MEMS-Based Micropacked Thermal Desorption GC/PID for In-Field Volatile Organic Compound Profiling from Hot Mix Asphalt

Abstract

Background: In response to the growing demand for the real-time, in-field characterization of odorous anthropogenic emissions, this study develops and uses a MEMS-based micropacked thermal desorption Gas Chromatography system coupled with a PhotoIonization Detector (GC/PID) for Hot Mix Asphalt (HMA) plant emissions. Methods: The innovative portable device, Pyxis GC, enables the high-sensitivity profiling of Volatile Organic Compounds (VOCs), particularly aldehydes and ketones, with sub-ppb detection limits using ambient air as the carrier gas. A comprehensive experimental design optimized the preconcentration parameters, resulting in an efficient, green analytical method evaluated via the Green Analytical Procedure Index (GAPI). Sorbent comparison showed quinoxaline-bridged cavitands outperform the conventional materials. Results and conclusions: The method was successfully deployed on site for source-specific sampling at an HMA plant, generating robust emission fingerprints. To assess environmental impact, a Generalized Additive Model (GAM) was developed, incorporating the process temperature and Sum of Odour Activity Values (SOAV) to predict odour concentrations. The model revealed a significant non-linear influence of temperature on emissions and validated its predictive capability despite the limited sample size. This integrated analytical–statistical approach demonstrates the utility of MEMS technology for real-time air quality assessment and odour dispersion modelling, offering a powerful tool for environmental monitoring and regulatory compliance.

1. Introduction

Odorous emissions stemming from anthropogenic activities may contribute to adverse health outcomes, diminished property values, and increased levels of public annoyance. Therefore, many jurisdictions classify odour as an atmospheric pollutant and regulate emissions and/or impacts from odour-generating activities at a national, state, or municipal level [1]. Asphalt Mixture Plants represent common sources of odour-active Volatile Organic Compounds (VOCs), due to the high temperature (ranging from 150 to 180 °C) of the final Hot Mix Asphalt (HMA) [2]. Given this context, the need for effective monitoring techniques becomes evident. A current monitoring technique for odour concentration (OC) is dynamic olfactometry [3], which reflects the human perception threshold to odorous compounds but presents some limitations, such as subjectivity, the definition of odour responsibility, and the impossibility of continuous monitoring. Analytical short-time air monitoring, using high sensitivity and specificity to calculate the relationship between the concentration of compounds and the relative odour concentration and anticipate the odour’s perception, is encouraged. Moreover, advancements in automation driven primarily by the growing demand for Green Analytical Chemistry (GAC) [4] have facilitated the evolution of person-portable Gas Chromatography (GC) instruments. Tackling the challenges of on-site airborne chemical exosomes required the implementation of diverse analytical strategies and technological advancements; this approach contributed to miniaturization and automation via online preconcentration, which can reduce the time and the cost of sample preparation [5].

The recent upswing of Micro-Electro-Mechanical Systems (MEMSs) technology has brought new availability and a range of automation to air monitoring. Even though the first work on a microchip-based chromatographic system was a miniaturized GC in 1979 [6], using MEMSs by etching channels into a substrate, this groundbreaking work had not led to further developments of related skills or technology until the 1990s, with the introduction of chip-based preconcentrators instead of an injector to increase sensitivity and selectivity when the solute concentration is below the detection limit of the detector [7]. Since the first preconcentrating microstructures by Frye-Mason et al. [8] in 1999, many works have been carried out, combined with a wide range of adsorbents capable of generating sharp injection plugs [9,10,11,12]. Water condensation problems can be prevented by sampling VOCs on traps filled with hydrophobic materials kept at room temperature [13]. As no single adsorbent material can effectively capture VOCs across a broad volatility spectrum [14], multi-layer sorbent traps comprising various combinations of hydrophobic materials have been introduced to enable VOC collection under varying humidity conditions. Among them, those based on graphitic carbons and poly-2,6-diphenyleneoxide porous polymers (Tenax) are widely used as non-specific sorbents for the quantitative thermal desorption of polar and non-polar components up to C20 [15], despite the low mass of the artefacts provided [16]. This explains the recommended by the US Environmental Protection Agency (EPA) TO-17 Method of multi-layer tubes filled with graphitized Tenax (Tenax-GR) and Carbopack B set in series. Similar performances are provided by traps filled with combinations of Carbopack C and Carbopack B [17]. Traps containing Carbograph 5 are also proposed for the collection of very volatile VOCs at an ambient temperature [13].

