Aromaverse: Multi-Domain Olfactory Insights

• Aromaverse is a multi-domain concept unifying immersive XR perfume retail, structured odor perception spaces, and astrochemical aromatic inventories.
• It features a virtual environment with scent blending via hardware-software synchronization and customizable fragrance experiences, enhancing social and purchasing dynamics.
• In computational frameworks, Aromaverse supports molecular generation, odor prediction, and spectroscopic analysis of aromatic species in interstellar clouds.

Aromaverse is a term used in recent arXiv literature for several technically distinct but olfaction-centered constructs. In virtual reality and e-commerce, it denotes a virtual immersive multi sensorial experience: an immersive 3D multiplayer environment augmented with olfactive experience where users can experience and customize perfumes, share customized fragrances, and shop socially in XR (Dubey et al., 9 Sep 2025). In computational olfaction, it denotes the structured space of odor perceptions induced by small molecules, organized by hierarchical odor taxonomies over descriptors (Sajan et al., 11 Aug 2025). In astrochemistry, it denotes the population of aromatic and other cyclic molecules found in cold, dense molecular clouds (McCarthy et al., 2021). This suggests that “Aromaverse” functions less as a single standardized technical object than as a recurrent label for organized odor worlds spanning XR systems, molecular odor space, and aromatic chemistry.

1. Polysemy and conceptual scope

Recent usage of the term falls into three explicit definitions: immersive perfume-centered XR, structured molecular odor space, and aromatic chemistry in astronomical environments (Dubey et al., 9 Sep 2025, Sajan et al., 11 Aug 2025, McCarthy et al., 2021).

Domain	Use of “Aromaverse”	Representative context
XR retail	Virtual immersive multi sensorial experience	Perfume experience, customization, and social buying
Molecular odor modeling	Structured space of odor perceptions induced by small molecules	Hierarchical taxonomy over odor descriptors
Astrochemistry	Population of aromatic and other cyclic molecules in molecular clouds	Cold, dense interstellar chemistry

The XR meaning is directly operational: users enter a retail-like 3D space, sample and customize fragrances, and share olfactive events with companions. The molecular-ML meaning is representational: molecules are embedded into a descriptor space structured by expert or data-driven taxonomies. The astrochemical meaning is compositional: the term names a chemical domain characterized by planar or nearly planar ring systems, asymmetric-top rotational spectra, and cyclization-based formation pathways.

A plausible unifier is systematic organization of odor-related phenomena across scales. In the XR case, the organization is social and interactive; in computational olfaction, it is descriptor- and model-centric; in astronomy, it is molecular and spectroscopic. The literature does not present a single canonical ontology linking these senses.

2. Aromaverse as virtual social immersive multi-sensory commerce

The most explicit system named Aromaverse is a virtual immersive multi sensorial experience for perfume retail (Dubey et al., 9 Sep 2025). Its hardware stack comprises a standalone head-mounted display running the application without tethering to a PC, a custom olfactory display physically connected to the headset via USB-C or Bluetooth, and a Wi-Fi or local-area network link supporting real-time voice chat and synchronized scene updates. The software stack uses Unity3D or Unreal for the virtual retail space, avatar animations, UI panels, and teleport/joystick locomotion; Photon or a similar networking layer for state synchronization and voice chat; and a scent-control API exposing functions such as InitializeDevice(), SetFragranceMix(topNoteID, midNoteID, baseNoteID, α, β, γ), EmitScent(duration_ms), and CalibrateChannel(channelID, intensityLevel).

The data-flow pipeline is specified in four stages. A user taps a virtual perfume bottle and, for customizable items, adjusts three sliders corresponding to top, mid, and base notes. Unity serializes the chosen note IDs and relative weights $(\alpha,\beta,\gamma)$ and broadcasts them to peers. The local scent-control API translates the request into device-specific voltage or pulse-width commands and queues an emission sequence. The hardware then heats or micro-pumps precise quantities of each note cartridge and delivers the blended aroma to the user’s nostrils over a defined time window such as $5\,\mathrm{s}$.

Fragrance representation is reduced to a normalized three-vector,

$F = [T,M,B]^T,$

where $T$, $M$, and $B$ are the relative intensities of top, mid, and base notes in $[0,1]$. The system computes a linear blend,

$$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$

In the reported implementation, the simplest convex combination sufficed, and no higher-order interpolation or volatility dynamics were modeled. Before each session, the three vaporization channels were calibrated through a lookup table

$$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$

followed by a one-point sniff check by the experimenter to verify perceptual linearity.

