WorksComputer Science
Physics-Informed Graph Neural Network Surrogates for Turbulent Nanoparticle Dispersion in Dental Clinical Environments
Abstract
Dental aerosol procedures produce sub-50 micrometre nuclei that can remain airborne for long periods in enclosed clinics, creating pathways for airborne pathogen transmission. Reynolds-Averaged Navier-Stokes (RANS) simulations with Euler-Lagrange particle tracking capture this transport accurately but require very long run times per scenario, which precludes real-time clinical decision support in 3D. We present the Eulerian-Lagrangian Graph Interaction Network (ELGIN), a physics-informed graph surrogate that jointly predicts carrier-flow dynamics on the OpenFOAM polyhedral mesh and the per-parcel motion of the polydisperse spray cloud. ELGIN couples a multi-head Graph Transformer with Jacobi-preconditioned learnable pressure projection and a turbulence-closure head to a sigmoid-gated Lagrangian Interaction Network through differentiable inverse-distance mesh-parcel coupling, and advances parcels with a symplectic Stormer-Verlet integrator. A four-stage physics-informed curriculum stabilises 260-step autoregressive rollouts without gradient explosion. A parameter sweep with foam-extend 4.1 OpenFOAM reactingParcelFoam across clinically relevant ventilation rates and handpiece spray speeds provides CFD ground truth. This article reports a single-case demonstration in which both ELGIN and a Lagrangian-only baseline (M0) are trained and evaluated on Sweep_Case_03 of a twenty-case sweep; full 16/2/2 retraining is in progress and will replace all reported metrics. On this case, ELGIN tracks the foam-extend particle cloud much more closely than M0: mean parcel displacement error falls from 19.56% to 16.20% of room width and cloud radius-of-gyration error from 9.85% to 6.58%. A 26-second rollout completes in ~64 s on a 4 GB GPU, approximately 37x faster than the foam-extend reference pipeline, toward per-appointment infection-risk screening once the multi-case checkpoint is in place.
Introduction
Dental procedures such as high-speed drilling ( rpm), ultrasonic scaling, and air-polishing routinely generate polydisperse bioaerosols with droplet nuclei in the to range.[1, 2] These particles remain airborne for 30–90 minutes in poorly ventilated clinical rooms and carry bacteria (M. tuberculosis, oral streptococci), viruses (SARS-CoV-2, hepatitis B), and fungal spores that infect patients, dental workers, and bystanders.[4] The COVID-19 pandemic dramatically reinvigorated concern about airborne pathogen transmission in enclosed clinical spaces,[5, 6] motivating quantitative tools that can predict where aerosols migrate under given ventilation conditions and identify infection hot-zones in real time.
Computational Fluid Dynamics (CFD) has been the primary quantitative tool for complex flow phenomena for over six decades.[7] For turbulent aerosol transport the Euler–Lagrange framework is widely used: the carrier-phase turbulent flow is resolved by the Reynolds-Averaged Navier–Stokes (RANS) equations with a – SST (Shear Stress Transport) closure,[8] while individual particle trajectories are obtained by integrating Newton’s second law under aerodynamic drag (with Cunningham-corrected Stokes relaxation time),[9] gravity, turbulent dispersion via a Discrete Random Walk (DRW) model,[10] and Brownian diffusion[11] for sub-micrometre nuclei. Despite this physical fidelity, a single RANS + Lagrangian-parcel simulation of a dental treatment room requires roughly 40 min of single-core wall time on a modern workstation in the present foam-extend 4.1 reference pipeline (SI Sec. S7), rendering Monte Carlo sweeps over ventilation rates, room layouts, and spray angles intractable within clinical timescales. Published CFD studies of dental and indoor bioaerosol transport[3, 4, 12] have characterised the risk qualitatively but none provides a tool capable of real-time personalised guidance during a dental appointment.
The past decade has witnessed explosive growth in machine-learning surrogates for CFD. Convolutional Neural Networks (CNNs) on Cartesian grids,[13] embedded learned closures,[14] Fourier Neural Operators,[15] DeepONet,[16] and Physics-Informed Neural Networks (PINNs)[17] all achieve – speed-ups but share a fundamental limitation: they operate on fixed-resolution structured grids or hand-crafted collocation points and cannot generalise to the unstructured topology-varying meshes characteristic of complex clinical geometries. Graph Neural Networks (GNNs) overcome this limitation by representing the simulation domain as a graph in which mesh cells or particles are nodes with feature vectors , edges carry edge feature vectors , and information propagates through rounds of message passing:
where is the message passed from node to node at round , is the dimension of , are learnable multi-layer perceptrons, is the neighbourhood of node , and is a permutation-invariant aggregation operator.[18] This Message Passing Neural Network formalism is the formal generalisation of finite-volume stencil updates to arbitrary graph topologies, and the relational inductive bias[19] aligns naturally with the locality of physical interaction laws.
The Encode-Process-Decode (EPD) paradigm of Sanchez-González et al.20 (the Graph Network Simulator, GNS) and the mesh-based MeshGraphNets of Pfaff et al.[21] provide the architectural foundation for most GNN-CFD surrogates. For Lagrangian particle simulation the GNS achieves rollout errors below 5% at the speed of Smoothed Particle Hydrodynamics (SPH) and Material Point Method (MPM) solvers on granular benchmarks. Recent extensions address ocean wave dynamics,[22] dense granular suspensions,[23] plasma particle-in-cell,[24] SE(3)-equivariant momentum conservation[25] (Special Euclidean group in three dimensions), and probabilistic uncertainty quantification through diffusion graph networks. Physics-informed loss terms,[17] attention aggregation,27,28 and Hamiltonian or Lagrangian decoders29–31 have all been proposed to mitigate the persistent weakness of vanilla GNS: purely data-driven training does not guarantee energy or momentum conservation, and rollouts drift over long horizons.[25]
Two concurrent works independently highlight the value of Eulerian–Lagrangian hybridisation for GNN surrogates. COGRI32 augments a GNS backbone with a lightweight convolutional Eulerian grid for global context aggregation, projecting particle features to a grid, applying CNN updates, and mapping the result back to the particle domain, achieving 57% better rollout accuracy over plain GNS with only 13% inference overhead. DeepLag33 integrates Lagrangian particle tracking into Eulerian field prediction via cross-attention EULAG blocks, using tracked particles to guide the evolution of the Eulerian velocity field. Both methods use learned Eulerian representations; neither conditions on a pre-computed physical flow field, nor addresses multi-physics particle dynamics. For RANS-conditioned GNNs on Eulerian fields, PIGNN-CFD34 trains a physics-informed GNN on OpenFOAM RANS outputs to predict urban wind fields on unstructured meshes at CFD speed, demonstrating that RANS-mesh graphs are viable training inputs, but does not track Lagrangian particles or handle particle physics. In the aerosol domain, GLAD35 applies the GNS framework to atmospheric aerosol microphysics (sulfuric acid condensation and coagulation in the PartMC-MOSAIC model), demonstrating GNS generalisation to particle chemical dynamics. GLAD operates without a background flow field and targets composition changes rather than spatial transport, leaving the problem of turbulent indoor bioaerosol dispersion unaddressed. A comprehensive comparison of GNN-CFD methods, methodologies, advantages, and limitations is provided in the Supporting Information (§ 21, Table S5).