The introduction of MEMS-based micropacked GC columns allowed for the separation of unsaturated isomers of light hydrocarbons, and standardized methods [18] as the detectors for chip-based GC. Detection has consistently represented a major challenge in the field of MEMSs, primarily due to the need for highly sensitive techniques capable of analyzing ultrasmall sample volumes. The miniaturization of the Thermal Conductivity Detector (TCD) started with the first micro GC in 1979 [6], and since then, several studies have been published in this area [19]. Yang et al. [20] proposed a miniaturized planar Flame Ionization Detector (FID) in 2000, where the oxygen–hydrogen flame burns inside a glass–silicon chip. Many sensors, such as the chemiresistor array and Metal Oxide (MOX) sensors, have been reported for chip-based GC [20,21]. The proposal of the Photo Ionization Detector (PID) as a selective detector is recent [22,23,24,25,26,27]. Conventional detectors can suffer from large dead volumes and thus slow response times. Therefore, proposals relating to optical detectors used in micro GCs based on surface plasmon resonance [28], ring resonators [29], Fabry–Perot interferometry [30], diode laser spectroscopy [31], optomechanical sensors [32], photonic crystal slab [33], and photoacoustic spectroscopy [34] have been offered. Person-portable GC, in contrast to conventional transportable chromatographs, are typically smaller, lighter, and more energy-efficient, often powered by batteries and using compact gas tanks. While some person-portable devices weigh around one kilogram, others are built on chips using MEMS technology. To be considered truly portable, a system must be easily carried or moved. This can, of course, be very subjective, but objective frameworks of the Health and Safety Executive of the United Kingdom [35] and the International Organization for Standardization [36] set out several helpful guidelines for safe lifting capacity, depending on the task at hand and the posture of the individual. Gałuszka et al. [37] defined a portable system as one ideally weighing no more than 10 kg. It follows that the instrument should be a monolithic system, integrating all the necessary computers and electronics into a single housing. A summary of person-portable GC instruments is in

Table 1. Summary of person-portable GC instruments available on the market. Including information about pre-concentration system, detector type, ruggedness, LOQ, dimensions, and manufacturer.

	Instrument	Pre-Concentrator	GC Column	Rugged	LOQ	Dimensions (cm)	Weight (kg)	Producer
GC/MS	Torion T-9	None	Capillary	Rugged	ppb-ppt	38.1 × 39.4 × 22.9	14.5	PerkinElmer Inc. (Hopkinton, MA, USA) Link: PerkinElmer Inc. https://content.perkinelmer.com/library/prd-torion-t-9-portable-gcms.html (28 March 2025)
	HAPSITE ER	None	Capillary (15 M, Rtx-1 MS, 0.25 mm × 0.1 µm)	Rugged	ppm to-ppt	46 × 43 × 18	18	INFICON (Bad Ragaz, Switzerland) Link: https://www.inficon.com/en (28 March 2025)
	Griffin G510	On-board	Capillary	Rugged, classified asIP65	ppb or less	33.7 × 33.7 × 40	16.3	Teledyne FLIR LLC (Wilsonville, OR, USA) Link: https://www.flir.it/ (28 March 2025)
	Enviroprobe	None	Capillary (es. DB-1, 5 m, 0.5 mm, 0.25 µm)	Rugged	ppb	15 × 12.7 × 1.7	1.3	FemtoScan Corporation (Bountiful, UT, USA) Link: http://www.femtoscan.com/index.htm (28 March 2025)
GC/FID	NCMS 6300	On-board (low-temperature preconcentration)	Capillary	Rugged	0.5 ppb	Data not available	Data not available	Nutech Instruments (Richardson, TX, USA) Link: https://www.nutechinst.com/ (28 March 2025)
	PetroAlert® 9100	None	Packed, micro-packed, or capillary columns; Specific to application	Rugged	ppm	13 × 48.26 × 41	13.6	AMETEK MOCON (Brooklyn Park, MN, USA) Link: https://www.ametekmocon.com/ (28 March 2025)
GC/TCD	Micro GC 990	None	Capillary (WCOT, PLOT, micropacked)	Rugged	WCOT 05 ppm; PLOT 2 ppm Micropacked 10 ppm	28.28 × 14.5 × 32.9	7.3 (15.6 cabined)	Agilent (Santa Clara, CA, USA) Link: https://www.agilent.com/ (28 March 2025)
	Micro GC Fusion	None	Capillary	Rugged	1 ppm	46.2 × 19.6 × 25.4	6.2	INFICON (Bad Ragaz, Switzerland) Link: https://www.inficon.com/en/products/chemical-detection-and-utility-monitoring/micro-gc-fusion-gas-analyzer (28 March 2025)
	DynamiQ	None	MEMS	Rugged, classified as IP65	0.5 ppm	28.9 × 12.2 × 25.8	15	QMicro-Sensirion (Enschede, NL, USA) Link: https://www.qmicro.com/about/ (28 March 2025)
	I-Graph X	None	MEMS	Rugged	1 ppm	31 × 12 × 29	4.7	I-GRAPHX GmbH (Reinbek, DE, USA) Link: https://www.i-graphx.com/ (28 March 2025)
	DPS Micro-TCD GC	None	Capillary	Rugged	ppm	20 × 15 × 10	8	DPS Instruments Europe GmbH (Bad Honnef, DE, USA) Link: https://www.dps-instruments.com/ (28 March 2025)
GC/PID	FROG-4000 VOCAM	On-board	MEMS	Rugged	sub-ppb	25.4 × 19.1 × 37	2.2	Defiant Technologies Inc (Albuquerque, NM, USA) Link: https://www.defiant-tech.com/ (28 March 2025)
	eGC	Selective-trap	Capillary (e.g., 0.53 mm × 10 m)	Rugged	0.5 ppb	44.5 × 49.5 × 21.3	19.9	ENMET (Ann Arbor, MI, USA) Link: https://enmet.com/ (28 March 2025)
	GC PID	Selective-trap	Capillary	Rugged	<1 ppb	36 × 30 × 15	7	PAS Technology Deutschland GmbH (Magdala, DE, USA) Link: http://www.pas-technologies.com/(11 Feburary 2025)
	PyxisGC	Selective-trap	MEMS	Rugged	0.01 ppb	42 × 62 × 21	19	Pollution (Budrio, IT, USA) Link: https://www.pollution.it/it/ (28 March 2025)
	Baseline® 9200	Selective-trap	Capillary	Rugged	sub-ppb	13 × 48.26 × 56.9	13.6	AMETEK MOCON (Brooklyn Park, MN, USA) Link: https://www.ametekmocon.it/products/gaschromatographs/9200dualdetectorgc (28 March 2025)
	PID-Analyzer	None	Capillary	Data not available	ppb	Data not available	Data not available	IUT Technologies GmbH (Berlin, DE, USA) Link: https://iut-technologies.de/ (28 March 2025)
	NovaTest P300	Selective-trap	Capillary	Rugged	sub-ppb	47 × 35.7 × 17.6	10	Nanova Environmental, Inc. (Columbia, MO, USA) Link: https://nanovaenv.com/ (28 March 2025)
	312 Portable GC	Selective-trap	Capillary	Rugged	ppb	47.5 × 35 × 19	12.2	PID ANALYZERS, LLC (Sandwich, MA, USA) Link: https://hnu.com/ (28 March 2025)
	GC 866 microVOC-Trap	Selective-trap	Capillary	Rugged	ppt	48.2 × 60.0 × 18.0	17	Chromatotec® (Saint-André de Cubzac, France) Link: https://chromatotec.com/ (28 March 2025)