The virtual environment is organized into scent-zones containing virtual bottles on pedestals. Users choose simple humanoid avatars and display names upon login; avatar gestures signal “scent trial,” “like,” or “share.” Shared olfactive experience is implemented by sending an RPC to all connected peers when one user emits a scent, after which each headset mirrors the emission in real time. A floating panel shown on gaze displays a thumbnail image, current notes, price, and three draggable sliders, alongside “Buy” and “Share” buttons.

Evaluation used a within-subjects protocol with $N=13$ participants $5\,\mathrm{s}$0, ages $5\,\mathrm{s}$1–$5\,\mathrm{s}$2, under two conditions: alone and with a companion. Each participant received a virtual budget of $5\,\mathrm{s}$330–$5\,\mathrm{s}$4120–$5\,\mathrm{s}$5+33.84\%$5\,\mathrm{s}$6+7.41\%$5\,\mathrm{s}$7+2.35\%$5\,\mathrm{s}$8+2.44\%$5\,\mathrm{s}$9+10.14\%$. No inferential statistics were reported. Qualitative positives included “Novel way to try fragrances,” “Customization was fun,” “Less crowded than physical stores,” and “Social advice helped me decide,” while concerns centered on scent precision, headset comfort, and controller navigation. The study concludes that social presence boosts immersion, imagination of smells, and purchase satisfaction, while also identifying calibration drift, long-term comfort, and ergonomics as unresolved constraints.

3. XR olfactory rendering, synchronization, and media orchestration

Several adjacent systems describe hardware and synchronization strategies that are directly compatible with an Aromaverse-type XR stack. In the Virtual Vauxhall Gardens case study, the University of Leicester Olfactory Display comprises a VR Host PC, an Olfactory Controller with an Arduino-class microcontroller and MIDI interface, and a wearable Olfactory Delivery Module with odor column, PTFE tubing, and a face mask with nostril inlets (Pearce et al., 19 May 2025). The transport layer uses MIDI over USB at $F = [T,M,B]^T,$0, pump flow is controlled over $F = [T,M,B]^T,$1 per head by PWM duty cycle, and up to 16 units can be daisy-chained on a single MIDI bus, supporting up to 128 discrete odor channels in principle. The internal controller is an event-driven finite-state machine mapping NoteNumber to odor channel and NoteVelocity to duty cycle $F = [T,M,B]^T,$2.

The concentration model is first-order:

$F = [T,M,B]^T,$3

with $F = [T,M,B]^T,$4 and $F = [T,M,B]^T,$5. After release, concentration decays as

$F = [T,M,B]^T,$6

Because measured command round-trip was $F = [T,M,B]^T,$7 while odor onset lag was approximately $F = [T,M,B]^T,$8, pre-triggering was used:

$F = [T,M,B]^T,$9

This work reported strong auditory presence and control, exceptionally high UEQ attractiveness, and an anticipation effect in which 60% of a sham group falsely reported smelling.

A second XR line of work examined scented virtual agents with a headset-mounted module clamped to a Meta Quest 2 and driven by an Arduino Nano plus HC-06 Bluetooth module (Han et al., 14 Sep 2025). The device used one ventilation fan and four scent fans holding cotton pads pre-soaked with woodsy, floral, or unpleasant odors. A Unity C# component opened a Bluetooth serial connection, sent “ON” when the agent stopped in front of the user, and sent “OFF” when the agent left; the script aligned a 7 s scent-on window to the agent’s text-bubble display, followed by 5 min of active ventilation between trials. Mixed-design ANOVA showed that unpleasant odors overpowered visual emotional expression and gender cues, while pleasant scents produced no reliable appeal boost over a no-scent baseline. Future engagement was predicted by affinity alone through

$T$0

A third line addressed semantic planning before physical delivery through a two-stage video-to-scent pipeline (Wang et al., 27 Jan 2026). Stage 1 used Google Gemini 3.Pro to extract a JSON visual timeline from video frames sampled at 1 fps; Stage 2 used GPT-5.2 to convert that timeline into structured olfactory events constrained by a fixed odor schema of about 20 classes, three intensity levels, and timing constraints. In Study 1, the system-generated plans outperformed over-inclusive and naive baselines under a Friedman test, with mean rank $T$1 and first-place rate $T$2. In Study 2, system plans exceeded the over-inclusive baseline on immersion, distraction reduction, coherence, and ease-to-imagine, with all Holm-corrected Wilcoxon $T$3. These results indicate that semantic planning can be intelligible and preferable even before hardware scent delivery is present.