| Method | Methodology | Key finding | Advantage (+) Limitation (−) |
| CNN surrogate[15] | U-Net on Cartesian grid; predicts steady , | FV; MAE < 2% on bluff bodies | + Fast; simple. − Structured grid only; no unstructured mesh |
| ML-inside-FV[16] | CNN flux corrections inside classical FV solver; JAX | 40–80×; 8–10× coarser grid at DNS accuracy | + Retains solver stability. − Structured grid; not portable to new geometries |
| PINNs[17] | MLP minimising NS residuals at collocation points | Accurate for laminar inverse problems from sparse data | + Physics-consistent. − Spectral bias; expensive; intractable for 3-D turbulence |
| MeshGraphNets[18] | EPD GNN on native FV/FE mesh; multi-scale hierarchy | 400×; 0.3% error on cylinder wake | + Unstructured mesh; variable topology. − Long-rollout drift; |
| GNN wall model[19] | GNN: near-wall profiles → ; DNS-trained LES | Beats Smagorinsky for –5200 | + Generalises across . − High-fidelity training data; wall layer only |
| DA-GNN20 | GNN + FreeFEM++ adjoint; RANS residual constraint | RANS field from 5 sensors; < 5% error on bluff bodies | + Physics-consistent; sparse data. − Requires adjoint solver; steady flows only |
| GNS (TF)[21] | Radius graph; EPD + noise augmentation; Euler rollout | SPH/MPM; < 5% RMSE (sand, water, goop) | + General; variable particle count. − No carrier phase; energy non-conservation |
| GNS (PyTorch)[22] | PyTorch/PyG; multi-GPU DDP; inertial-frame bias | 5000× MPM; < 5% error; 8-h training on A100 | + Open source; HPC scalable. − Same physical limits; no aerosol physics |
| Dynami-CAL[23] | SE(3)-equivariant edge frames; momentum conserved by construction | >1000 stable steps vs. ∼200 for GNS; 3-D granular | + Exact momentum; rotation-invariant. − Higher cost; dry granular only |
| Diffusion GNN[24] | Score-based diffusion in graph latent space | Full UQ on 3-D wing pressure; calibrated intervals | + Probabilistic; calibrated UQ. − High training cost; slow inference |
| PIC-GNN[25] | Hybrid graph: particle nodes + Eulerian field nodes | 100× larger ; two-stream plasma instability | + Coupled particle-field. − 1-D plasma only; no turbulent aerosol physics |
| Ocean GNS26 | Variable- GNS; wall nodes as fixed particles | > 1000× SPH; run-up < 3% error | + Flexible time step; harbour geometry. − Free-surface only; no dispersed particles |
| Baseline GNS (present, M0) | Radius graph (); EPD; isotropic Gaussian noise augmentation; Euler rollout; analytic Cunningham drag node feature | Lagrangian-only baseline; metrics reported in Table III of the main text | + Simple; fast; competitive at 300 ep. − No carrier-phase signal; slight KE over-estimation |
| ELGIN (present) | Hybrid Eulerian–Lagrangian: Graph Transformer + Jacobi-PCG pressure projection on the polyMesh; LSTM history; rotation-invariant local-frame geometric edges (Cartesian drag channel); Störmer–Verlet integrator; per-case airInlet conditioning; polyMesh-derived and ; IDW-mapped to every parcel | Joint Eulerian + Lagrangian prediction on coupled mesh and particle graphs (IDW exchange); metrics reported in Table III of the main text | + Geometry- and inlet-conditioned hybrid GNN for dental-room aerosol; near-wall sticking flag from wall-distance features. − Requires polyMesh extraction; quasi-2-D extruded geometry |
| Method | Methodology | Key finding | Advantage (+) Limitation (−) |
Table S5. Literature summary of GNN and ML methods for CFD.
A systematic examination of the GNN-CFD literature reveals that existing Lagrangian GNN surrogates address dry granular flows,20,36 dense suspensions,[23] free-surface waves,[22] plasma particle-in-cell,[24] or atmospheric aerosol microphysics without spatial transport.35 RANS-conditioned GNNs have been demonstrated for Eulerian wind-field prediction on urban OpenFOAM meshes,34 and hybrid Eulerian–Lagrangian architectures have been explored with learned (rather than physically pre-computed) Eulerian representations.32,33 No previous work reports a GNN surrogate for turbulent bioaerosol spatial transport in enclosed clinical spaces that is conditioned on a pre-computed physical carrier-phase flow field. This gap is consequential, because dental aerosol physics in this regime involves several effects that prior GNN benchmarks do not jointly cover.
Spray droplets are polydisperse (1 µm to 50 µm), so aerodynamic relaxation times span orders of magnitude with . The ventilation Stokes number (ratio of the Stokes relaxation time at mean diameter to , with matching the CFD domain height; full definition in SI Sec. S5) remains across the sweep, i.e. tracer-like on the mean-flow time scale. Sub-micrometre nuclei exhibit Cunningham slip-correction factors that strongly modify drag. Saffman shear-lift is significant for and larger; Brownian diffusion dominates below and requires a stochastic model that deterministic GNS formulations omit. The carrier flow is an Eulerian field tightly coupled to obstacle-laden geometry (dentist, patient, walls, ceiling supply, lateral pressure outlet), so dispersion depends on wall-normal proximity, swirl, and recirculation. Clinical deployment further requires calibrated risk intervals on infection-relevant metrics.
This paper addresses these gaps through a hybrid Eulerian–Lagrangian GNN surrogate framework, demonstrated here on the representative case Sweep_Case_03 of a twenty-case foam-extend 4.1 reactingParcelFoam parameter sweep of a two-dimensional dental treatment room.
The present results are a single-case demonstration that illustrates the architecture, training curriculum and evaluation pipeline on one realisation of the parameter grid; the production checkpoint analysed below is trained and rolled out on the same case, so the numbers should be read as proof-of-concept fidelity rather than as a generalisation study. Full 16/2/2 (train/validation/test) retraining on the complete twenty-case dataset is in progress and results will be updated once the multi-case rollouts are available.
Two model variants are developed and compared: (i) a GNS baseline (M0) following the Encode–Process–Decode paradigm of Sanchez-González et al.,20 with uniform Lagrangian message aggregation and no Eulerian carrier-field conditioning; and (ii) ELGIN, the principal contribution of this work, which solves the carrier flow on the OpenFOAM polyMesh with a multi-head Graph Transformer processor and a Jacobi-preconditioned learnable pressure projection, replaces uniform aggregation with sigmoid-gated attention on the particle graph,27 exposes every parcel node to the projected RANS velocity, the local turbulent kinetic energy, the obstacle-aware distance-to-wall, and the wall-normal vector interpolated from the mesh, broadcasts a per-case airInlet velocity vector as a global conditioning feature, and integrates particle positions with a symplectic Störmer–Verlet step.
The four primary contributions of this work are: (i) a hybrid Eulerian–Lagrangian GNN surrogate for polydisperse dental bioaerosol that jointly predicts the divergence-corrected RANS velocity field and the particle trajectories on the same OpenFOAM polyMesh; (ii) a full-timeline data-extraction protocol with persistent origID tracking and an alive-mask training objective that correctly handles variable parcel counts arising from injection and deposition; (iii) explicit, name-aware boundary-condition encoding (nine semantic classes including dentist and patient obstacles), polyMesh-derived distance-to-wall and wall-normal features, and a per-case airInlet velocity conditioning vector designed to let a single trained model generalise across ventilation rates and handpiece spray speeds (the present single-case checkpoint exercises the conditioning mechanism; demonstration of cross-case generalisation is deferred to the planned 16/2/2 retraining); and (iv) a four-stage training curriculum (Eulerian pre-training, particle supervised, PDE-informed joint, backpropagation-through-time rollout fine-tuning) that stabilises long-horizon GNN rollouts for the dissipative multi-physics dental problem.
The remainder of the paper is organised as follows. Section II summarises the CFD data-generation methodology (detailed equations and solver settings are in SI Sec. S1). Section III presents the two GNN model variants. Section IV outlines the four-stage training curriculum (detailed loss formulation in SI Sec. S3). Section V reports headline rollout metrics, snapshot comparisons, and the clinical Breathing Zone Exposure (BZE) outcome, with extended results, non-dimensional analyses, dispersion statistics, and computational benchmarks in SI Secs. S4–S7. Section VI concludes.
CFD DATA GENERATION
Geometry, governing physics, and parameter sweep
The computational domain represents a simplified two-dimensional cross-section of a dental treatment room (4.0 m × 3.0 m, Figure 1a), using Cartesian coordinates with vertical coordinate . The two-dimensional cross-section is adopted to balance physical representativeness with the computational cost of generating a twenty-case sweep for GNN training. The domain contains a 0.2 m-wide ceiling supply inlet (downward velocity ); a lateral pressure outlet at ; rectangular obstacles representing the dentist and patient; and a 3 mm dental-handpiece nozzle (, in the typical range for high-speed dental sprays) at the patient’s oral cavity, aimed horizontally toward the dentist with injection speed and cone half-angle . A breathing zone (BZ) is defined at the dentist’s head height for infection-risk metric evaluation; the Breathing Zone Exposure (BZE) computed from this volume is specified in Sec. V D. The full set of boundary conditions, mesh-independence study, and inlet turbulence formulae are tabulated in SI Sec. S1.