.

Therefore, to appreciate the fingerprint of the odour rather than the single molecule, the key attributes of a micro GC system—coupled with non-mass detectors—are its compactness, automaticity, reduced power consumption, minimal maintenance costs, and suitability for field analysis.

The global micro GC market was valued at 256 million in 2022, and is projected to reach United States Dollar (USD) 437 million by 2029, at a Compound Annual Growth Rate (CAGR) of 8.5% during the forecast period. The first commercial micro GC product released by the C2V Company dates back to 2010, and it was called C2V-200 [38]. Since then, only a few commercialization experiences have been presented to date, and the remaining projects are within research groups [33,39]. The challenges faced by this work include the in-field and laboratory characterization of the odorous emissions from HMA via MEMS technology via a new, innovative micro GC monitoring system. To contribute to the growing use of this technology, this study explores the air sampling for the preconcentration, and the related analysis performed by GC/PID. So, to reduce the number of experiments needed for the optimization of the preconcentration, and thus the resources utilized, the development of the method was performed by applying the Design of Experiment (DoE) approach. Using the GAPI approach, the conformity principles of GAC were evaluated for the developed method. The analysis of olfactometry nuisance in this study combines field measurements with subsequent modelling to assess the impact of odorous emissions. Field sampling was conducted over several days at the site of the final HMA production, capturing variations in odour intensity and composition. Following the field data collection, a statistical modelling based on Generalized Additive Models (GAMs) approach was employed to predict the odour concentration emitted by channelled emissions of Asphalt Mixture Plants, offering a comprehensive understanding of the spatial and temporal distribution of the olfactory nuisance. This integrated methodology aims to provide insights into the underlying factors contributing to the perceived olfactory impact, facilitating more effective management and mitigation strategies.

2. Materials and Methods

2.1. Analytical Instruments: Micro GC, Preconcentration, MEMS Packed Column

Pyxis GC (Pollution, Budrio, Italy) combines the MEMS of microfluidics for selective preconcentration and GC separation (Figure 1), with the peak quantification given by the miniaturized PID (10.6 eV), allowing VOC monitoring (detection limit 0.01 ppb of benzene) using ambient air as the carrier gas.

