Intelligent wearable olfactory interface for latency-free mixed reality and fast olfactory enhancement

The study addresses the limitations of existing olfactory interfaces—namely perceptible latency (second-level response times), bulkiness, and limited odor channels—that hinder immersive applications such as VR/AR/MR and clinical olfaction training. Olfaction is critical to human interaction with the environment, influencing hazard detection, social behavior, mood, and memory. Despite this importance, olfactory display technologies have lagged behind haptics, leading to technical stagnation. The authors propose a wearable, miniaturized olfactory interface using compact odor generators (OGs) and AI to achieve millisecond-scale response, low power, and high channel density for latency-free mixed reality and efficient olfactory enhancement, spanning applications from entertainment and education to therapy and human–machine interfaces.

Prior wearable olfactory interfaces rely on arrays of odor generators with mechanisms such as piezoelectric atomizers and controllable phase-change odorous wax. These systems typically exhibit second-level response times, causing noticeable latency during use. State-of-the-art odor generation often requires bulky setups, limited odor options, and expert operation, constraining accessibility and portability. Compared with extensive progress in haptic feedback technologies, olfactory technologies have received less attention, contributing to slow advancement. Performance gaps identified in the literature include slow response, high power consumption, large form factor, and low channel counts, motivating the need for sub-second response, centimeter-scale form factors, tens of OGs per array, and low power consumption for true wearability and responsiveness.

Device design and operation: Miniaturized odor generators (OGs) use a multilayer stacked structure: top PET (0.1 mm) with a 4×3 breathing-hole array as the odor outlet; a mechanical actuation unit comprising a 0.5 mm-thick magnet and an underlying Cu coil (300 turns, 50 µm wire) to electromagnetically deflect a PET cantilever and open/close the holes; an odorous cotton/adhesive layer (0.1 mm) as a chemical reservoir; a PI (2 µm)/Au (200 nm)/Cr (40 nm) heating platform with a 1×0.5×0.5 mm thermistor to control temperature (35–55 °C); and a 3D-printed epoxy ring as support. The OG functions as a solenoid valve to modulate odor release rate by opening/closing breathing holes and uses the heater to control odor generation rate via temperature. A custom control panel supplies 1.7 V (coil) and 9.0 V (heater), measures temperature via ADC, and uses decoders to convert DC to AC for programmable magnetic actuation (duty cycle, amplitude, frequency). Mechanical and electrical characterization: Finite element analysis confirmed strains below elastic limits under 8.8 kPa external pressure and during 29° bending. Electrical optimization studied breathing-hole patterns, heating temperature, and AC drive parameters (frequency, amplitude, duty cycle). Ethanol (95 vol%) served as the test odorant measured by a commercial sensor (TGS823). Temperature was stabilized at setpoints (35–55 °C) with small fluctuations, and response times to reach setpoints were recorded. Odor concentration dependence on heating temperature, coil-drive amplitude, duty cycle, and ambient airflow (2.76–6.61 m/s) was measured. Response time for odor termination was assessed by closing the breathing holes and monitoring ethanol decay. Leakage tests with holes closed were performed across temperatures and distances. Wearable systems and control electronics: Flexible PCB-based two-channel and 32-channel control panels enable wireless manipulation (Bluetooth) of OGs, independent duty-cycle control, and temperature regulation. Power was provided by Li-ion batteries (500–1500 mAh) with full-load operation benchmarks. The 32-channel forearm-mounted interface uses PDMS encapsulation and FPCB interconnects; a microcontroller (ATMEGA328P-MU), shift registers (SN74HC595), multiplexers (MAX4691), and a 24-bit ADC (ADS1220) manage coil actuation and temperature sensing. Mixed reality (MR) algorithm and positioning: A positioning system (DWM1000) tracked real-world odor source (OS) coordinates, mirrored in Unity for MR; VR headset tracked user location and heading. A latency-free strategy predicted user motion and commanded OG activation to ensure odor threshold is reached when the user arrives. Odor dispersion was modeled with Gaussian plume dispersion (for horizontal wind cases) and asymmetric diffusion with convective vertical flow (for still-air cases), using Briggs dispersion parameters and Pasquill stability classes. The system triggers OGs when the potential distance from odor plume to user is within the user’s exploration potential accounting for OG response time and transport time. Olfaction training protocols: Two training paradigms were used. (1) Random Odor Generation (ROG): volunteers wore the 32-channel device and identified randomly presented odors (32 types across fruits, spices, drinks, bakeries, herbs, others). Experimental group had a 1-hour pretraining (2 minutes per odor) prior to testing; control group did not. Recognition rate and reaction time were recorded over three sessions. Standard clinical olfaction test concepts (e.g., UPSIT, CCCRC, BAST-24, Sniffin’ Sticks) motivated metrics. (2) AI-driven List-wise Recommendation Framework (LRF): a reinforcement learning agent modeled as an MDP recommended per-odor training time allocations to maximize expected cumulative reward emphasizing final recognition rate. Ten volunteers trained and were tested daily for seven days under LRF. Fabrication: Heating electrodes were fabricated on PI via sputtered Au/Cr and photolithography; PI encapsulation layers were patterned by RIE. Thermistors were attached with silver paste and glue; cotton, coils, PET chambers, and magnets were assembled with adhesives and 3D-printed epoxy rings. Flexible circuits were fabricated on Cu/PI with gold plating; components (Bluetooth, MCU, passives, crystal, OGs) were soldered; PDMS encapsulation was cast and cured. User studies and assessments: Volunteer tests included 10-person sensory evaluations for breathing-hole optimization and odor thresholds, 25 volunteers for training efficacy (11 experimental with pretraining; 14 controls), and a 10-volunteer cohort for LRF evaluation over seven days. A post-COVID-19 anosmia patient underwent a two-week personalized training protocol. Mental health was assessed pre/post training via PHQ-9 and HADS.

