Abstract

The photonic nose is an emerging class of optical sensing systems designed to mimic the olfactory capabilities of a human nose. Evolving from conventional chemical and gas sensors, photonic noses leverage optical phenomena to achieve high sensitivity and fast, label-free analysis of chemical volatiles. This review provides an in-depth analysis of the evolution and current state of photonic nose technologies, particularly focusing on their integration with artificial intelligence (AI) and machine learning (ML). We first discuss key optical sensing and fabrication methods, including colorimetry, refractive index sensing, spectroscopy, and integrated photonic devices. Then, the role of ML algorithms in photonic noses is highlighted, and the integration of photonic noses into cloud-to-edge computing systems is also explored, demonstrating intelligent microsystem designs capable of on-chip real-time analytics and distributed data processing. Additionally, we highlight representative application scenarios where AI-driven photonic noses show significant advantages, including environmental monitoring, early-stage medical diagnostics, and ensuring food quality and safety. A concise comparative analysis between photonic noses, electronic noses, and analytical instruments is provided. Finally, this review identifies the remaining challenges in AI-driven photonic noses and offers insights into future development pathways toward smarter, miniaturized, and more robust photonic sensing systems.

Introduction

The ability to detect and identify complex odors and gaseous analytes is crucial in a wide range of fields, including environmental safety, healthcare, food science, and national security1,2,3. Traditional gas sensors and electronic noses (e-noses) have long been used to sense volatile compounds4. These e-noses typically consist of an array of chemical sensors (e.g., metal-oxide semiconductors5, conductive polymers6, quartz microbalances7) with partially overlapping selectivity, coupled with pattern recognition algorithms that interpret the resulting “smell” fingerprints. While successful in many applications, classical e-noses face persistent challenges8, including sensor drift, limited selectivity, and susceptibility to environmental factors like humidity and temperature.

Over the past decade, photonic nose technology has emerged as a powerful complement to conventional e-noses9,10. Rather than relying on purely electrical or chemical transducers, a photonic nose uses optical sensors and spectroscopy to detect chemicals based on their interaction with light11. By measuring unique absorption spectra or refractive index changes, photonic noses can achieve higher sensitivity, faster response, and improved long-term stability12. In essence, photonic noses mimic the biological olfactory system using photonic devices where multiple optical sensing elements act as olfactory receptors, capturing a chemical fingerprint that algorithms then interpret13,14, like the brain processes olfactory signals.

Early demonstrations of photonic noses were built on advances in optical gas sensing, such as non-dispersive infrared (NDIR) sensors and laser absorption spectrometers15,16, which target characteristic infrared absorption lines of specific gases (e.g., CO2 or CH4). Currently, the photonic noses concept has been expanded to integrate numerous optical sensors on a single platform to detect more complex mixtures and a wider variety of volatiles. For instance, integrated photonic chips can combine arrays of waveguides or microresonators, each designed or functionalized for particular analytes17,18. When exposed to an odor sample, these sensing elements yield an aggregate optical response (an “odor fingerprint”) that can be matched against a reference library19. Advances in photonic integration and microfabrication are making these systems increasingly compact and scalable, steering them toward practical intelligent microsystems that integrate both sensing and processing on a single chip.

Crucially, the rise of the photonic nose has gone hand in hand with developments in artificial intelligence (AI) and machine learning (ML). Just as the human brain interprets neural signals from olfactory receptors, AI algorithms are now harnessed to analyze the complex optical signals from photonic sensors20,21,22,23. Machine learning has proven indispensable for improving gas sensing selectivity and correcting sensor drift24,25. AI-enhanced photonic noses can resolve subtle differences in chemical mixtures, untangle overlapping spectral signals, and even quantify individual components in a multi-analyte sample26,27. Alongside this progression, improvements in computational resources have popularized a distributed approach to data processing. Heavy computations, such as large-scale analytics or model training, can be performed in the cloud, while lightweight AI models can operate on-edge (adjacent to the sensor) for real-time detection28,29,30,31,32. Despite these advances, AI-enabled photonic noses are still in their infancy, and merging AI with photonic nose technologies presents numerous complexities. These include issues such as sensor integration and scalability, the need for comprehensive odor datasets, and the challenge of achieving real-time, low-power operation. Significant progress has been made in addressing these challenges, though diverse technical approaches and perspectives still remain.

