data, or internal data gathered on a similarly clustered schedule.
2. Index-versus-concentration conflation. Several commercial products (Bosch BSEC, Sensirion’s Gas Index Algorithm) deliberately output a unitless relative “index” rather than an absolute ppm value, because the underlying sensor cannot support a metrologically defensible absolute field output [30]. Feeding such an index into a ppm-based decision threshold — the exact failure mode Reading the specs warns against — is a category error the vendors themselves have flagged by design choice.

Use for geologic H2

Field assessment

An array of covariate-logging, moderately-selective, cheap sensors plus a regression/classification layer is a legitimate way to build a dense, low-cost pre-screening grid, and the calibration-transfer literature shows the historically prohibitive maintenance cost of such a grid is a genuinely improving, actively researched problem. But no commercial product, academic paper, or field-deployment report found in this literature describes an H2-specific electronic nose validated in a natural-hydrogen exploration or soil-gas context; this is an unfilled gap, not a mature capability Hestia can buy off the shelf. Use it as a screening flag over a wide grid, not as a standalone quantitative result.

Where this is real. A heterogeneous array plus a regression/classification layer is a legitimate way to build a dense, low-cost pre-screening grid — flagging candidate anomaly zones across a wide survey area faster and more cheaply than deploying a reference-grade instrument at every point.

The calibration-transfer results above ([13][14]) show the maintenance cost of keeping such a grid’s models current — historically the central objection to any large array deployment — is genuinely improving: a 99.3% reduction in required recalibration samples [13] could make a distributed multi-station program logistically viable where it previously was not. Regression against co-logged temperature, RH, and pressure is not optional; it is the demonstrated mechanism (46–96% variance reduction on commercial parts; the 87.7%→99.3% resolution improvement above — both unverified, see Environmental covariance & cross-sensitivity) by which a raw array reading becomes interpretable, directly extending the covariate-regression framework in Environmental covariance.

Where this is not yet real for Hestia’s use case. No source found here describes an H2-specific electronic nose, or a MOX/electrochemical array plus ML pipeline, validated in a natural-hydrogen exploration or soil-gas context. The commercial “AI gas sensor” products (BME688, ZMOD4410/4510) are explicitly TVOC/indoor-air-quality-optimized, not H2-calibrated: the ZMOD4410’s own datasheet states it responds “impressively, but not selectively” to H2 [20], and neither manufacturer claims a validated absolute H2-ppm output without extensive customer-side training against application-specific reference data [17][19].

The academic e-nose literature is overwhelmingly oriented toward food quality, breath diagnostics, and indoor air/VOC identification — not trace-gas geochemistry. The field literature that does target natural-hydrogen soil-gas exploration — MONHyTOR and the broader instrument landscape covered in Soil-gas surveys — uses single electrochemical sensors, not machine-learning arrays, as its tool of record: MONHyTOR is built around one electrochemical H2 sensor (0–1000 ppm, ±5% full-scale error, 1 s sampling) [27], not an array. Isotopic fingerprinting (see Isotopes & fingerprinting) offers an orthogonal, chemically grounded route to gas-source discrimination worth weighing against a bespoke array/ML build for the same decision.

The realistic, defensible design, synthesizing the above with MOX chemiresistors’s independent conclusion that the field community has already converged on electrochemical cells over MOX for primary quantitative soil-gas H2 work:

1. Use a heterogeneous array (MOX + electrochemical, at minimum) — heterogeneity is what actually improves $\kappa(\mathbf{S})$, regardless of the specific gas target.
2. Mandatorily co-log temperature, RH, and barometric pressure — and, where practical, He as an inert co-tracer per Helium & co-tracers — at every node, as explicit regressors from day one.
3. Treat any array/ML output as a screening flag, not a quantitative result: a tripped anomaly threshold should trigger a follow-up measurement with a reference-grade instrument, mirroring existing single-sensor MOX/EC workflows.
4. Budget explicitly for periodic recalibration against ground truth as an ongoing operational cost — modern calibration-transfer/domain-adaptation methods can substantially reduce, not eliminate, it.
5. Do not present array/ML readings as absolute ppm concentrations without periodic calibration against a reference instrument in the actual target matrix (soil headspace with CO2/CH4/H2S/He co-occurrence) — no source reviewed validates this specific matrix.