Taken together, these works imply a layered Aromaverse architecture: low-latency actuation, pre-triggered synchronization to compensate for diffusion lag, explicit wash-in and wash-out management, and increasingly semantic scheduling of olfactive events.

4. Aromaverse as structured molecular odor space

In molecular machine learning, the Aromaverse is formally defined as the structured space of odor perceptions induced by small molecules, organized so that semantically or perceptually related odor descriptors occupy adjacent regions (Sajan et al., 11 Aug 2025). Let $T$4 be the set of odorant molecules and $T$5 the set of descriptors; the Aromaverse is the pair $T$6, where $T$7 is a hierarchical taxonomy over $T$8. The expert taxonomy was built from 14 curated human-odor datasets from Pyrfume, yielding 617 raw descriptors after merging identical spelling variants. These were organized into 557 source-based descriptors and 60 olfactory qualities, with 16 top-level scent families and 31 sub-classes. A parallel data-driven taxonomy used the 146 most frequent descriptors in a merged molecular dataset of approximately 6,700 molecules and clustered them with agglomerative hierarchical clustering using Euclidean distance, Ward linkage, and $T$9.

The predictive effect of organizing the descriptor space was quantified with logistic regression, random forest, and XGBoost. The reported XGBoost macro-averaged test metrics were:

Representation	AUROC	F1 score
Original, 146 labels	0.604	0.268
Expert taxonomy	0.684	0.496
Data-driven taxonomy	0.698	0.513

Randomized groupings preserving bin sizes achieved AUROC $M$0 and F1 $M$1, so both expert and data-driven taxonomies lay above the random baseline. Error analysis emphasized Burden-matrix eigenvalue features such as BCUTZ-1h and BCUTc-1l, and a pear-odorant case study showed that higher-level classes still captured dominant fruity and green facets for molecules such as Poirenate, Pearadise, Helvetolide, and Quince Ester.

A complementary study on first-principles odor prediction argued against a universal descriptor basis for smell (Zanineli et al., 10 Apr 2026). It constructed structural, vibrational, and electronic descriptors from DFT and RDKit, then trained a separate random forest classifier for each of 113 odor labels. Per-odor accuracy ranged from about 0.9 for “catty,” “almond,” and “camphoreous” to about 0.5 or below for “tobacco,” “mushroom,” and “metallic.” Feature importance analysis found odor-specific regimes: cinnamon was dominated by vibrational features, camphoreous by structural features such as rotational constants and ring counts, alliaceous by electronic charge descriptors, and musk by a balanced combination of vibrational, structural, and electronic descriptors. The paper concluded that no single descriptor class universally dominates odor prediction.

These results suggest that an Aromaverse in the molecular sense is not a flat descriptor inventory. It is a structured, heterogeneous odor space in which taxonomic grouping improves prediction, but the governing physicochemical determinants remain odor-specific rather than universal.

5. Generative and predictive computational frameworks

A second computational usage of Aromaverse is platform-oriented: a modular environment for generating molecules, filtering them, predicting odor-likeliness, assigning odor labels, and exposing interpretability artifacts (Sharma et al., 30 Jan 2025). One such blueprint combined graph autoencoders, variational graph autoencoders, adversarially regularized variants, diffusion models on graphs, and graph transformers. Generated molecular graphs were converted to RDKit molecules and passed through sanitization enforcing valence limits, ring-size constraints, and uniqueness via canonical SMILES. Fragrance-likeliness screening used a logistic-regression model with the explicit equation

$M$2

trained with SMOTE and 5-fold cross-validation. Reported performance was ROC AUC $M$3, recall $M$4, precision $M$5, accuracy $M$6, and F1 $M$7. A GNN head then predicted 138 odor probabilities, while SHAP identified $M$8 as the top global contributors.

A more tightly coupled generative framework combined a SMILES VAE with QSAR supervision to generate synthetically viable odorants (Pearce et al., 28 Dec 2025). The encoder used four 1D-CNN layers, the decoder used stacked GRUs, and a QSAR head predicted $M$9 from the latent variable $B$0. The VAE objective was

$B$1

with final settings $B$2, $B$3, batch size $B$4, and Adam with $B$5 and $B$6. A logistic regression QSAR model trained on 1,615 odorous and 309 non-odorous molecules supplied odor probabilities. On generation, rejection sampling yielded 100% valid molecular graphs and 94.82% uniqueness. Latent structuring was supported by a Fréchet ChemNet Distance of approximately 6.96 between generated molecules and known odorants, compared with approximately 21.6 for the ChemBL baseline. Bemis–Murcko analysis found 74.43% uncharted scaffold hops, and retrosynthesis via AiZynthFinder reported 100% solvable candidates with average 2.89 steps.