The use of flow inversion in the pre-concentrator between the sampling and the desorption phase ensures sharp injection peaks into the packed GC column, optimizing the chromatographic separation. The robustness, enabling monitoring from −40 °C to 50 °C, was represented by dimensions (420 mm × 620 mm × 210 mm) and weight (19 kg), with batteries that can be connected to the electric mains or photovoltaic cells. Pyxis GC was equipped with an auto-check system to verify and eventually correct the instrumental drift. Remote control for data access was performed by the Pollution Guardian IoT Service (Pollution, Budrio, Italy), and notifications via SMS, email, or a dedicated app were applied. The wind speed and direction, barometric pressure, relative humidity (RH), air temperature, and solar radiation were obtained by the Pyxis GC weather control unit. Additionally, the pre-concentration efficiency of quinoxaline-bridged cavitand (QxCav) [40], Carbograph 1 (Lara, Formello, Italy), and Activated Porous Carbon Fibre (APCF) by Labtech (Labtech s.r.l., Rome, Italy) [41], used as a sorbent micropacked on the MEMS pre-concentrators, was assessed in terms of mass loading, trapping time, desorption temperature, and desorption time. For the MEMS GC packed column, the material used in our investigation was non-porous Carbograph 2SAP (Lara) with a surface area of 13 m²/g. Carbograph is obtained via the graphitization process from different kinds of carbon black derived from mineral oil or natural gas, heated in an oxygen-poor environment.

The initial column temperature was set to 50 °C for 50 s, then reached 180 °C in 180 s for a total analysis time of 600 s. Ambient air at a flow rate of 8.0 mL/min was used as the carrier gas.

For confirmation of the considered aldehydes and ketones detected in the field, the Pyxis GC was coupled with a Varian 320-QpQ-MS (Varian Inc.) detector. The MEMS GC packed column was connected to the injector via retention gaps (1 m, ID 0.25 mm, part # 10009, Restek, Bellefonte, PA, USA) and to the MS via a particle trap (1 m, ID 0.25 mm, part # 19774, Restek). Helium was used as the carrier gas at 3 mL/min.

2.2. Training and Calibration: Dynamic Atmosphere Standard by ATIS

Calibration of the Pyxis GC was carried out using a dynamic atmosphere system (ATIS, Supelco, Merck KGaA, Saint Louis, MO, USA).

Subsequently, the configured instrument was employed to determine the odour fingerprint of the emissions from the HMA, specifically the conventional Surface Layer (SL) made (w/w) of 5% bitumen (50/70 type), 65% aggregate, 8% filler, and 22% Reclaimed Asphalt Pavement (RAP). The RH of the HMA-SL declared by the manufacturer was less than 1%. For the analysis, ten grams of the HMA-SL sample was loaded into a 40 mL crimp-top amber glass vial (Supelco, Merck KGaA, Saint Louis, USA, product no. 23189), with a closure-type screw top vial, a closure-type white polypropylene hole cap, and a PTFE/silicone septa equipped with two 18-gauge needles (1.219 mm external diameter), one of which was connected to Pyxis GC. The vial was then heated at the bottom at 160 °C and cooled at the head with compressed air at 25 °C.

For the calibration of the instrument, stock solutions of butanal, pentanal, hexanal, heptanal, octanal, 2-butanone, 2-hexanone, and 3- heptanone in methanol were prepared and were stored at −20 °C.

Starting from stock solutions, a mixed working solution containing the diluted substances in methanol was prepared. A five-level calibration curve was prepared for each compound by adding the appropriate volume of STD work solution, and the standard atmospheres were generated through the adsorbent Tube Injector System (ATIS, Supelco, Bellefonte, PA, USA), in order to obtain the desired airborne concentration. The analysis was carried out using the following instrumental parameters: a sampling time of 100 s, a flow rate sampling of 250 mL/min, a desorption time of 10 s, and a desorption temperature of 120 °C (obtained via chemometric optimization in Section 3.2). The analyte concentrations for each calibration-generated atmosphere are reported in

Table 2. Concentration levels, reported in µg/m³ for each of the five calibration solutions.

Compound	Level 1	Level 2	Level 3	Level 4	Level 5
Butanal	0	200	500	700	1000
Pentanal	0	200	500	700	1000
Hexanal	0	200	500	700	1000
Heptanal	0	200	500	700	1000
Octanal	0	200	500	700	1000
Butan-2-one	0	100	300	400	600
Hexan-2-one	0	100	300	400	700
Heptan-3-one	0	100	300	400	700

.

The calibration curves were obtained by plotting the peak area ratios (PARs) between the analyte and IS quantitation ions versus the nominal concentration of the calibration solution. A linear regression analysis was applied to obtain the best-fitting calibration curve. The limits of detection and quantitation (LOD and LOQ) were calculated according to the ICH guidelines, using the approach based on the standard deviation of the blanks and slope of the regression [42].