Downhole. Everything in MOX chemiresistors and Downhole deployment about MOX’s unsuitability for the high-temperature, high-pressure, no-local-electronics-preferred downhole environment applies with at least equal force to an array built from multiple MOX elements plus onboard pattern-recognition compute — more parts, more heat, more electronics, compounding the same intrinsic-safety and thermal-qualification problems. Array/ML approaches here are a surface and near-surface tool only.

Vendors, instruments & standards

Category	Name	Key specs / caveats	Ref
Commercial IC, on-device AI	Bosch Sensortec BME688	MEMS MOx “gas-scan” + T/RH/P; detects VOC/VSC/CO/H2 “in the ppb range” but needs BSEC or customer-trained AI-Studio for usable output; not an inherently calibrated H2-ppm sensor	[17][18][19]
Commercial IC, embedded ML	Renesas ZMOD4410	MEMS MOx + CMOS ASIC, pre-trained firmware, TVOC/IAQ-marketed; own datasheet: responds “impressively, but not selectively” to H2	[20]
Commercial IC, embedded ML	Renesas ZMOD4510	Same platform, O3/NO2 channel paired with ZMOD4410 for cross-compensation; 10-yr qualified lifetime (mfr. claim)	[21]
Commercial IC, index output	Sensirion SGP41	Unitless VOC/NOx Index (1–500) via open-source Gas Index Algorithm, explicitly not absolute concentration; field validation $R^2=0.27$ absolute, $R^2=0.85$ unit-to-unit	[22][23]
Commercial photonic/biochemical array	Aryballe NeOse Advance	64 peptide biosensors on a silicon-photonic Mach-Zehnder array, ML-matched to an “olfactive database”; odor/VOC market, no H2-specific validation	[24][25]
Public benchmark dataset	UCI Gas Sensor Array Drift Dataset	16 MOX sensors, 6 gases, 36 months, 10 time-batches; shown in 2021 to contain a temporal-clustering confound likely inflating published accuracies in ≥17 papers	[10][11]
Public benchmark dataset	Wörner/Eimler/Pein-Hackelbusch one-year e-nose dataset	62 SnO2 nanowire sensors, 12 months, co-logged T/RH; clearest published drift-vs-repeatability quantification (25.7–40.8% vs. 0.38–1.21% CV)	[15]
Academic research group	SENSOR Lab, Brescia	TCO/virtual-array and pattern-recognition strategies for non-selective MOX elements	[29]
Academic program (historical)	Caltech carbon-black/polymer array program	1990s cross-reactive chemiresistor work; early resolving-power and drift-correction comparisons	[6][7]
Academic methodology	Resolving-power framework	Hypervolume-ratio method for array resolving power and feature selection, validated on TMO MOX + CO/humidity mixtures	[5]
TinyML tooling	Edge Impulse	Commercial ML-ops platform for deploying gas-classification models to microcontroller hardware; illustrative, not H2-specific	[32]

Standardization landscape. Fuller regulatory context is in Standards & regulations; the e-nose-specific picture is thin and explicitly acknowledged as immature by its own reviewers [16].

Standard	Assessment
NTA 9055:2012 (Netherlands),	the first e-nose standardization effort, was assessed as “extremely synthetic and too generic.”
VDI 3518-3:2018 (Germany)	is the most complete guideline to date — three measurement tasks, ≤30% deviation from reference-gas value — but its fixed minimums may bar prototype-stage systems.
EN 14181/EN 15267	were built for automatic measuring systems, not e-noses, but their QAL1/QAL2/QAL3 framework is a mature uncertainty-evaluation precedent.
CEN TC/264 WG41 (est. 2014)	pursues a technology-agnostic “black-box” approach, still preliminary;
UNI 1605848 (Italy, proposed)	offers the most flexible, performance-declaration philosophy reviewed; and the
IEEE Standards Association	is developing a machine-olfaction performance-verification standard, unconfirmed as of the underlying review [16].

State of research & practice

Three threads dominate current activity beyond the 1982/1994 conceptual foundation [1][2]:

Classical pattern recognition remains the workhorse, but deep learning is closing in. Reviews catalogue PCA, LDA, PLS, SVM, RF, KNN, and shallow ANN as still dominant in practice, mostly because available datasets are small, but document a clear trend toward CNN/LSTM architectures outperforming classical methods by ~10% on average once enough data exist [3][4]. Information-theoretic feature selection — the resolving-power framework, tracing through Gardner-Bartlett and the Caltech carbon-black program [5][6] — is the more rigorous end of this spectrum, explicitly quantifying when adding sensors or features helps versus merely adds overfitting risk.