This review aims to summarize key advancements in the evolution of photonic nose technology, particularly emphasizing the complexities of integrating AI into optical sensing systems (Fig. 1). First, we introduce the core technologies underpinning photonic nose systems, focusing on the fundamental principles of optical gas sensing. Next, we examine the various AI models and architectures used to process photonic nose signals, detailing the evolution from post-sensing intelligence to cloud-based and edge-based processing methods. Subsequently, we explore representative applications in environmental monitoring, medical diagnostics, and food quality assessment, illustrating how post-sensing, cloud, and edge intelligence are tailored to specific use cases. A comparative analysis is then presented to highlight the advantages of photonic noses, e-noses, and other sensor paradigms. Finally, we review the primary challenges remaining in AI-driven photonic sensing and look ahead to the future of photonic noses.

Fig. 1: Roadmap of AI‑driven photonic noses.

Early gas sensors (including colorimeters, refractive sensors, absorptive sensors, and spectroscopic sensors) paved the way for subsequent innovations. With advancements toward high-throughput sensors, distributed nodes, and on‑chip photonic integrated circuits, the post‑sensing intelligence, cloud-based processing, and edge-based intelligence have progressively been realized

Principle and device

Optical sensing principle

Following the long-established history of optical chemical sensing, photonic nose systems detect gases and volatile analytes through direct optical interactions between light and matter. Four major optical sensing mechanisms underpin the core technologies of p-nose systems, namely sensors based on colorimetry, refractive index changes, optical absorption, and spectroscopy (Fig. 2).

Fig. 2: Principles of optical gas sensing.

a Gases consist of various groups (c) and display distinct properties, such as refractive index (n), absorption coefficient (A) and absorption spectrum (α(λ)). b Colorimetric sensing mechanism: the gas groups react chemically with dye molecules, resulting in a detectable color change. c Surface plasmon polariton (SPP)-based methods: these techniques exploit the sensitivity of SPP to variations in refractive index (n). d NDIR methods: detection is achieved by measuring the absorption of infrared light by gas molecules. e Spectroscopic methods: for instance, surface-enhanced Raman spectroscopy (SERS) leverages Raman scattering for molecule detection. f Various output signal types corresponding to these mechanisms are also illustrated

Colorimetric sensors utilize measurable color changes triggered by specific chemical reactions with target gases33,34. Such sensors employ dyes or quantum dots (QDs) as indicators35, exemplified by cadmium telluride (CdTe) QDs (Fig. 2b). Certain gaseous analytes (e.g., formaldehyde) can quench their fluorescence intensity36, and the degree of quenching directly correlates with gas concentration, typically described by the Stern–Volmer equation.

$${I}_{0}/I=1+{K}_{SV}[Q]$$

(1)

where I0/I is intensity ratio of fluorescence before and after exposure in gases, [Q] represents quencher (gas) concentration, and KSV is Stern–Volmer quenching constant. Within the linear range defined by this equation, a simple calibration curve can relate the measured intensity ratio to absolute gas concentrations. However, at higher analyte concentrations or when multiple quenching pathways coexist, deviations from linearity may occur. Therefore, modified mathematical models are required for accurate quantification under such circumstances. It should be noted that the effectiveness of colorimetric sensors relies heavily on the selection and optimization of dyes or quantum dots tailored specifically to the target analyte.