Figure 1. (a) Dental treatment room: 4 m × 3 m 2-D cross-section with every OpenFOAM boundary patch labelled (ceilingInlet, outlet, floor, ceiling, leftWall, rightWall, dentist, patient, and symmetric frontAndBack empty extrusion), the dental-handpiece nozzle, and the clinician breathing zone (BZ). (b) Wall-clock time for one autoregressive surrogate rollout spanning the stored 26 s trajectory segment versus the foam-extend reactingParcelFoam reference pipeline for the same segment.
The carrier-phase airflow is modelled by the incompressible RANS equations with the Menter – SST closure.[8]
where is the Reynolds-averaged mean velocity, and are the kinematic and turbulent eddy viscosities, is the modified kinematic pressure, and is the particle back-reaction. The eddy viscosity is closed by the SST formula and reported in full in SI (29).
Each parcel of instantaneous diameter and mass obeys the simplified Maxey–Riley equation:37
where is the parcel velocity, the Cunningham-corrected Stokes drag, the McLaughlin–Saffman shear-lift, the gravitational force with gravitational acceleration , the Brownian stochastic (thermal) force, and the turbulent dispersion force from the DRW model. The full closed forms of each term, the Cunningham slip-correction factor, and the McLaughlin-corrected Saffman expression are given in SI Sec. S1.3, and a force-magnitude regime map across is presented in SI Sec. S4.
Parcels are sampled from a Rosin–Rammler distribution (, ) and the foam-extend solver evaporates each parcel quasi-steadily according to Wells’ law,38
where is the initial droplet diameter, and are the air and droplet densities, respectively, is the vapour mass diffusivity, and is the Spalding mass-transfer number at 50% relative humidity. The full Wells closure and its implementation in reactingParcelFoam are documented in SI Sec. S1.3.5. The training data are extracted at without further evaporation post-processing, and the GNN is configured to treat the parcel diameter as a static per-parcel input feature inherited from the initial Rosin–Rammler sampling (diameter dynamics are not predicted by the surrogate). This is consistent with the relatively small fractional change in over the typical advection time-scale resolved by the snapshots, and avoids polluting the learned dynamics with an additional time-scale that is not needed for displacement and dispersion prediction; the Wells reduction in inertia is implicitly absorbed into the trajectory targets generated by the foam-extend solver.
Each case is integrated as a two-stage foam-extend 4.1 pipeline: steady-state RANS via simpleFoam, followed by transient compressible reacting parcel transport via reactingParcelFoam for 30 s of physical time using PIMPLE (merged PISO–SIMPLE) pressure–velocity coupling. Twenty cases span a 4 × 5 factorial grid over and at fixed , providing 261 quasi-steady-state frames per case (extracted on at ) split 16/2/2 (train/validation/test) by random shuffle with a fixed seed. The OpenFOAM polyMesh contains owner cells with nine named boundary patches (ceilingInlet, outlet, floor, ceiling, leftWall, rightWall, dentist, patient, and the frontAndBack empty patch enforced by the 2-D extrusion). Figure 1(a) sketches the associated patch layout, nozzle, and breathing zone. Detailed solver settings, the full case list, and dataset statistics are provided in SI Table S2–S4. For each timestep, representative parcel trajectories are tracked with persistent origId identifiers so that every node in the Lagrangian graph corresponds to the same physical parcel across all frames (deposited or out-of-domain parcels are masked out by an alive-mask flag rather than re-indexed).
| Parameter | Value |
| Stage 1 fluid solver | simpleFoam (steady, incompressible) |
| Stage 2 fluid+parcel | reactingParcelFoam (transient, compressible) |
| Mesh (raw) | Structured quad., cells (25 blocks) |
| Mesh (active fluid) | 7 704 cells after dentist/patient obstacle removal |
| Cell size | |
| Turbulence model | – SST3 |
| Thermophysics | hsPsiMixtureThermo<reactingMixture<gasThermoPhysics> |
| Species | , (NASA-7 polynomials, no reactions) |
| Time stepping | Adaptive PIMPLE, , |
| Typical | |
| Physical time | |
| Output interval | (adjustableRunTime) |
| Raw snapshots / case | over the full |
| Snapshots used | 261 in (quasi-steady window) |
| Parcel injection rate | 5000 parcels/s (continuous, ) |
| Volumetric flow rate | |
| Total injected mass | per case |
| Wall interaction | escape (deposition: parcel removed) |
| Outlet treatment | Pressure outlet, escape |
| Evaporation | Wells’ -law (LiquidEvaporation) |
| Stage 1 convergence | Residuals |
| Wall-clock per case | (single core, foam-extend 4.1; 38–42 min logged for Sweep_Case_01/17) |
Table S2. Numerical solver settings for the twenty-case foam-extend 4.1 pipeline used to generate the GNN training data.
GNN Architectures
The ELGIN surrogate operates on a dual-graph representation: a static Eulerian mesh graph derived from the foam-extend polyMesh, and a time-varying Lagrangian particle graph rebuilt at every rollout step. The two graphs are coupled through a sparse cross-graph inverse-distance-weighted (IDW) interpolation operator (Sec. III A). The two model variants compared in this paper – GNS (M0) and ELGIN – differ by the carrier-phase information they consume, by Lagrangian message aggregation (uniform sum vs. sigmoid-gated attention), and by the position integrator, but share the same Encode–Process–Decode (EPD) backbone 39. The Lagrangian side employs a long short-term memory (LSTM)39 velocity-history encoder and an optional variational autoencoder (VAE)-style acceleration decoder (Sec. III C). The full ELGIN is described first (Secs. III A–III C); M0 is obtained by switching off specific feature flags (Sec. III D).
Figure 2 gives an end-to-end schematic of the ELGIN pipeline. (Panel 1) depicts the OpenFOAM ground truth: RANS fields , , , and on the polyMesh together with the sprayed Lagrangian parcels. The dual-graph representation extracted from this data couples a static Eulerian mesh graph on cells, equipped with boundary-condition embeddings, distance-to-wall , wall-normal , and the inlet-velocity vector , to a time-varying Lagrangian radius graph over persistent-origId parcels with connectivity radius m. Panel 2 shows the dual sub-network architecture. The Eulerian branch is an Encode–Process–Decode stack of Graph Transformer blocks based on the masked-attention dot-product formulation of Shi et al. 40, followed by a turbulence-closure head and a Jacobi-preconditioned learnable pressure projection (Eq. (16)). The Lagrangian branch combines an LSTM velocity-history encoder, an Interaction-Network processor with sigmoid-gated attention (Eq. (21)), an optional VAE-style acceleration decoder (§ 3 C), and a symplectic Störmer–Verlet integrator (Eq. (20)). Information is exchanged through a differentiable inverse-distance-weighted operator with nearest cells (Eq. (12)). Panel 3 outlines the four-stage curriculum used for both M0 and ELGIN. Stage 1 is Eulerian fluid pre-training. Stage 2 is one-step particle supervision with input noise. Stage 3 is PDE-informed joint training with continuity, momentum, turbulence, angular, and Kullback–Leibler (KL) terms. Stage 4 is back-propagation through time (BPTT) rollout fine-tuning. The architectural and training hyperparameter (, , , batch size, per-stage epoch counts, learning rates, BPTT unroll length, noise scales) are provided in Table I.

Figure 2. Overview of the ELGIN workflow: dual-graph extraction from OpenFOAM RANS and parcel data, coupled Eulerian and Lagrangian sub-networks with differentiable inverse-distance exchange, and the four-stage training curriculum (the same backbone and schedule apply to the M0 baseline). Full notation, loss terms, and equation references are given in § 3.
| Parameter | Value |
| Optimiser | AdamW,41 weight decay |
| LR schedule | Cosine annealing () |
| Total epochs | 300 (both variants) |
| Stage allocation | ELGIN: 60/120/60/120/60; M0: 75/150/75 (Stage 1 skipped) |
| Batch size | 8 |
| Hidden dimension | 64 |
| Message-passing steps | (Lagrangian); (Eulerian) |
| Lagrangian connectivity | 0.10 m (ELGIN); 0.30 m (M0 baseline) |
| Cross-graph IDW neighbours | 4 |
| Pressure-projection PCG iters | 20 (Stages 1–2); 50 (Stages 3–4); Jacobi-precond. |
| History length | 5 positions (= 4 velocity diffs); LSTM-encoded in ELGIN |
| Model time step | 0.1 s (= CFD snapshot interval ) |
| Lagrangian integrator | Störmer–Verlet (ELGIN); Euler (M0) |
| Position-noise scale | (isotropic Gaussian) |
| BPTT unroll steps | 5 (Stage 4) |
| BPTT loss weight | 0.7 |
| BPTT rollout noise | m |
| Stochastic decoder KL weight | (decoder off in production checkpoint) |
| GPU | NVIDIA Quadro P1000 (4 GB VRAM) |
Table 1. Training hyperparameters common to both model variants (20-case run).