2.3. Experimental Plan on MEMS Preconcentrators

The data were collected using Microsoft Excel and processed using the Chemometric Agile Tool (CAT), an open source and R-based software (3.0.0) [43]. The application of a D-Optimal design allowed for the evaluation of the effect of four factors on MEMS preconcentrators using MEMS trapping materials with a multivariate approach, only performing nine experiments. From the preliminary results, it was possible to exclude the possibility—compared to the ones of the quinoxaline-bridged cavitand (QxCav)—of using both Carbograph 1, with a far inferior performance, and the APCF, which was too fragile. Moreover, the sampling times higher than 100 s were not considered during the optimization of the method, because this would have enhanced the duration of the sample analysis too much. To optimize the analysis, the effects of four factors were studied in three levels each: the sampling time (x₁) at 25, 50, and 100 s, the flow rate sampling (x₂) at 250, 350, and 450 mL/min, the desorption time (x₃) at 10, 40, and 80 s, and the desorption temperature (x₄) at 120, 140, and 180 °C. To build the model, the interactions among the factors were omitted, since the authors already knew that the variables considered are naturally linked. The relative experimental plan with the experiments performed is reported in

Table 3. Experimental matrix with the corresponding experimental plan for MEMS preconcentrators.

	Experimental Matrix				Experimental Plan
Exp#	x1	x2	x3	x4	Sampling Time (x1)	Flow Rate Sampling (x2)	Desorption Time (x3)	Desorption Temperature (x4)
1	0	−1	−1	−1	50 s	250 mL/min	10 s	120 °C
2	1	1	0	−1	100 s	450 mL/min	40 s	120 °C
3	−1	0	1	−1	25 s	350 mL/min	80 s	120 °C
4	1	0	−1	0	100 s	350 mL/min	10 s	140 °C
5	−1	−1	0	0	25 s	250 mL/min	40 s	140 °C
6	0	1	1	0	50 s	450 mL/min	80 s	140 °C
7	−1	1	−1	1	25 s	450 mL/min	10 s	180 °C
8	0	0	0	1	50 s	350 mL/min	40 s	180 °C
9	1	−1	1	1	100 s	250 mL/min	80 s	180 °C

.

To optimize the method, three responses were selected. The intensities, as PID signals reported in mV, were selected for three key analytes, namely hexanal (y1), butanal (y2), and octanal (y3) (as shown in Section 3.2. The aim was to maximize the sensitivity in terms of the intensities for all the analytes. Each experiment was performed in triplicate and the models were computed at the concentration of 50 ppb.

2.4. Greenness Evaluation

The environmental friendliness of the analytical method was evaluated by applying the GAPI tool [44] (moGAPI 2024)(performed with the software available in [45]). The GAPI was utilized to highlight that the method is safer for the operators and complies with green chemistry principles. Each pentagram of the GAPI [44] represents a specific factor: the pictogram represents a step of the analytical protocol and the colour scale from red to green indicates a high or low environmental impact.

2.5. Model Specification and Fitting

Channelled emissions of an Asphalt Mixture Plant were evaluated for OC determination (n = 15) using dynamic olfactometry following standardized procedures in accordance with EN 13725:2022 [3]. VOC concentrations were quantified using Pyxis GC, and the temperature (°C) was recorded concurrently. The sample size was selected to ensure the adequate representation of emission variability and to comply with the 30 h analytical time window mandated by the standard for odour emission evaluation (EN 13725). The Odour Activity Value (${OAV}_i$) [46] was computed by dividing its VOC concentration ($C_i)$ by its established Odour Threshold (${OT}_i$) [47]:

$${\displaystyle \frac{C_i}{{OT}_i}}={OAV}_i$$

The Sum of (${OAV}_i$) (SOAV) represents an aggregate predictor variable uncorrelated with other predictors of interest such as the process temperature.

$$\sum _{i=1}^n{\displaystyle \frac{C_i}{{OT}_i}}=SOAV$$

A GAM [48] was formulated to predict odour emissions using a Gamma family distribution of the response variable with a logarithm function link.

$$\log \left(Odour Concentration\right)={\beta }_0+{\beta }_{1}SOAV+f(Process Temperature)$$

Diagnostic Evaluation

Basis Dimension (k) Check: A diagnostic (commonly referred to as the “k-index”) was employed to verify that the chosen spline basis dimension was sufficient. An excessively low k would over-smooth the data, while an excessively high k could lead to overfitting.

Goodness-of-Fit Metrics: The Deviance-Explained and Adjusted R² measured how much of the odour emission variability was captured by the model.

GCV Score: This provided a comparative index for model complexity.

Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC): These were also monitored to ensure an appropriate balance between model fit and parsimony.