- Performance of odor generators (OGs): - Millisecond-level control: odor termination response time reduced to 70 ms by closing breathing holes via electromagnetic actuation. - Temperature control: stable heating at 35–55 °C with small fluctuations (±0.33 °C at 35 °C to ±0.72 °C at 55 °C); 12-hour stability at 55 °C with ±0.15 °C fluctuation and 21,600 open/close cycles. - Odor concentration relationships: peak ethanol concentration increases linearly with temperature (e.g., ~2102 ppm at 35 °C to ~2705 ppm at 55 °C). Increasing coil-drive amplitude raises concentration from 635 ppm (20.9 mW) to 2689 ppm (84.8 mW). Higher duty cycles yield higher steady concentrations with corresponding stabilization times: 1047 ppm (20% duty, 0.99 s), 1575 ppm (40%, 3.26 s), 1992 ppm (60%, 5.09 s), 2717 ppm (80%, 7.58 s). AC frequency (0.5, 1, 2 Hz) had negligible effect on peak concentration (~2717 ppm) with similar temperature stability (±0.6–0.7 °C). - Airflow effects: surrounding airflow reduces concentrations to 422 ppm (2.76 m/s), 339 ppm (4.12 m/s), 259 ppm (6.61 m/s) from ~2717 ppm at 0 m/s. Closing holes sharply decreases concentration; higher ambient airflow accelerates decay. - Leakage with holes closed: at 0 cm distance and 55 °C, maximum leakage ~255 ppm (below human identification threshold of 350 ppm); at 1.5 cm, leakage 49–62 ppm, below perception; during release, ~1491 ppm at 1.5 cm. - Optimized breathing-hole pattern (lowest solid area ratio 0.914) produced highest stabilized ethanol concentration (~2649 ppm) and highest perceived intensity in volunteers. - OG array density up to 0.75 units/cm² using corner-to-corner layout with 2 mm height gap; magnets operate normally when center-to-center distance >1.5 cm. - Wearable and system-level metrics: - Two-channel controller powered by 500–1500 mAh Li-ion batteries supports full-load operation: ~2.7 h (500 mAh), 5 h (900 mAh), 5.5 h (1500 mAh). Odor reservoir (0.04 mL per OG) yields 0.26–2.7 h operation depending on odorant. - 32-odor support spanning fruits, spices, drinks, bakeries, herbs, and others; wiring concealed in realistic models; wireless control via Bluetooth; MR integration with positioning system (slope 0.998 vs ruler). - Latency-free MR: - AI algorithm predicts user motion and odor transport using Gaussian plume/convective models, commanding OGs so odor reaches threshold (e.g., 5 ppm in demo) upon user arrival. Compared to prior OG with 1.4 s response, current OGs reduce unnecessary operation and energy consumption (e.g., OS2 operation 4.1 s vs 15.1 s prior system) and eliminate perceived latency. - Olfaction training and enhancement: - ROG protocol: with 1-hour pretraining, experimental group’s average recognition rate reached 0.70 with reaction time 6.2 s at third test; controls without pretraining averaged 0.41. More training sessions improved recognition and reduced reaction times (e.g., 0.59@9.0 s → 0.65@8.1 s → 0.70@6.2 s across three sessions). Mental health measures improved modestly (HADS A 5.5→5.2; PHQ-9 3.2→2.3). - LRF (AI recommendation) over 7 days in 10 volunteers improved category-wise recognition rate and reduced training time: fruits 0.47@18.75 min → 0.90@6.10 min; spices 0.58@11.25 min → 0.87@5.10 min; drinks 0.44@9.38 min → 0.80@5.10 min; bakeries 0.44@9.38 min → 0.86@4.10 min; herbs 0.33@5.62 min → 0.83@2.60 min; others 0.70@5.62 min → 0.97@0.80 min. Near-saturation on day 5 (e.g., fruits 0.81@7.10 min; spices 0.72@6.70 min; drinks 0.72@6.60 min; bakeries 0.86@6.10 min; herbs 0.83@3.40 min; others 0.97@2.90 min). LRF improved recognition faster than ROG (fitted slope 0.54 vs 0.40). - Post-COVID-19 anosmia case: recognition improved from ~0.19–0.25 (days 1–2) to 0.38@19.61 s → 0.72@11.83 s (days 3–6) and 0.78@9.42 s → 0.97@3.8 s (days 7–14). Without training (tri-weekly checks), performance stabilized around 0.63–0.72 recognition and 8.78–14.49 s reaction time.