VIANA addressed a different but complementary target: perceived odor intensity (Queiroz et al., 1 Apr 2026). Its “tri-pillar” architecture fused a GCN-derived structural embedding, a PCA-compressed Principal Odor Map embedding retaining 95% of semantic variance, and a Hill-law dose-response head. The output was

$B$7

where $B$8 is the saturation asymptote, $B$9 the half-maximal concentration, and $[0,1]$0 the Hill coefficient. Using 209 molecules and about 20,900 total samples generated from synthetic dose-response curves, VIANA achieved peak test $[0,1]$1 and test MSE $[0,1]$2. The forward pass cost was reported as less than 10 ms on modern GPU or embedded-AI hardware, and the framework explicitly supported user-specific calibration by shifting $[0,1]$3 through reference odorants.

Across these frameworks, the computational Aromaverse is an overview of molecular generation, validity control, odor likelihood, multi-label semantics, interpretability, and dose-response modeling. A plausible implication is that the term marks an emerging interface between cheminformatics and digital olfaction rather than a single algorithmic lineage.

6. Astronomical and astrochemical Aromaverse

In astrochemistry, “Aromaverse” names the aromatic and cyclic molecular inventory of cold interstellar environments (McCarthy et al., 2021). The term refers specifically to the population of aromatic and other cyclic molecules found in cold, dense molecular clouds such as TMC-1, distinguished from highly unsaturated carbon chains by planar or nearly planar geometries with delocalized $[0,1]$4 electrons, asymmetric-top rotational spectra, and formation routes involving cyclization. Since 2018, five CN-functionalized rings have been identified in TMC-1 via Green Bank Telescope centimeter-wave observations combined with laboratory rest frequencies: benzonitrile, 1- and 2-cyanocyclopentadiene, and 1- and 2-cyanonaphthalene. Reported benzonitrile constants were $[0,1]$5, $[0,1]$6, $[0,1]$7 with $[0,1]$8 quadrupole coupling about $[0,1]$9, and its column density in TMC-1 was approximately $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$0. The review highlighted barrierless gas-phase routes such as

$$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$1

and

$$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$2

together with isotopic-labeling strategies for constraining cyclization mechanisms.

A galactic-scale usage appeared in a JWST/NIRCam WFSS survey explicitly described as the “so-called Aromaverse” of aromatic and aliphatic PAH bands at redshifts $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$3–$$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$4 (Lyu et al., 25 Feb 2025). The survey analyzed 200 galaxies with usable spectra, detected the $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$5 band at $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$6 in 88 out of 187 galaxies and the $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$7 band in 37 out of 159 galaxies, and found a strong correlation between $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$8 and $$F_{\text{blend}} = \alpha F_T + \beta F_M + \gamma F_B, \qquad \alpha+\beta+\gamma=1,\ \alpha,\beta,\gamma\ge 0.$$9 with Pearson $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$0 and Spearman $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$1. The mean $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$2 was $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$3, while $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$4 ranged from about 0.05 to 0.58 with median about 0.19, corresponding to aliphatic carbon fractions of about 0.8%–8.3% and median about 2.9%. The $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$5 ratio exhibited a drop by a factor of at least 10 below metallicities $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$6–$$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$7, whereas $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$8 showed no significant dependence on redshift, stellar mass, metallicity, or morphology but decreased with several SFR tracers.

A related JWST/MIRI study detected spatially varying $$I_{\text{chan}} = f(\text{raw\_level}), \qquad \text{raw\_level}\in\{0,\ldots,255\},$$9 PAH emission in the galaxy SPT 0418-47, less than 1.5 billion years after the Big Bang (Spilker et al., 2023). It measured a global $N=13$0, order-of-magnitude spatial variations in equivalent width, and a factor of about 5 variation in $N=13$1 across the source. The authors interpreted the resolved variations as evidence for geometric offsets between PAHs and large dust grains or strong local fluctuations in the UV radiation field.

This astronomical usage shifts Aromaverse away from human olfactory rendering and toward the chemistry, spectroscopy, and spatial distribution of aromatic carbonaceous matter. The shared terminology is therefore analogical rather than architectural: the “world” being mapped is a chemical and radiative environment rather than an XR scene or a learned odor manifold.