Drift correction and calibration transfer is the most active open research area here. The trajectory runs from early multivariate drift correction (CC, OSC) in the late 1990s/2000s [8][9], through classifier-ensembles built around the 36-month UCI benchmark [10], to Direct-Standardization transfer [12], to deep transfer learning [13] and unsupervised domain adaptation [14] — consistently toward reducing the labeled-data burden of staying calibrated, the most important trend for any unattended, long-duration field deployment.

A counterpoint: the 2021 finding that the field’s most-used public benchmark contains a temporal-clustering confound that likely inflated a decade of published accuracy claims [11] argues for skepticism toward any single benchmark number, including several quoted above.

Commercial productization has shifted the problem from sensing physics to closed-source software, with real transparency costs. Bosch (BME688 + BSEC/AI-Studio) and Renesas (ZMOD + embedded AI) both ship, functionally, a bare MOX sensor plus a proprietary compensation layer. Sensirion instead abandoned absolute-concentration output for an openly published but explicitly relative Gas Index Algorithm [22] — an honest, if limited, response to the field-accuracy problem [23]. Aryballe is a structurally different, optical/biochemical bet on the same paradigm, aimed at a different market [24][25].

TRL honesty. Reviewers of the broader e-nose field are consistently candid that commercial adoption lags the academic literature — sensor noise, drift, calibration burden, and inconsistent data-handling protocols are cited as barriers to real-world implementation, with standardization bodies still calling the field “not mature enough” for rigid performance standards [16].

Applied to Hestia’s niche: zero commercial products and zero field-deployed academic studies combine an H2-targeted array with ML discrimination in a geologic-hydrogen context. The nearest analogues are the general e-nose/TMO/drift-correction literature transplanted from indoor air quality, food safety, and medical diagnostics, and mobile methane monitoring [26] — encouraging in direction, but a different gas, range, and interferent environment than Hestia’s soil-gas H2 problem.

Open questions

1. No source validates any array/ML gas-discrimination pipeline in the actual Hestia-relevant matrix (soil headspace with co-occurring CH4/CO2/H2S/He, hypoxic microzones, wide diurnal T/RH swings). Commission or find such a study before committing engineering budget to a bespoke array/ML station.
2. Determine whether the domain-adaptation/transfer-learning drift-correction results [13][14] generalize to intermittent, remote, low-power field deployment, versus the continuously-powered indoor/lab conditions under which they were demonstrated.
3. Quantify, in-house, the actual conditioning of a candidate array’s sensitivity matrix $\mathbf{S}$ for H2/CH4/CO2/H2S/He before investing in an ML layer — a large $\kappa(\mathbf{S})$ is a hardware/materials problem no software fixes.
4. Decide build-vs-buy on the ML/compensation layer: in-house PLS/RF regression is transparent but needs data-science capacity; a vendor “AI” module deploys faster but is opaque, not H2-validated out of the box, and creates a closed-source dependency.
5. Run a controlled re-analysis check on any internally collected calibration dataset for the temporal-clustering confound documented by Dennler et al. [11]; randomize gas-exposure and background-sampling order during calibration-data collection.
6. Investigate whether a heterogeneous MOX+EC(+TCD) array plus a simple PLS/RF regression meaningfully outperforms a single well-characterized electrochemical cell for Hestia’s decision threshold (~1 ppm-to-percent anomalies against a ~0.5 ppm background) — no study answers this directly, and the marginal value may be smaller than the added drift-model maintenance burden justifies.
7. Reconcile the “index output” convention (Bosch BSEC, Sensirion Gas Index Algorithm) with Hestia’s ppm-based decision thresholds explicitly in any SOP that uses such a module.
8. Bench-test a temperature-cycled/TMO virtual-array approach specifically for H2 vs. CH4 vs. CO discrimination at Hestia-relevant concentrations (1 ppm–percent) — TMO/TCO is demonstrated for VOC/general trace-gas discrimination, but no direct H2-focused TMO selectivity study was found.