Dual-graph construction
The polyMesh of owner cells defines nodes (one per cell centroid) and undirected edges through the internal face connectivity. At every internal face, the owner–neighbour pair is added in both directions, giving a fully antisymmetric processor. The polyMesh boundary patches are parsed by a name-aware classifier into nine semantic classes: interior, inlet (ceilingInlet), outlet, wall (leftWall, rightWall), floor, ceiling, dentist, patient, and the symmetryPlane/empty 2-D extrusion class. Each cell adjacent to a non-interior, non-empty boundary face inherits the corresponding boundary-condition (BC) identifier (first-wins so that the front/back empty patch never overwrites a physical wall, dentist, patient, inlet or outlet). For each cell we compute the true geometric distance to the nearest wall-type face (walls + dentist + patient) directly from the polyMesh,
where is the set of obstacle and wall boundary faces (walls, floor, ceiling, dentist, patient), is the centroid of cell , and is the midpoint of face , together with the unit wall-normal vector pointing from the cell centroid towards the closest wall-face midpoint. The polyMesh axis-aligned bounding box defines the domain bounds that constrain the rollout (§ 3 C).
Each Eulerian node carries the feature vector
where is the per-channel -score of the five RANS state variables; is the two-sided signed distance to the four bounding-box walls; is a learned look-up table over BC identifiers ; and is the per-case airInlet velocity vector parsed directly from the OpenFOAM 0/U dictionary, broadcast to every node so that a single trained model generalises across the ventilation grid. Eulerian edges concatenate a six-dimensional geometric descriptor (the outward face unit normal , the face area scaled by , the Euclidean distance between owner and neighbour cell centroids scaled by , and the unit separation vector in two components) with a four-dimensional learned face-type embedding, giving . The length is the nominal room width and matches the reference scale used for trajectory-error metrics in § 6 A.
At each timestep , the tracked parcels (uniquely identified by their persistent foam-extend origId) form a directed radius graph,
where is the position of parcel at time . Both orientations and are retained to preserve antisymmetry under relative velocity. The connectivity radius is set by the median nearest-neighbour spacing in the spray core to keep the mean node degree below 20 while still resolving aerodynamic wake interactions, and avoids the over-squashing pathology of dense graphs.[19]
Each Lagrangian node concatenates an LSTM-encoded velocity history, geometric box-SDF features, a learned particle-type embedding, the analytic Cunningham-corrected Stokes-drag acceleration, the log diameter, the local turbulent kinetic energy interpolated from , and – importantly – the cell-interpolated distance-to-wall and unit wall-normal pulled from Eq. (7). Concretely,
where the LSTM history encoder embeds the most recent finite-difference velocities into a -dimensional state, is a discrete particle-type label feeding , and tildes mark training-set-normalised channels. Edges in use a local-frame geometric representation whose first three components are invariant under planar rotations and translations of the spray cone, while the fourth (the analytic drag acceleration) is retained in the global Cartesian frame:
with and the relative bearing in the source-particle local frame. The composition is thus rotation-invariant in the geometric descriptor and rotation-covariant in the drag channel; full SE(2)-equivariance would require expressing in the same local frame, which is a straightforward extension reserved for the planned full twenty-case retraining.
Eulerian-to-Lagrangian feature transfer at any continuous parcel position uses an inverse-distance-weighted interpolation over the nearest cell centroids,
where denotes any scalar field component carried on the mesh, is its value at cell , is the centroid of cell , is the set of the nearest cell centroids to , and is a small positive constant for numerical stability, which is differentiable and respects the polyMesh’s unstructured topology (no rectangular bilinear assumption). The same operator transports and from cells to parcels for the Lagrangian wall feature in Eq. (10).
Eulerian sub-network and pressure projection
The Eulerian sub-network advances the carrier flow on as
where stacks the five RANS fields on each mesh cell at training step and is the predicted residual increment from the network with parameters . The increment is produced by an EPD network whose processor is a stack of multi-head graph-transformer blocks with heads each; node updates use small multi-layer perceptron (MLP) maps denoted below.
with hidden dimension , residual connections and layer normalisation. Here is the Eulerian node embedding, the edge feature on the owner–neighbour link in , , , , and are learnable attention projections for head , are attention weights, the concatenation stacks head outputs, and is the set of graph neighbours of node on . This is the canonical softmax dot-product attention of graph-transformer architectures40,42 adapted to the message-passing setting of Battaglia et al.[19], and is consequently distinct from the sigmoid-gated graph attention network (GAT) used on the Lagrangian side (§ 3 C).
A learnable turbulence closure head provides an explicit eddy-viscosity prediction at every cell, regularised towards the algebraic SST formula (Eq. (29)) so that the network does not have to re-discover the closure from scratch. A learnable pressure-projection module reduces the discrete divergence of the predicted velocity update by solving a graph-Poisson problem
where is a scalar pressure-correction potential on the mesh graph and is a graph-edge approximation of the finite-volume gradient operator (it uses the same edge set as the message-passing processor; it is not the OpenFOAM div(phi) face-flux operator with non-orthogonal corrections). The projection therefore enforces zero net signed-edge flux on the graph stencil and is consistent with, but not identical to, an exact finite-volume divergence-free constraint. The discrete graph-Laplacian is inverted by a fixed number of Jacobi-preconditioned conjugate-gradient (PCG) iterations on ; the inner solve is fully differentiable and adds no mesh-structure assumption beyond what the graph already encodes. The combined Eulerian sub-network (encoder, graph-transformer blocks, decoder, turbulence head, pressure projection) is pre-trained in Stage 1 of the curriculum (Sec. IV) and held nearly frozen thereafter.
Lagrangian sub-network
The Lagrangian processor is a stack of Interaction Network blocks with the same encode–process–decode skeleton as the Eulerian side but operating on :
where is the Lagrangian message from parcel to parcel , the aggregated message, the parcel node embedding, is the set of Lagrangian neighbours of parcel under Eq. (9), selects between uniform sum aggregation and sigmoid-gated attention, controlled by the model variant (§ 3 D); is the scalar gate logit produced from the Lagrangian edge encoding . An optional probabilistic (VAE-style) decoder43 maps the final node latent state to a normalised acceleration via a head pair and the reparametrisation trick, with the associated KL divergence added to Eq. (22) with weight .
Particle positions are advanced using a symplectic Störmer–Verlet kick–drift–kick scheme,31
Here scales decoded accelerations into physical units consistent with the ground-truth training-set statistics; and are parcel position and velocity. The integration preserves a discrete symplectic two-form31 and prevents the energy drift typical of forward-Euler GNS rollouts on conservative flows.20 Predicted positions are clipped to the polyMesh bounds at every step, and a geometry-aware deposition flag is raised whenever falls below an adhesive threshold, where is the shortest distance from parcel to the rectangular domain bbox; once raised, the parcel’s velocity is set to zero for the remainder of the rollout, mimicking the foam-extend stick wall interaction.
The Lagrangian network is designed so that closed-form particle physics need not be learned from scratch: at every step the analytic Cunningham-corrected Stokes drag and the SST turbulent-kinetic-energy field enter the node and edge features (Eq. (10), (11)); a discrete-random-walk (DRW) eddy-interaction model adds an analytic turbulent kick scaled by .
Two model variants
The two GNN surrogates compared in this work are configurations of the same hybrid backbone, summarised in Table II:
| Model | Lagrangian aggregation | Carrier-phase conditioning |
| GNS (M0) | Uniform sum | none |
| ELGIN | Sigmoid attention | full Eulerian + pressure proj. |
Table II. Architecture comparison for the two GNN surrogates. is the number of Lagrangian message-passing blocks; the Eulerian sub-network and pressure projection are only active in ELGIN.
GNS (M0): Lagrangian-only baseline with sum aggregation, a flat velocity-history input (no LSTM), simple relative-position edges, forward-Euler integration, and no carrier-flow conditioning. This is a faithful reproduction of the canonical GNS of Sanchez-González et al.20``. The M0 connectivity radius is set to , larger than the ELGIN value below, so that the purely Lagrangian neighbourhood reaches across the spray cone and partially compensates for the absence of an Eulerian context; reducing to the ELGIN value gave noticeably worse M0 validation MDE in preliminary tests because parcels lose neighbours as the cloud disperses. This choice is therefore conservative in favour of M0.