To check the collinearity and the concurvity and to evaluate the potential multicollinearity between the predictors included in the GAM, the Pearson correlation coefficient and the concurvity measure were captured. The Pearson correlation matrix for the predictors SOAV and process temperature revealed a correlation (r = 0.7), indicating a certain degree of linear association between these variables. Furthermore, the concurvity in the GAM was assessed. A posterior predictive check—where the empirical data densities are compared to model-predicted densities (Figure 2)—helped confirm that the model approximates the observed odour emissions reasonably well.

3. Results and Discussion

To date, there are few analytical methods [49] for the determination of the odour-active compounds of HMA; the gap especially occurs for in-field monitoring. A person-portable MEMS-based micropacked thermal desorption GC/PID system was evaluated as a potential alternative to conventional methods for the determination of odorous compounds, offering a simple, rapid, sensitive, and solvent-free approach for HMA fingerprinting. These compounds are represented by aldehydes and ketones [48], having an Odour Threshold (OT) that decreases markedly to sub-ppb as the number of carbon atoms increases.

3.1. Analytical Set up Optimizatiom: Sorptive Material Evaluation and Method Calibration

The sorptive material of the preconcentrator must selectively adsorb one or more molecules of interest over a time necessary to concentrate the chemical compound in the adsorptive material [41]. The optimization of the adsorption and desorption performance—established between a suitable adsorbent, low power consumption, and simple fabrication technology—is crucial for achieving a high preconcentration factor, and is then able to be refocused at the head of the GC column. Subsequently, the sorptive layer is rapidly heated to induce thermal desorption, generating sharp peaks with relatively high analyte concentrations directed toward the connected sensor or detector. This process enables the purification and preconcentration of analytes from a large air volume, thereby enhancing detection efficiency.

The dimensions of MEMS-based micropacked technology in relation to preconcentrators led us to consider the comparison of three unconventional trapping materials, namely APCF, QxCav, and Carbograph 1. Paris et al. [40] point out the surface area of granular-active carbon (usually of about 1000 m²/g) versus fibrous APCF (2000 m²/g, density of 0.095 g/cm³, and fibre diameter 10 μm), and compared tubes in APCF and classic multilayers tubes (Carbograph 1TD plus Carboxen 1003), showing that APCF provides better results with lower standard deviations when compared to others, even with different sampling volumes. The Authors [40] also note that APCF, being a fibre, allows the tubes to be packed more homogenously for sampling than the current activated granular-active carbons, resulting in a superior adsorption/desorption capacity. The selective complexation of aromatic hydrocarbons by QxCav, driven by multiple π–π and CH–π interactions between the aromatic guest and the deep, hydrophobic cavitand cavity [50,51,52], has been leveraged to develop a low-cost system capable of sub-ppb detection limits for VOCs, even in the presence of other airborne pollutants and with high resistance to humidity [53].

Although aldehydes and ketones generally require, for trace analysis, a derivatization of the carbonyl group or the use of mass detectors when considered as such [54,55], the miniaturized PID (10.6 eV) has proven in lab tests to be particularly sensitive (

Table 4. Substances, information about IE and eV, and limits of quantitation (LOQ) in µg/m³ for each compound.

Chemical Name	CAS	IE, eV	Lamp Type (RF)	LOQ
			10.6 eV	µg/m3
Butanal	123-72-8	9.86	1.7	5.0
Pentanal	110-62-3	9.74	1.5	5.3
Hexanal	66-25-1	9.72	1.2	4.9
Heptanal	111-71-7	N.D.	N.D.	5.4
Octanal	124-13-0	N.D.	1.10	5.8
Butan-2-one	78-93-3	9.51	0.96	2.8
Hexan-2-one	591-78-6	9.34	0.8	3.3
Heptan-3-one	106-35-4	9.02	0.73	3.4

) for the molecules considered in this work, and capable of generating a particularly significant HMA fingerprint.

Graphitized Carbon Blacks (GCBs), such as Carbograph, Carbopack, and Carbotrap, are broadly utilized non-selective carbon sorbents in trace-level VOC analyses, primarily due to their low artefact production. Uniquely among hydrophobic materials, they offer performance comparable to that of certain molecular sieve adsorbents. GCBs exhibit a spectrum of adsorption strengths—from very weak to medium and strong—with the most retentive forms displaying some degree of microporosity, thereby positioning them as intermediates between standard GCBs and carbonized molecular sieves. Carbograph 2SAP has proven to be excellent for the rapid separation of ultra-volatiles and volatile C2-C7 VOCs, as shown in Section 3.3.

3.2. Chemometric Analysis

As already mentioned in Materials and Methods, a D-Optimal design was applied to investigate all the possible interactions among the variables involved during the sampling and detection of three aldehydes with the Micro-Electro-Mechanical System, performing only nine experiments. As can been seen from Table 3, all the variables were studied at three levels each, with the sampling time (x₁) between 25, 50, and 100 s, the flow rate sampling (x₂) between 250, 350, and 450 mL/min, the desorption time between (x₃) 10, 40, and 80 s, and the desorption temperature (x₄) between 120, 140, and 180 °C.