Refractive index sensors detect gases by sensing slight changes in the refractive index of the medium surrounding the optical device37, which is essentially the speed at which light travels. Surface plasmons (SPPs) are a typical application case38. An SPP is an electromagnetic wave bound to the interface between a metal (permittivity εm) and a dielectric (εd = n2). Its resonance condition is set by the in-plane wave-vector:

$${k}_{SPP}=\frac{2\pi }{\lambda }\sqrt{\frac{{\varepsilon }_{m}{\varepsilon }_{d}}{{\varepsilon }_{m}+{\varepsilon }_{d}}}$$

(2)

Where λ is the free space wavelength. Adsorption of gas molecules onto the metal surface changes n (and thus εd), leading to a measurable shift in the resonance wavelength λres. Differentiating Eq. (2) gives the first-order wavelength shift:

$$\Delta \lambda ={S}_{SPP}\Delta n,{\text{and}},{S}_{SPP}={\lambda }_{res}\frac{{\varepsilon }_{m}}{{({\varepsilon }_{m}+{\varepsilon }_{d})}^{3/2}}$$

(3)

Here SSPP = dλ/dn is the bulk sensitivity, typically on the order of several hundred nm per refractive‑index unit (RIU) for noble‑metal films in the visible or near‑IR. According to Eq. 3, when gas molecules bind or aggregate on a metal surface, they cause a measurable shift in the SPP resonant wavelength or optical phase (Fig. 2c). Similar refraction-based sensing approaches include ring resonators and Mach-Zehnder interferometers39, whose resonance or interference conditions are also shifted by changes in the refractive index. Selectivity is commonly enhanced by functional over‑coatings engineered to adsorb target molecules preferentially, such as polymer films, metal–organic frameworks, and nanoporous layers. These coatings magnify the local refractive‑index perturbation Δn and thus amplify the wavelength or phase shift predicted by Eq. (3).

Absorption-based sensors detect gases by measuring the optical absorption intensity of gas molecules at specific wavelengths. A representative example is the NDIR sensor (Fig. 2d)15,40, widely used for detecting gases such as carbon dioxide (CO2) and methane (CH4) at room temperature and 1 atm. This sensor typically consists of an infrared light source and a detector equipped with optical filters designed specifically to isolate the strong absorption bands of target gases. The governing relationship is the Beer–Lambert law and can be expressed in logarithmic form:

$$A\equiv -{\text{log}}\left(\frac{I}{{I}_{0}}\right)=\varepsilon cl$$

(4)

where I, I0, c, l, ε represent incident intensity, transmitted intensity, gas concentration, optical path length, and molar absorptivity (cm2 mol−1 log−1). Typically, the detector measures I0/I, and a pre-calculated calibration curve converts this ratio to a concentration c. Sensitivity can be increased by expanding l or by employing narrowband quantum cascade or light-emitting diode (LED) sources tuned exactly to the strongest molecular line. Due to their robust structure, high stability, and ease of deployment, NDIR sensors provide rapid and reliable quantitative gas analysis without the need for chemical reaction components, making them suitable for diverse environmental conditions. Optical absorption sensors typically exhibit sub-second response times, primarily constrained by gas exchange within the measurement cell and the detector’s electronics. However, variations in temperature and pressure can influence the molar absorptivity (ε); thus, integration of thermistors and pressure sensors is common practice to enable automatic calibration of Eq. (4).

Spectroscopy-based sensors offer comprehensive and powerful optical detection methods41, providing complete absorption or scattering spectra over a broad wavelength range42,43,44,45,46,47. This capability enables accurate chemical fingerprinting, even in complex mixtures. Surface-enhanced Raman spectroscopy (SERS) is one notable example of such technologies (Fig. 2e, f)48. For purely electromagnetic enhancement, the enhancement factors G can be expressed by the ratio of SERS intensity ISERS relative to normal Raman intensity I0:

$$\frac{{I}_{SERS}}{{I}_{0}}\approx {\left(\frac{|E({\bf{r}})|}{|E(0)|}\right)}^{4}$$