By introducing a compact solenoid-valve OG with electromagnetic actuation of breathing holes and a controlled heating platform, the system overcomes longstanding challenges of olfactory interfaces: perceptible latency, bulk, and limited channels. Millisecond-scale control of odor release (70 ms) eliminates delay between user action and olfactory perception in mixed reality when paired with predictive AI and plume transport modeling. Electrical optimization demonstrates precise, low-jitter temperature control and parameterized concentration control via duty cycle and amplitude, enabling rapid, energy-efficient modulation without relying on slow thermal transients. System-level demonstrations confirm zero-latency MR experiences and reduced operational overhead versus prior OGs with second-level response. In olfactory training, the 32-channel wearable enables rapid switching among many odorants directly on the forearm, significantly improving recognition rates and reducing reaction times with even brief pretraining. The reinforcement learning-based list-wise recommendation further accelerates learning by personalizing training time allocation per odor, achieving near-saturation within days. The approach also benefits clinical recovery, as evidenced by the post-COVID-19 anosmia case. Collectively, the findings establish a general, wearable platform bridging olfaction, AI, and MR to enhance user immersion, education, and therapy.

The work presents AI-driven, wearable olfactory interfaces built on miniaturized odor generators that achieve millisecond-level response, precise concentration control, stability over prolonged operation, and high array density with 32-odor support. Integrated with predictive modeling and positioning, the system enables latency-free mixed reality and efficient, personalized olfactory training, demonstrating substantial improvements in recognition, reaction time, and potential psychosocial benefits. The platform opens a path toward establishing olfaction as an information channel in VR/AR/MR metaverse, remote education, and clinical treatment. Future research directions include deeper integration with physiology and neuroscience for adaptive biofeedback, exploration of robust odor blending strategies, reduction of auxiliary hardware bulk (e.g., positioning modules) for fully skin-mounted MR, and material advances to further lower power and size while expanding odor libraries.

- Positioning system bulk: The commercial positioning module (DWM1000) is too large for seamless integration with skin-mounted olfactory interfaces, preventing fully wearable, latency-free MR in such configurations. - Odor blending: Attempts to blend two odors into a new stable percept were inconsistent (probability 0.19–0.64) due to variable distances from OGs to nose, differing diffusion rates, and ambient airflow; the system was configured to emit one odor at a time for training. - Gas leakage: With breathing holes closed, minor ethanol leakage (~255 ppm at 0 cm and 55 °C) can be perceivable; users are advised to keep >1.5 cm distance to avoid perceiving leakage (49–62 ppm at 1.5 cm). - Control via heating temperature: Adjusting concentration by changing heater temperature incurs second-level response (>4 s), so rapid control relies on duty cycle and amplitude modulation of the electromagnetic actuation. - Modeling simplifications: Odor transport modeling neglected turbulent airflow and used time-averaged plume envelopes under slow laminar conditions; performance in complex, turbulent environments was not addressed. - Measurement choice: Ethanol was used as the test odorant for sensor compatibility; generalization to other odorants’ quantitative concentration profiles was not directly measured with the same instrumentation.