The model describing the sensitivity of the analytes (y₁, y₂, and y₃) can be considered parallel, meaning that the experimental conditions that allowed the highest response for the three analytes are the same. This can be easily seen from Figure 3, which reports the response surfaces obtained from the models; in fact, the highest responses for all the analytes can be achieved when the factors x₂, x₃, and x₄ are at the lowest level (−1) and the factor x₁ is at the highest level (+1).

So, the optimization of the analytical method can be obtained when working with a sampling time of 100 s (x₁ = 1), a flow rate sampling of 250 mL/min (x₂ = −1), a desorption time of 10 s (x₃ = −1), and a desorption temperature of 120 °C (x₄ = −1). The same results can be assessed by comparing the contour plots elaborated from the models for the intensities of the aldehydes. For example, in Figure 4, the contour plots regarding the factors x₂ and x₄ for the three analytes in different colours are reported. Also in this case, it is easy to indicate that to maximize the analyte responses, the two variables must be set at the lowest level (−1). The same can be carried out with the contour plots for the factors x₁ and x₃ (not reported), with the only difference that, in this case, to maximize the sensitivities, the factor x₁ must be set at the highest value (1), and the other one at the lowest value (−1).

3.3. Greenness Evaluation

As can be easily assessed in Figure 5, this method has a complete green character, with the sample storage and pre-treatment being unnecessary, having a solvent-free method, and requiring a low amount of energy for each sample. The use of filtered ambient air as a carrier gas results in no need for carrier gas cylinders, further reducing the operational costs and environmental impact on the proposed MEMS-based GC/PID system.

The first pentagram on the bottom left (A) refers to the sample collection and storage, while the pentagrams on the upper left (B) and the upper right (C) account for the sample pre-treatment and the amount of reagent/solvent used. The last one, (D), on the bottom right, instead involves the quantity of generated waste, the hazard for the operator, and the energy requested for the analysis.

The evaluation of the green character of an analytical protocol has gained great importance; in our opinion, the developed method represents the gold standard for this kind of application and it can provide a benchmark against which to compare similar analytical scenarios.

3.4. Method of Field Application

On-field sampling was conducted with the developed instrument and settings for fifteen days (temperature ranged from 18 to 22 °C, and relative humidity ranged from 60 to 79%) directly at the source, specifically at the stack within a production facility. The data presented in

Table 5. Descriptive statistics of substances identified at the chimney level.

	Butanal	Butan-2-One	Pentanal	Hexanal
	μg/m3
Average	227.62	117.39	129.97	153.01
Dev. Std	113.91	58.58	69.28	85.23
Median	224.28	115.91	123.99	140.73
	Hexan-2-one	Heptanal	Heptan-3-one	Octanal
	μg/m3
Average	62.98	105.7	52.79	46.27
Dev. Std	41.94	74.01	57.15	34.88
Median	45.08	73.04	19.46	42.1

and in the graph (Figure 6) below are derived from these sampling activities. From the data obtained in the field, a characteristic fingerprint of the HMA can be observed, like the one obtained in the training phase of the instrument with aldehydes and ketones, which represent the most present compounds in the profile, with particular attention on hexanal.

3.5. GAM Approach

This study demonstrates the utility of a GAM approach in modelling odour emissions, particularly when the data exhibit non-linearity and collinearity issues among predictors. Although the sample size used in the GAM was limited (n = 15), it was considered sufficient to capture variability in the odour concentration, while respecting the 30 h analytical window mandated by EN 13,725 [3]. The model incorporated penalized thin-plate regression splines with a low basis dimension (k = 3) to avoid overfitting. Additionally, goodness-of-fit diagnostics, concurvity checks, and posterior predictive evaluations confirmed the model’s ability to approximate observed odour patterns despite the limited dataset. While the relatively small sample size may limit the generalizability of the findings to broader populations or other industrial settings, the model still provides robust insights within the scope of the observed operational conditions. The constrained complexity of the smooth terms and the use of an appropriate distribution (Gamma family with log link) contributed to a stable model fit, suggesting that the approach remains valid for exploratory and predictive purposes in similar contexts. The fitted Gamma model with a log link provided meaningful insights into how the overall SOAV and temperature can be used to predict odour emission rates. SOAV was computed as the sum of the OAV values of butanal, pentanal, hexanal, heptanal, octanal, 2-butanone, 2-hexanone, and 3-heptanone.

The Shapiro–Wilk test confirmed that the response variable, odour concentration (OC), was not normally distributed (p-value < 0.01). Scatterplots between OC and each predictor (i.e., VOC concentrations and process temperature) did not reveal any clear linear trends. The processing temperature refers to the actual temperature recorded at the point of emission during HMA production, as measured concurrently with odour sampling using the Pyxis GC system. This temperature is a key operational variable affecting the volatilization of odorous compounds.