(5)

where E(r) is the local electric field at the molecule and E0 is the incident field. The fourth power dependence arises because both the excitation and the re-radiated Raman fields are enhanced by the local surface plasmon. Hot spot field factors |E/E0| (102–103) yield enhancement factors of 108–1012, enabling single molecule detection under optimized conditions. Because each molecule exhibits a unique set of Raman shifts, SERS provides highly selective identification down to trace concentrations. Another important technique is photoacoustic spectroscopy (PAS)49, which exploits periodic thermal pulses produced by gas molecules upon absorption of modulated optical radiation (room temperature and 1 atm). These thermal pulses generate acoustic waves detectable by highly sensitive microphones. For a closed cylindrical cell in the fundamental acoustic mode, the steady state photoacoustic signal SPAS is proportional to the absorption coefficient α(λ): ({S}_{PA}=C{P}_{0}\alpha (\lambda )), where P0 and C represent incident optical power and cell constant. Because α(λ) is directly related to molecular concentration c via Beer–Lambert absorption, quantitative gas analysis is achieved using Eq. (5). With high-power quantum cascade lasers and low-noise micro-electromechanical system (MEMS) microphones, modern PAS systems routinely achieve sub-ppm and even ppb detection limits in palm-sized modules.

Additionally, the application of functional materials has significantly enhanced the performance of optical sensing technologies. Through precise engineering of polymers, metal-organic frameworks (MOFs)50,51, and nanoporous structures, selective capture and enrichment of specific analytes can be efficiently achieved52. Their impact is clear when viewed through sensitivity and selectivity. The affinity of the functional material for the gas increases the local analyte concentration csurf above the gas-phase value c. If the partition constant is K = csurf/c, then the effective sensitivity scales as Seff = K‧S0 (unenhanced sensitivity). MOFs exhibit K ≫ 1 for their target gas, giving a proportional boost in signal. In terms of selectivity, competitive adsorption follows a Langmuir isotherm53, that is, (\theta =\frac{{k}_{ep}p}{1+{k}_{ep}p}), where kep and p are the equilibrium constant of gas and concentration of the gas, respectively. A functional material is tailored to increase the KEP so that the fraction of adsorbent surface covered by the target gas increases. For instance, when polyhexamethylene biguanide (PHMB) is coated on a silicon-based optical resonator, both the sensitivity and selectivity of the sensor are improved.

Fabrication of optical sensors

The fabrication of optical sensors includes bottom-up (colloidal synthesis, self-assembly of nanoparticles, etc.)54,55,56,57,58,59,60 and top-down (lithography-based patterning, etching processes, etc.)61,62,63 nanofabrication. Next, we introduce the fabrication of representative optical sensors in detail.

Quantum-dot-based gas sensors employ nanoscale semiconductor or carbon dots as their sensing layer. Typically, quantum dots (QDs) are synthesized via colloidal chemical methods64,65, using materials such as metal chalcogenides (e.g., PbS, CdSe) or carbon/graphene quantum dots (Fig. 3a)66. After synthesis, various deposition techniques are utilized to fabricate the sensors. For instance, QD inks can be deposited onto interdigitated electrodes through drop-casting or spin-coating67, typically followed by mild annealing to remove residual solvents. QDs may also be incorporated into matrix materials (such as polymer blends, metal-oxide composites, and MOFs) to form robust sensing films68. Compared to traditional semiconductor processes, this solution-based fabrication method is relatively simple and compatible with low-temperature processing.

Fig. 3: Optical gas sensors.

a CdTe quantum dot-based gas sensor. Reproduced with permission from the Springer (2016)66. b SPP-based gas sensor. Reproduced with permission from the Springer (2022)69. c Fiber-based gas sensor. Reproduced with permission from the Multidisciplinary Digital Publishing Institute (2020)73. d NDIR gas sensor. Reproduced with permission from the Springer (2020)40. e SERS-based gas sensor. Reproduced with permission from the Springer (2024)95. f Surface enhanced infrared spectroscopy (SEIRAS)-based gas sensor. Reproduced with permission from the Springer (2023)21. g Integrated waveguide-based gas sensor. Reproduced with permission from the De Gruyter Brill (2021)125

Early surface plasmon polariton (SPP)-based gas sensors commonly employed a configuration involving a thin metallic film deposited on a prism via physical vapor deposition, which supports collective electron oscillations at the metal–dielectric interface (Fig. 3b)69. Higher sensitivity SPP-based gas sensors leverage nanostructured surfaces to enhance surface plasmon resonance (SPR). These nanostructures include metal hole arrays and nanograting structures fabricated via lithography70, or nanoporous gold films obtained through chemical dealloying methods71, all of which enhance gas adsorption and plasmon propagation. Furthermore, metal nanoparticles or nanodisks, deposited via self-assembly, nanoimprinting72, or electron-beam lithography, are meticulously designed in size and shape to tune their resonance wavelength, significantly enhancing sensitivity towards specific gases. Overall, fabrication approaches for SPP gas sensors range from top-down lithography for creating periodic nanoarrays to bottom-up colloidal nanoparticle assembly and templated metallic growth.