ELGIN: the full hybrid model. Lagrangian message aggregation uses sigmoid-gated graph attention (Eq. (21)) rather than uniform sums, consistent with Veličković et al.27``; the Eulerian sub-network of § 3 B runs at every step, the Lagrangian processor consumes the projected RANS velocity through the analytic drag/Saffman/turbulent kinetic energy (TKE) features, the LSTM history encoder (Eq. (10)) and rotation-invariant local-frame edges (Eq. (11)) replace the minimal M0 inputs, the distance-to-wall and wall-normal interpolated from the polyMesh enter both the Eulerian and the Lagrangian features, the per-case airInlet vector is broadcast as global conditioning, and the symplectic Störmer–Verlet integrator of Eq. (20) replaces forward Euler. A tighter Lagrangian radius suffices here because the cell-interpolated RANS velocity already supplies the long-range carrier information, so the network does not need a wider parcel neighbourhood to recover it.
The attention gate in ELGIN takes the form
with learned row vector and weight matrices , and denoting feature concatenation. The sigmoid (rather than softmax) yields an absolute relevance score in and lets the model down-weight distant neighbours; in the spray core the weights concentrate on the nearest fluid-coupled parcels, recovering an implicit Oseen-type distance decay.44 Crucially, neither model predicts a per-step diameter change: enters Eq. (10) as a static per-parcel input feature inherited from the Rosin–Rammler initial distribution, and the surrogate is a position–velocity predictor only. This is a deliberate design choice motivated by two observations. First, sub-10 droplets dominate clinical aerosol exposure risk, and for that fraction the per-step fractional change in is below 1% (SI Sec. S1.3.5), so the static- approximation is essentially exact at the snapshot grid. Second, the foam-extend trajectory targets already absorb the inertia reduction caused by Wells’ law (with full 50 droplets evaporating to their nuclei in , SI eq:S20), so the GNN learns a position–velocity mapping that implicitly inherits the diameter drift from the target data without having to track it explicitly. Diameter dynamics, if required by a downstream infection-risk calculation, can be applied as an analytic post-processing step on the predicted trajectories.
Training Protocol
Both model variants share the same four-stage curriculum (panel 3 of Figure 2). The AdamW optimiser41 with weight decay and cosine-annealing learning-rate schedule is used throughout, on an NVIDIA Quadro P1000 GPU (4 GB VRAM) using PyTorch 2.x45 and PyTorch Geometric.46 The full epoch budget of 300 epochs is distributed across the four stages as 60 : 60 : 120 : 60, allocating the bulk of the budget to the PDE-informed joint Stage 3 and the BPTT rollout fine-tuning Stage 4: The production checkpoint analysed in Sec. V is trained on the single representative case Sweep_Case_03 of the twenty-case CFD design space.
Stage 1 (Eulerian fluid pre-training, 60 epochs, ): The Eulerian sub-network of the ELGIN is trained on the static RANS snapshots to predict the residual increment of on the mesh graph, with the SST regulariser on the turbulence-closure head and the discrete continuity residual after pressure projection (Eq. (16)). The Lagrangian sub-network is frozen during this stage, and the stage is skipped entirely for the M0 variant.
Stage 2 (one-step particle supervised, 60 epochs, , Eulerian frozen): The Eulerian stack stays frozen while the Lagrangian network learns one-step acceleration from teacher forcing.20 Training examples are trajectory segments of length that respect persistent origId pairing and the parcel alive mask. Isotropic Gaussian noise ( in normalised units) is added to velocities as augmentation. The Stage 2 objective is the alive-masked mean squared error on acceleration. Optionally, samples are weighted with a multi-Gaussian kernel-density-estimation reweighting so that spray-cone, near-wall, and recirculation parcels contribute more than the bulk of the diffuse cloud.
Stage 3 (PDE-informed joint, 120 epochs, ): Both fluid and particle sub-networks are jointly trained with the total loss
The nonnegative scalars , , , , , and are loss weights. is the mean-squared mismatch between predicted and ground-truth parcel accelerations (alive-masked, variance-normalised). is the mean squared discrete divergence on the Eulerian mesh graph . is the scale-normalised finite-volume residual of the RANS momentum equation on . penalises departure from the algebraic SST relation linking predicted , , and eddy viscosity . is a weak penalty on spurious step-to-step changes in the cloud’s depth-averaged angular momentum about its centroid. is the Kullback–Leibler divergence from the optional VAE acceleration decoder (§ 3 C); it is identically zero for the deterministic production checkpoint analysed in Sec. V, where the VAE head is disabled. The default weights are (with active only when the VAE decoder is enabled). The momentum residual is normalised by , with reference speed as in SI Sec. S3, to make the loss scale-independent across mesh refinements; the finite-difference time derivative uses consistently with the CFD snapshot interval. No evaporation loss is included, since the surrogate does not predict diameter changes (Sec. III D). Closed-form expressions for , , , , , and are given in SI Sec. S3.
Stage 4 (BPTT rollout fine-tuning, 60 epochs, ): Autoregressive rollout over steps with back-propagation through time (BPTT)20,47 and gradient checkpointing every five steps. The Stage-4 loss is a 70/30 weighted average of the multi-step rollout and one-step losses (), and the position-noise scale is doubled relative to Stage 2 (to ) to expose the network to rollout-distribution inputs. In addition, a Gaussian noise of amplitude is injected between BPTT steps to simulate long-horizon covariate shift.20
Training-loss convergence curves over the 300-epoch curriculum are monitored on the held-out validation split. For the M0 baseline, which skips Eulerian pre-training Stage 1, the same total of 300 epochs is spent in Stages 2–4 as 75 : 150 : 75 (folding Stage 1’s allocation into the Stage 2–4 ratio 60 : 120 : 60), with and left at their defaults. The complete hyperparameter list is reported in Table I.
Results and discussion
Evaluation metrics and headline performance
Quantitative performance is assessed with three complementary metrics on the held-out rollout (260-frame autoregressive trajectory, of physical time). The production checkpoint analysed in this section is trained and evaluated on the representative case Sweep_Case_03 of the twenty-case CFD design space.
The mean displacement error (MDE) is
where is the surrogate position, the CFD reference position of parcel , the number of alive parcels, and (room width). The time-averaged MDE is the primary trajectory-fidelity statistic.
The kinetic-energy ratio (KE-ratio) is
where is the predicted velocity, the CFD reference velocity, and each sum runs over alive parcels at time . The ratio diagnoses numerical kinetic-energy dissipation independently of any systematic translation error; KE-ratio is exact conservation.
The radius-of-gyration error (Rg-err) is
where is the cloud centroid for whichever particle coordinates are inserted into the sum, and the superscripts pred and GT indicate that is built from surrogate-predicted and CFD positions, respectively (the same alive parcels as in Eqs. (23)–(24)). The metric quantifies how well the surrogate reproduces the spatial spread of the aerosol cloud, which directly affects zone-averaged exposure estimates in infection-risk models.
Table III summarises the headline rollout metrics for the purely Lagrangian GNS baseline (M0) and the full ELGIN. Overall, ELGIN delivers substantially better trajectory and cloud-shape fidelity than M0, at the cost of a slightly lower kinetic-energy ratio (see below). Rollout-resolved traces of the same three diagnostics (MDE, cloud , instantaneous KE-ratio) for Sweep_Case_03 are plotted in Figure 3 alongside the headline scalars in Table III; § 6 C interprets each panel in detail.
| Model | MDE (%) | KE-ratio | Rg-err (%) |
| GNS (baseline; M0) | 19.56 | 1.057 | 9.85 |
| ELGIN | 16.20 | 0.659 | 6.58 |
Table III. M0 vs. ELGIN after a 260-frame autoregressive rollout (s).
Figure 3. Rollout diagnostics for M0 and ELGIN on Sweep_Case_03. (a) Mean displacement error (Eq. (23)) for M0 (red, with per-particle band) and ELGIN (green); reference m. (b) Radius-of-gyration trajectories: foam-extend CFD reference (blue), ELGIN rollout (green), and M0 rollout (dashed red). (c) Instantaneous kinetic-energy ratio following Eq. (24); KE-ratio = 1 is exact conservation.