Given the known influence of temperature on the volatilization of odorous compounds, and its potential non-monotonic effects, we opted for a Generalized Additive Model (GAM) that incorporates a spline-based smooth term for the process temperature. This allows for the flexible fitting of complex, non-linear relationships such as thresholds, plateaus, or saturation effects.

To validate the shape of the non-linear trend, we compared the GAM fit with a LOESS (Locally Estimated Scatterplot Smoothing) model—a non-parametric smoother widely used for exploratory data analysis. The purpose of this comparison, shown in Figure 7, is to ensure that the GAM, despite its more structured statistical formulation, aligns well with the trend revealed by the data-driven LOESS curve. The consistency between the two models strengthens the confidence in the GAM’s ability to accurately capture the underlying temperature–emission relationship.

Thus, we relied on a GAM to build a regression model that utilizes the process temperature and SOAV.

The response variable, OC, fit a Gamma distribution, reflecting the continuous and strictly positive nature of odour emission data and accounting for possible heteroskedasticity. A logarithmic link ensures that predictions remain positive. The Sum of Odour Activity Values (SOAV) was included in the model as a parametric linear term, as its relationship with odour concentration (OC) did not exhibit non-linear behaviour. In contrast, the process temperature was modelled using a penalized regression spline, denoted as f(), to account for its non-linear effects on odour emissions (Figure 6). A thin-plate regression spline with a moderate basis dimension (k = 3) was employed to ensure a balance between model flexibility and overfitting risk, with Generalized Cross-Validation (GCV) scores used to optimize the spline’s complexity. Concurvity analysis revealed a worst-case value of 0.8, suggesting a relatively strong dependency between predictors, indicative of potential collinearity. However, the concurvity values specific to the smooth term for temperature remained in the moderate range (0.4–0.5), supporting the model’s robustness.

The SOAV term exhibited a statistically significant positive effect on odour emissions (p < 0.05), indicating that as the combined odorant load increases, the overall emission rate also rises. The process temperature (smooth term) showed a non-linear relationship (p < 0.001), with an estimated effective degree of freedom of approximately 1.9, suggesting a near-quadratic trend. The model indicated an initial rise in emissions with increasing temperature, followed by a plateau or slight decline at higher temperatures (Figure 8).

Figure 7 and Figure 8 reinforce the predictive reliability of the GAM and emphasize the process temperature as a manageable operational variable for odour mitigation. The results we obtained are shown in Figure 9, where prediction values and observed odour concentrations are confronted. Overall, the model has a Mean Absolute Prediction Error (MAPE) of approximately 39%. Thus, this model can be used to obtain an estimate of the odour impact at the receptor by predicting the estimate of the odour concentration emitted from channelled emissions via Pyxis GC, and using it as the input of dispersion models such as AERMOD [56] or CALPUFF [57].

Residual plots were inspected to check for patterns or trends that might suggest misspecification in the model (e.g., leftover heteroskedasticity or unmodeled interactions). The residuals were normally distributed, thus, no patterns were found that could imply the misspecification of the model.

4. Conclusions

Micro Gas Chromatography systems have experienced significant growth over the past two decades, evolving into a powerful analytical tool for the real-time monitoring of gaseous emissions. The method developed in this study, based on a MEMS-enabled micro GC/PID system, represents a major innovation in the field of odour emission analysis. Its novelty lies in the integration of multiple technical components that have not previously been used together: miniaturized and directly heated MEMS devices for both the preconcentrator and the separation column; a hydrophobic quinoxaline-bridged cavitand as a highly selective sorbent material; and graphitized carbon (Carbograph 2SAP) as the active stationary phase within the MEMS GC column. Moreover, the system operates using filtered ambient air as the carrier gas, enabled by the oxygen tolerance of both the cavitand and the graphitized carbon under operational conditions.

The instrument, whose preconcentration process was optimized via D-Optimal experimental design, was tested on real HMA samples, and proved capable of generating detailed emission fingerprints composed predominantly of aldehydes and ketones—compounds known for their low Odour Thresholds and significant olfactory impact.

Concerning the greenness, the method demonstrated excellent environmental compatibility, as evaluated by the GAPI, confirming its minimal ecological footprint and alignment with green chemistry principles.

Finally, the field data acquired were incorporated into a GAM, which effectively captured the non-linear relationship between the odour concentration, process temperature, and the Sum of Odour Activity Values (SOAV). Despite a limited sample size, the model showed predictive capacity, underscoring the potential of micro GC systems not only as analytical tools, but also as integral components of environmental modelling workflows. This approach paves the way for the deployment of advanced, miniaturized instrumentation in environmental monitoring, regulatory compliance, and proactive odour management strategies.