Optical fiber gas sensors are based on standard fibers that undergo specialized processing for gas detection. A common approach is to expose the fiber to gases, enabling detection. For instance, when silicon-based anti-resonant hollow-core fibers are exposed to a methane and carbon dioxide mixture, their resonance wavelength shifts (Fig. 3c)73. To enhance the interaction between light and the target gas, several strategies are employed. First, from a materials perspective, a thin gas-sensitive layer (such as a polymer film or nanoporous material) can be coated on the fiber surface to increase gas adsorption74. At the device level, fabricating a tapered fiber section is an effective method75. By heating and stretching the fiber to form a reduced-diameter region, the evanescent field can interact directly with the surrounding gas or nanomaterials deposited on the fiber surface. Additionally, microstructured fibers, such as photonic crystal fibers with air holes76, can serve as gas sensors by allowing gas to permeate the air holes, thereby extending the interaction along the entire fiber length. The fabrication of these fibers involves specialized drawing techniques to form hollow cores or microcapillaries. Overall, the fabrication focuses on post-processing techniques (such as laser grating writing, etching, and side polishing) along with thin-film deposition methods like dip-coating, sputtering, and self-assembly77,78,79,80,81,82.

Conventional NDIR sensors typically utilize discrete components, including broadband infrared sources, gas chambers with reflective inner walls, and detectors such as thermopiles or pyroelectric sensors83,84. Recent advancements have focused on integrating NDIR sensors within MEMS structures85,86,87,88,89,90,91,92. This integration involves on-chip microfabrication of infrared sources (micro-hotplate emitters or infrared LEDs) and miniaturized detectors (thermopiles or microbolometers) coupled with optical filters93. For example, fully integrated MEMS-based CO₂ NDIR sensors have been demonstrated, incorporating MEMS-heated IR emitters combined with on-chip thermopile detectors arranged inside micromachined optical cavities. Such cavities are commonly etched into silicon or alternative substrates, occasionally employing hollow substrate-integrated waveguides or folded optical paths to minimize sensor size94. Optical filters can incorporate metasurfaces, endowing detectors with intrinsic wavelength-selective capabilities (Fig. 3d)40.

Surface-enhanced Raman scattering (SERS) substrates consist of engineered noble metal nanoparticles designed to amplify Raman signals via plasmonic “hotspots”. Traditional preparation methods involve electrochemical oxidation-reduction or chemical etching to create roughened metal electrodes. Alternatively, colloidal silver or gold nanoparticles (10-100 nm) can be deposited onto solid supports, forming interparticle gaps that serve as hotspots and enhance sensitivity (Fig. 3e)95. Advanced top-down fabrication methods, such as electron-beam or nanoimprint lithography, enable periodic structured arrays, although their effectiveness is often less pronounced compared to slightly random, cluster-based structures96,97,98,99,100,101. Template-assisted methods, including deposition onto porous anodized aluminum oxide or salt-crystal templates, provide a viable compromise. Commercial SERS substrates commonly consist of thin evaporated silver or gold films (5-10 nm thickness) deposited onto dielectric supports, which evolve into fractal island-like morphologies ideal for SERS enhancement102.

Typical surface-enhanced infrared absorption spectroscopy (SEIRAS)103,104,105,106,107,108,109 substrates are composed of thin metallic structures deposited onto infrared-transparent supports110,[111](#ref-CR111 “Li, D., Wu, X., Chen, Z., Liu, T. & Mu, X. Surface-enhanced spectroscopy technology based on metamaterials.