The unconstrained GNS (M0) trained for 300 epochs reaches MDE = 19.56% of , i.e. a mean absolute displacement error of in a -wide room, together with KE-ratio and Rg-err . Longer training markedly improves this Lagrangian-only baseline relative to weaker checkpoints; the near-unity KE-ratio (1.057) in particular shows that aggregate parcel kinetic energy is matched with only modest over-estimation compared with under-trained M0 states.
By contrast, ELGIN cuts the mean displacement error to MDE of ( in absolute units, lower than M0) and the radius-of-gyration error to Rg-err $= 6.58%
( % lower). These gains align with supplying each parcel with interpolated RANS velocity and turbulence features at every step, in a flow where droplets follow the mean carrier quickly ( on the ventilation time scale, well inside the tracer-like regime defined by Balachandar and Eaton⁴⁸), while transport relative to that mean remains advection-dominated (turbulent Péclet number , with as in SI § 25), so accurate local advection drives long-horizon cloud statistics.
Cloud-snapshot and trajectory comparison
Figure 4 presents three-time-step particle-cloud snapshots for Sweep_Case_03, evaluated with the M0 and ELGIN production checkpoints.
For M0, the mean displacement error grows from MDE % of at s to % at s as the cloud enters regions of stronger carrier-velocity variation, so small one-step errors compound in the autoregressive rollout. By s the trajectory registers NaN positions on % of frames after parcels are flagged outside the domain bounding box (last fully-valid MDE %). The same figure shows ELGIN staying nearer the reference at those instants (MDE % and % of ) without the late-time mass loss to out-of-domain flags that limits M0, consistent with Table III. That M0 error growth and boundary escape match the autoregressive compounding expected for rollout surrogates²⁰,²⁵ and, in this Euler–Lagrange problem, the residual information bottleneck of a purely Lagrangian baseline, as discussed next.
The temporal error-growth pattern observed in Figure 3(a) has a clear physical origin rooted in the Euler–Lagrange dynamics of the problem. The M0 baseline operates in a purely Lagrangian mode: particle acceleration is inferred from inter-particle neighbourhood interactions, position history, and wall-proximity features alone, with no access to the background Eulerian air-velocity field. In the dental aerosol problem, aerodynamic coupling to the resolved carrier field is the dominant transport mechanism at the mesh scale: the ventilation Stokes number with m and m from the twenty-case sweep satisfies (SI § 24), so droplets track the resolved RANS velocity almost instantaneously on the time scale, while transport relative to that mean field remains advection-dominated (; Sec. VE and SI § 25). The velocity of this carrier phase varies strongly in space due to the ceiling inlet jet, the lateral exhaust outlet, and the deflection of the mean flow around the dentist and patient body obstacles. During the injection phase (), all particles are localised near the source nozzle, their separation is small relative to , and the neighbourhood graph is dense; the model correctly reproduces the near-source ballistic injection because the local flow conditions are largely uniform and well-represented in training data. As the cloud disperses and individual particles enter regions with strongly varying carrier-phase velocities (the dentist-side recirculation zone, the obstacle wakes, the ventilation throughflow corridor), the M0 graphs become sparser and the model must extrapolate beyond the spatial correlations learnable from particle kinematics alone. This produces the growing position scatter characteristic of Lagrangian surrogates applied to externally driven particle-in-fluid problems, and is qualitatively distinct from the numerical compounding error observed in the original particle-as-fluid GNS benchmarks (sand, water, goop) where inter-particle forces are themselves the dominant dynamics.20
The ELGIN eliminates this information bottleneck by exposing each particle node, through the analytic Cunningham-corrected drag and turbulent-kinetic-energy features (Eq. (10)), to the cell-interpolated RANS velocity and TKE produced by the Eulerian sub-network and pressure projection (§ 3 B). Because the Eulerian network operates on the obstacle-aware polyMesh (with dentist and patient no-slip patches), the predicted, divergence-corrected velocity field naturally encodes flow deflection around the dental staff and patient, including the recirculation zones that trap mid-range droplets at breathing-zone height. The dentist and patient boundary conditions are therefore handled implicitly in the ELGIN through the carrier-field features, so no explicit contact or obstacle-avoidance model is required; this is the primary mechanism behind the superior Rg-err and BZE fidelity reported in Table III.
A distinguishing feature of ELGIN relative to the Lagrangian-only M0 baseline is that it jointly advances a mesh-resolved carrier state with the parcel cloud. At each autoregressive step the Eulerian sub-network predicts the five RANS fields () on the OpenFOAM polyMesh (§ 3 B); the resulting rollout archive stores the full time series of velocity, modified kinematic pressure, turbulent kinetic energy (TKE), and on every cell together with the parcel trajectories. The M0 GNS variant has no Eulerian branch and cannot output any of these Eulerian quantities; only the particle positions and velocities produced by its Interaction Network.
Figure 5 visualises the comparison that is most directly informative for aerosol advection: the instantaneous speed on unstructured cell centroids with every parcel superimposed at matching physical times for Sweep_Case_03. Left: ELGIN fields from the predictive rollout; Right: OpenFOAM RANS reference velocities and CFD parcel coordinates on the same mesh and time grid. The pattern (nozzle-directed jet toward negative , ceiling supply, deflection past the clinician and patient, and egress toward the right-hand opening) matches between prediction and reference; residuals are concentrated in narrow shear layers downstream of the injector and obstacle edges, i.e. where the snapshots in Figure 4 also show the strongest differences in lateral cloud spread.
Figure 4. Particle cloud snapshots at , , and for Sweep_Case_03 from the twenty-case factorial sweep (exact , values in SI Table S2); Top row: foam-extend 4.1 reactingParcelFoam ground truth. Middle row: baseline GNS (M0). Bottom row: ELGIN. Colour encodes each particle’s radial distance from the injection source at the patient’s oral cavity (star marker at , ). Dashed rectangle: dentist breathing zone. The cloud disperses leftward with increasing time; M0 already shows MDE at and at , while ELGIN reaches MDE and at the same times respectively (Table III).
Time-resolved error and energy evolution
As plotted in Figure 3, panels (a)–(c) use the same Case 03 rollout archives as the other qualitative figures, with samples on (; the axis in the figure runs to 30 s but stored data end near for ELGIN). Panel (a) plots mean displacement error from Eq. 23; panel (b) plots built directly from instantaneous parcel positions (Eq. 25); panel (c) evaluates Eq. 24 using backward finite-difference velocities between consecutive stored positions at fixed , the same construction used to populate Table III. (a) The red trace and shaded band show the M0 baseline rollout on its archived CFD pairing, with the band giving the per-particle spread of the displacement error around the cloud mean at each frame. The green curve is the paired ELGIN rollout for the same case against the CFD reference released alongside it. After a short initial transient, the ELGIN curve lies below the M0 trace for essentially the remainder of the window; for example MDE vs. at , and for ELGIN at , while M0 accumulates NaN-masked parcels over the last of frames, with last fully-valid MDE near , all with as in Eq. 23. These pointwise values are consistent with the aggregate ordering in Table III and with monotonic rollout error growth in discrete GNS benchmarks.[25] [20] (b) The blue reference is CFD for Sweep_Case_03; green is the predictive ELGIN rollout and dashed red is the M0 rollout on the same case. ELGIN tracks the widening aerosol halo significantly more faithfully than M0 throughout the transient, consistent with the reduction in time-averaged -err reported in Table III. (c) uses the backward-difference velocities noted above. The M0 curve (300 epochs) hugs unity during the early part of the rollout but develops large late-time fluctuations once masking and positional errors make the sums in Eq. (24) ill-conditioned, even though the time-averaged ratio in Table III remains . The ELGIN curve shows a systematic downward drift to roughly 0.4–0.5 by late time, with time-mean , signalling under-estimation of parcel speed magnitudes. The improvement of ELGIN over M0 in MDE and -err therefore reflects primarily superior spatial trajectory fidelity from the RANS coupling, not kinetic-energy conservation per se.