Related pages

• Why H2 is hard
• Sensor selection guide
• Comparison charts
• Reading the specs
• MOX chemiresistors
• Electrochemical cells
• Plasmonic & metamaterial sensors
• UV/EUV & fluorescence
• Environmental covariance
• Helium & co-tracers
• Isotopes & fingerprinting
• Soil-gas surveys
• Downhole deployment
• Calibration & QA/QC
• Gas chromatography
• Standards & regulations

References

Full reference list (33 entries)

1. Persaud, K.; Dodd, G. “Analysis of discrimination mechanisms in the mammalian olfactory system using a model nose.” Nature, 299, 352–355 (1982). https://www.nature.com/articles/299352a0
2. Gardner, J.W.; Bartlett, P.N. “A brief history of electronic noses.” Sensors and Actuators B: Chemical, 18–19, 210–211 (1994). https://www.sciencedirect.com/science/article/abs/pii/0925400594870853
3. Zhai, Z.; Liu, Y.; Li, C.; Wang, D.; Wu, H. “Electronic Noses: From Gas-Sensitive Components and Practical Applications to Data Processing.” Sensors, 24(15), 4806 (2024). DOI: 10.3390/s24154806. https://pmc.ncbi.nlm.nih.gov/articles/PMC11314697/
4. Ye, Z.; Liu, Y.; Li, Q. “Recent Progress in Smart Electronic Nose Technologies Enabled with Machine Learning Methods.” Sensors, 21(22), 7620 (2021). DOI: 10.3390/s21227620. https://pmc.ncbi.nlm.nih.gov/articles/PMC8619411/
5. Fernandez, L.; Yan, J.; Fonollosa, J.; Burgués, J.; Gutierrez, A.; Marco, S. “A Practical Method to Estimate the Resolving Power of a Chemical Sensor Array: Application to Feature Selection.” Frontiers in Chemistry, 6, 209 (2018). DOI: 10.3389/fchem.2018.00209. https://www.frontiersin.org/journals/chemistry/articles/10.3389/fchem.2018.00209/full
6. “Quantitative Study of the Resolving Power of Arrays of Carbon Black−Polymer Composites in Various Vapor-Sensing Tasks.” Analytical Chemistry (ACS). https://pubs.acs.org/doi/abs/10.1021/ac971204+
7. “Comparison of analytical methods and calibration methods for correction of detector response drift in arrays of carbon black-polymer composite vapor detectors.” Sensors and Actuators B: Chemical. https://www.sciencedirect.com/science/article/abs/pii/S0925400504003752
8. Artursson, T. et al. “Drift correction for gas sensors using multivariate methods.” Journal of Chemometrics, 14, 711–723 (2000). https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/abs/10.1002/1099-128X(200009/12)14:5/6%3C711::AID-CEM607%3E3.0.CO;2-4
9. Padilla, M. et al. “Drift compensation of gas sensor array data by Orthogonal Signal Correction.” Chemometrics and Intelligent Laboratory Systems. https://www.sciencedirect.com/science/article/abs/pii/S0169743909001877
10. Vergara, A.; Vembu, S.; Ayhan, T.; Ryan, M.A.; Homer, M.L.; Huerta, R. “Chemical gas sensor drift compensation using classifier ensembles.” Sensors and Actuators B: Chemical, 166–167, 320–329 (2012). DOI: 10.1016/j.snb.2012.01.074.
11. Dennler, N.; Rastogi, S.; Fonollosa, J.; van Schaik, A.; Schmuker, M. “Drift in a Popular Metal Oxide Sensor Dataset Reveals Limitations for Gas Classification Benchmarks.” arXiv:2108.08793 (2021). https://arxiv.org/abs/2108.08793
12. Fonollosa, J.; Fernández, L.; Gutiérrez-Gálvez, A.; Huerta, R.; Marco, S. “Calibration transfer and drift counteraction in chemical sensor arrays using Direct Standardization.” Sensors and Actuators B: Chemical, 236, 1044–1053 (2016). https://www.sciencedirect.com/science/article/abs/pii/S0925400516307742
13. “Comparison of Transfer Learning and Established Calibration Transfer Methods for Metal Oxide Semiconductor Gas Sensors.” Atmosphere, 14(7), 1123 (2023). https://www.mdpi.com/2073-4433/14/7/1123