Clinical metric: Breathing Zone Exposure
The Breathing Zone Exposure (BZE) fraction is the share of the simultaneously active parcels that lie inside the dentist’s breathing-zone rectangle , at time . With the alive parcel count in the evaluator subgraph and the subset inside that rectangle, gle,
Figure 5. Eulerian on polyMesh centroids with parcel overlays: ELGIN rollout (left) vs. OpenFOAM RANS reference with matched CFD parcels (right); , , , Sweep_Case_03. Per row, colours share a cap set by maxpatches across both columns at that time. Grey: obstacle masks from boundary tagging; markers: ceiling inlet and lateral outlet.
matching the persistent-origId masks used for the headline trajectory metrics and Figure 6. The box is centred on the dentist’s head height directly above the obstacle ridge of the polyMesh and has dimensions ( horizontal, vertical) consistent with the typical clinician inhalation volume. Thus is a geometry-based surrogate, the fraction of evaluated parcels intersecting a fixed inhalation prism; it is not a complete inhaled-dose or site-specific infection-risk quantification, which would couple breathing-rate transients, filtration, and pathogen dose–response relations outside the present flow–particle surrogate.[5] The BZE time series is nonetheless the primary scalar clinical output used here for infection-risk screening alongside transport fidelity.
Figure 6. Breathing Zone Exposure (BZE) fraction vs. time for Sweep_Case_03 (, ). CFD ground truth (solid black), GNS (M0) baseline (dashed blue, same 1000-parcel subgraph as the ELGIN evaluator), and ELGIN (green) on the matched 1000-parcel evaluator. Breathing-zone rectangle: , . On this case the ELGIN reconstruction tracks the reference BZE envelope with root-mean-square error (RMSE) and absolute discrepancy at the reference peak.
For the rollout shown in Figure 6 (Sweep_Case_03: , ), the CFD reference produces a short-lived aerosol filament that intercepts the breathing zone with peak BZE , decaying to near-zero over the remainder of the rollout horizon as parcels leave the breathing-zone rectangle along the dentist-side recirculation streamlines. On the -parcel evaluator subgraph this peak corresponds to roughly five parcels simultaneously inside the breathing-zone rectangle, so the BZE fraction is intrinsically in the Poisson-counting-noise regime ( at the peak). Evaluated on the same 1000-parcel subgraph and identical time grid, the ELGIN surrogate produces a peak BZE (of order ten parcels), an absolute peak discrepancy of and a root-mean-square error of in BZE fraction across the rollout; in relative terms this is a factor-of-two over-estimate on a small base and is therefore meaningful only at the order-of-magnitude level on the present 1000-parcel evaluator. On the full -parcel CFD cloud the same BZE fraction is statistically much more robust, and will be reported alongside the multi-case retraining. The M0 baseline is evaluated on the same particle subset for visual comparison but accumulates a much larger BZE bias by late time (reflecting the spatial drift quantified in Table III via Rg-err); we therefore interpret M0’s BZE trace as illustrative of the information-bottleneck failure mode rather than as a clinically usable prediction.
The BZE characteristics identified here are consistent with the experimental aerosol observations of Micik et al.[2] and Harrel and Molinari,[1] the ventilation-rate dependence of peak BZE across the twenty CFD cases (§ 6 E) is consistent with the airborne pathogen concentration scaling reviewed by Morawska and Milton.[5]
Ventilation scaling and dispersion validation
To confirm that the CFD training data exhibit physically expected ventilation and dispersion behaviour, and therefore that the GNN surrogates inherit physically meaningful scaling laws, two non-dimensional analyses are summarised below. Extended results, additional figures, and full derivations are provided in SI Secs. S5 and S6.
Figure 7 reports the peak transient BZE as a function of the room ventilation rate, measured in air changes per hour (ACH), for all twenty CFD cases. A clear power-law decrease of peak BZE with increasing ACH is observed, with the slope of a single log–log least-squares fit across all twenty cases reported in Figure 7; the empirical exponent of order is weaker than the steady-state well-mixed-room scaling of 49, but this comparison is indicative rather than strict because the well-mixed prediction concerns the time-averaged concentration in fully developed steady state, while the present exponent describes the transient peak during the short BZE intercept event. The lowest ventilation level (, ACH , below the ASHRAE Standard 170 guideline of for airborne-infection isolation rooms50) corresponds to the highest peak-BZE values across the sweep, while the highest ventilation level (, ACH ) gives the lowest. The sub-linear scaling reflects the role of obstacle-driven recirculation zones that trap particles near the breathing zone for longer than the mean residence time, motivating the carrier-field conditioning of ELGIN that supplies recirculation information directly to each particle node.
Figure 7. Peak BZE vs. Air Changes per Hour (ACH) for all 20 CFD cases. Points coloured by spray velocity , shaped by spray angle . Vertical dashed line: WHO/ASHRAE 12-ACH healthcare standard.50 Dashed curve: power-law fit (slope ), weaker than the well-mixed prediction () due to obstacle-driven recirculation.
Figure 8 shows the non-dimensional longitudinal () and transverse () mean-squared displacements (MSDs) defined in SI Sec. S6 for all CFD trajectories, computed across the full reactingParcelFoam parcel dataset (5 300–7 220 parcels per case). Let be the MSD lag time (time since the reference frame) and its non-dimensional counterpart, with room height (SI Sec. S5). Both components display the classical transition from the short-time ballistic regime () to the long-time diffusive regime (),51,52 with the transition at – in order-of-magnitude agreement with the dimensional time scale derived from the ceiling-inlet specific-dissipation boundary condition (SI Tables S1–S2; we use as a proxy for the Lagrangian integral time scale, recognising that the canonical Pope expression differs by a constant factor[10]). Recalling the turbulent Péclet number from § 6 A, its value is essentially invariant across the sweep at (coefficient of variation ). This invariance is a direct consequence of the chosen inlet boundary conditions, which fix the turbulence intensity for every case so that , and therefore ; it is therefore a property of the chosen BC design rather than an emergent physical invariant. At advective transport nevertheless dominates turbulent diffusion by four orders of magnitude, confirming that aerosol fate is controlled by the mean flow pattern, a prerequisite for the GNN’s flow-feature-based prediction approach.
Additional analyses (Supporting Information)
A comprehensive set of additional analyses, supporting the main results above, is provided in the Supporting Information. SI Sec. S4 presents the force-magnitude regime map across , justifying the inclusion of Cunningham, Saffman, Brownian, and DRW models in the GNN framework. SI Sec. S5 complements Figure 7 with tabulated non-dimensional groups for all twenty cases and the Stokes–settling regime map coloured by peak Breathing Zone Exposure. SI Sec. S6 extends the Taylor-framework dispersion analysis introduced in Figure 8 with the longitudinal/transverse dispersion coefficients and and their scaling with ACH and , the dispersion anisotropy ratio , and the vertical drift decomposition. SI Sec. S7 documents inference wall times, end-to-end CFD versus surrogate speed-up, training wall time, peak GPU memory use,
Figure 8. Non-dimensional mean-squared displacement vs. non-dimensional lag time for all 20 CFD cases, coloured by ACH. Left: longitudinal . Right: transverse . Power-law guides (ballistic) and (diffusive) shown as dotted and dashed lines. The transition at – is consistent with the Lagrangian integral time scale . The turbulent Péclet number is invariant across the sweep (), confirming advection-dominated transport.
and comparison with published GNN/ML-CFD surrogates.
Conclusion
This work has presented ELGIN, a physics-informed hybrid Eulerian–Lagrangian Graph Neural Network surrogate for polydisperse dental bioaerosol dispersion in enclosed clinical spaces. On the representative case Sweep_Case_03 of a twenty-case foam-extend 4.1 reactingParcelFoam parameter sweep, ELGIN achieves a mean parcel displacement error of 16.20 % and a cloud radius-of-gyration error of 6.58 %, representing % and % reductions over the Lagrangian-only baseline (M0), respectively, while completing a 26 s rollout in s, approximately faster than the reference CFD solver. These numbers are a single-case demonstration; the analogous evaluation on the full 16/2/2 train/validation/test split is in progress and resulted will updated.
The accuracy gains stem from three interacting design choices. First, by solving the carrier flow on the OpenFOAM polyMesh through a Graph Transformer with a Jacobi-preconditioned pressure projection, ELGIN exposes every parcel node to the local RANS velocity, turbulent kinetic energy, and wall proximity, i.e. physical cues that a purely Lagrangian model cannot access. Second, geometry- and inlet-aware conditioning, combining nine semantic boundary-class embeddings, true wall-distance and wall-normal features, and a per-case inlet-velocity broadcast, supplies the information that a single trained model would need in order to generalise across the ventilation parameter sweep; demonstration of this generalisation in practice awaits the 16/2/2 retraining. Third, a four-stage curriculum (Eulerian pre-training, one-step supervised particle prediction, PDE-informed joint training, and BPTT rollout fine-tuning) provides a transferable stabilisation protocol for long-horizon multi-physics GNN surrogates.