14. Zhang, W.; Hu, S.; Zhang, Z.; Zheng, Y.; Wang, Q.J.; Lin, Z. “Unsupervised Attention-Based Multi-Source Domain Adaptation Framework for Drift Compensation in Electronic Nose Systems.” arXiv:2409.13167 (2024). https://arxiv.org/html/2409.13167v1
15. Wörner, J.; Eimler, J.; Pein-Hackelbusch, M. “Long-term drift behavior in metal oxide gas sensor arrays: a one-year dataset from an electronic nose.” Scientific Data, 12, 1628 (2025). DOI: 10.1038/s41597-025-05993-8. https://pmc.ncbi.nlm.nih.gov/articles/PMC12508210/
16. Cipriano, D.; Capelli, L. “Evolution of Electronic Noses from Research Objects to Engineered Environmental Odour Monitoring Systems: A Review of Standardization Approaches.” Biosensors, 9(2), 75 (2019). DOI: 10.3390/bios9020075. https://pmc.ncbi.nlm.nih.gov/articles/PMC6627819/
17. Bosch Sensortec GmbH. “BME688” datasheet, BST-BME688-DS000. https://www.bosch-sensortec.com/media/boschsensortec/downloads/datasheets/bst-bme688-ds000.pdf
18. Bosch Sensortec GmbH. “BME688” gas sensor product page. https://www.bosch-sensortec.com/en/products/environmental-sensors/gas-sensors/bme688
19. Bosch Sensortec GmbH. “BME AI-Studio” documentation. https://www.bosch-sensortec.com/software/bme/docs
20. Renesas Electronics. “ZMOD4410 Gas Sensor Module for TVOC and Indoor Air Quality,” datasheet. https://www.renesas.com/en/document/dst/zmod4410-datasheet
21. Renesas Electronics. “ZMOD4510 Gas Sensor Module for Nitrogen Dioxide and Ozone,” datasheet. https://www.renesas.com/en/document/dst/zmod4510-datasheet
22. Sensirion AG. “SGP41” product page; “Sensirion’s Gas Index Algorithm” (open-source). https://sensirion.com/products/catalog/SGP41; https://github.com/Sensirion/gas-index-algorithm
23. Krause, A.; Brooke, E. “How Accurate is the Sensirion SGP41 TVOC Sensor?” AirGradient blog, Dec. 10, 2024. https://www.airgradient.com/blog/accuracy-sensirion-sgp41/
24. Aryballe. “NeOse Advance” product page. https://aryballe.com/our-products/neose-advance/
25. CEA-Leti. “Aryballe offers a universal electronic nose.” News article. https://www.leti-cea.com/cea-tech/leti/english/Pages/What’s-On/News/Aryballe-offers-a-universal-electronic-nose.aspx
26. “Utilization of a Low-Cost Sensor Array for Mobile Methane Monitoring.” Sensors, 24(2), 519 (2024). https://www.mdpi.com/1424-8220/24/2/519
27. “In-soil hydrogen concentration monitoring using MONHyTOR: application for natural hydrogen exploration.” Advances in Geosciences, 67, 57 (2025). https://adgeo.copernicus.org/articles/67/57/2025/
28. “SnO2-Co3O4 nanocomposite sensor: Achieving ultra-selective hydrogen detection in mixed gas environments.” Sensors and Actuators B: Chemical (2024). https://www.sciencedirect.com/science/article/abs/pii/S0925400524012516
29. “Metal Oxide Gas Sensors, a Survey of Selectivity Issues Addressed at the SENSOR Lab, Brescia (Italy).” Sensors, 17(4), 714 (2017). https://pmc.ncbi.nlm.nih.gov/articles/PMC5421674/
30. Sensirion AG. “What is a metal oxide (MOX) sensor?” Information note. https://sensirion.com/media/documents/0083CDF4/6294DFEA/Info_Note_MOX_sensor.pdf
31. “OdorNet: A lightweight odor recognition method for TinyML in handheld electronic noses using spatiotemporal pseudo-images.” Sensors and Actuators B: Chemical. https://www.sciencedirect.com/science/article/abs/pii/S0925400525011694
32. Edge Impulse. “Sniffing Out Industrial Chemicals Using tinyML.” https://www.edgeimpulse.com/blog/sniffing-out-industrial-chemicals-using-tinyml/
33. “A Gas Leakage Detection Device Based on the Technology of TinyML.” Designs (MDPI), 11(2), 45 (2023). https://www.mdpi.com/2227-7080/11/2/45