The present results carry four important caveats. First, the production checkpoint is trained and rolled out on a single case (Sweep_Case_03) of the twenty-case sweep; the full 16/2/2 retraining required to demonstrate cross-case generalisation is under way. Second, all training data derive from a two-dimensional cross-sectional domain; extension to full 3-D geometry is the most critical structural next step. Third, the carrier flow is frozen at its steady-state RANS solution, excluding ventilation transients and occupant-induced unsteadiness; coupling ELGIN to time-evolving LES snapshots would remove this constraint. Finally, surrogate accuracy has been validated against CFD reference only; optical particle counter or phase-Doppler anemometry measurements in a clinical mock-up[1, 2] are required before clinical deployment.
Despite these limitations, the single-case results indicate that physics-informed hybrid Eulerian–Lagrangian GNN surrogates can faithfully reproduce the multiscale dynamics of clinical dental aerosol dispersion at speeds compatible with per-appointment risk screening, providing an architecture and training protocol on which the full multi-case study, and future three-dimensional, experimentally validated surrogates for infection-risk management in healthcare settings, can be built.
Acknowledgments
Takshak Shende thanks Professor Ian Eames and Professor Emad Moemendarbary (University College London) for their leadership of the EPSRC-funded programme that framed airborne dispersal and infection-relevant transport questions in dental clinical settings. Takshak Shende acknowledges support as a research fellow on that programme. This work was supported by the Engineering and Physical Sciences Research Council under grant EP/W009889/1.
Author Declarations
Conflict of Interest
The authors have no conflicts to disclose.
Ethics Approval
This study is a computational and numerical investigation involving no human or animal subjects. No ethics approval was required.
Use of artificial intelligence tools
During manuscript preparation, the authors used artificial-intelligence assistants (large language-model-based tools) for language editing and for code debugging support. Scientific ideas, methodology, numerical experiments, results, and interpretation were developed by the authors.
Author Contributions
Takshak Shende: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Project administration; Resources; Software; Validation; Visualization; Writing: original draft; Writing: review and editing. Viktor Popov: Resources; Supervision; Writing: review and editing.
DATA AND CODE AVAILABILITY
The ELGIN source code, the foam-extend 4.1 reference CFD case (dentalRoom2D), pre-trained model weights, and animated rollout comparisons are openly available at https://github.com/takshakshende/ELGIN under the MIT licence. Detailed CFD methodology, force-magnitude analysis, non-dimensional and dispersion analyses, and computational benchmarks are provided in supporting information Secs. S1–S7.
Supporting Information for: Physics-Informed Graph Neural Network Surrogates for Turbulent Nanoparticle Dispersion in Dental Clinical Environments
Takshak Shende; Viktor Popov
This Supporting Information provides additional detail on the CFD methodology, force-model derivations, full literature comparison, training loss formulation, non-dimensional and dispersion analyses, and computational benchmarks that complement the main manuscript. The present work reports a single-case demonstration on Sweep_Case_03; the planned 16/2/2 multi-case retraining will update all surrogate metrics.
CFD methodology: detailed formulation
Boundary conditions
The two-dimensional dental treatment room (, main text Figure 1) is discretised on a structured Cartesian quadrilateral mesh of nominal cells (uniform cell size ). After removing cells that coincide with the dentist and patient obstacle patches, 7 704 active fluid cells are retained; these form the nodes of the Eulerian graph used by the GNN. A mesh-independence study on three resolutions (, , ) confirmed that the time-averaged breathing-zone particle concentration differs by less than between the medium and fine meshes, justifying the medium resolution for the training sweep. The complete boundary-condition assignments for each physical quantity are listed in Table S1.
| Boundary | Particles | ||||
| Ceiling inlet | FV | ZG | FV | FV | – |
| Side outlet | ZG (pressure-inlet) | FV 0 | ZG | ZG | escape |
| Walls | no-slip | ZG | FV 0 | NW | wall-deposition |
| Nozzle inlet | FV | ZG | FV | FV | injection |
Table S1. Boundary conditions applied in all foam-extend 4.1 cases.
Note. “FV” = fixedValue; “ZG” = zeroGradient; “NW” = standard wall function. = unit vector in the (downward) direction; = nozzle direction unit vector at cone half-angle . ; with (inlet mixing length) and (– SST closure constant).
Carrier-phase governing equations
The carrier airflow is modelled using the incompressible Reynolds-Averaged Navier–Stokes (RANS) equations with the – Shear Stress Transport (SST) turbulence closure of Menter,[1] which captures both the free-stream turbulence in the room interior and the near-wall adverse pressure gradients behind the dentist obstacle:
with the Favre-averaged mean velocity, the modified kinematic pressure, the kinematic viscosity of air at 20 °C, the turbulent eddy viscosity, and a momentum source term from particle back-reaction (active only in the coupled regime). The eddy viscosity is closed by:
where is the strain-rate tensor, the corresponding strain-rate magnitude, is the Menter blending function, and .[1]
Particle-phase governing equations
Each airborne particle of instantaneous diameter , density (liquid water at 20 °C), and mass is tracked individually in the Lagrangian frame. Under the dilute-phase assumption (), the simplified Maxey–Riley equation reads:[2]
with the parcel velocity, , and , , , and specified in the subsections below. The neglected Basset history force, added-mass force, and Faxén correction are all and are standard to omit for saliva droplets in air.[2] Equation (S4) is integrated by the reactingParcelFoam parcel tracker using a semi-implicit first-order scheme with the built-in momentumTrackingTime and maxCo controls, so that the effective parcel sub-step never exceeds the carrier Courant cap (, ) and is further clipped to a fraction of the local Stokes relaxation time for sub-micrometre droplets where becomes small.
1. Stokes drag with Cunningham correction
Particles in the Stokes regime (particle Reynolds number ) satisfy
where is the Stokes relaxation time, is the Cunningham slip-correction factor,[3] with the empirical coefficients (1.257, 0.400, 1.10) taken from the Millikan-apparatus calibration of Allen and Raabe.[4] is the dynamic viscosity, and is the molecular mean free path of air at 101.325 kPa, 293 K. The Cunningham factor rises from at to at , enhancing the effective drag coefficient and thereby increasing diffusive transport of sub-micrometre residues. For non-negligible particle Reynolds number near the high-speed nozzle exit the foam-extend SphereDrag sub-model uses the standard Schiller–Naumann drag correlation[5, 6]
which reduces smoothly to the Stokes expression in Eq. (31) as .
Saffman shear-lift force
Particles with near surfaces experience a Saffman shear-lift in the small- limit4
where points from the wall toward the flow interior, and is the McLaughlin correction factor5 with . For the dental-aerosol parameter range ( for , near-wall shear ) we have for which to better than 5%, so the Saffman formula is used directly. Near the patient obstacle, can reach , producing lift forces of order – for – particles and deflecting mid-size droplets toward the breathing zone.
Brownian motion for sub-micron particles
For the Einstein–Smoluchowski random displacement is:6
with and . At each particle sub-step, an independent Gaussian displacement is added to the deterministic trajectory. For a nucleus, per timestep, comparable to the deterministic carrier-airflow displacement.
Turbulent dispersion (Discrete Random Walk)
Turbulent velocity fluctuations unresolved by the steady RANS solution are sampled from Pope’s Discrete Random Walk (DRW) model:7
with a fluctuation correlation time (), beyond which a new realisation is drawn independently.
Wells’ -law evaporation in the foam-extend reference
Freshly generated saline spray droplets evaporate quasi-steadily inside the foam-extend LiquidEvaporation sub-model according to:8
with vapour mass diffusivity and Spalding transfer number at 50% relative humidity. Substituting the model parameters gives
A droplet therefore reaches its equilibrium nucleus diameter after
and reaches complete evaporation in approximately . These times are of the order of, but somewhat longer than, the GNN snapshot interval , so for the largest droplets the per-step diameter change is not negligible; the practical implication for the surrogate is discussed below.
a. Treatment of inside the GNN. The hybrid GNN surrogate of this paper does not predict diameter changes. Instead, the parcel diameter is treated as a static per-parcel input feature in (10) of the main text, inherited from the Rosin–Rammler initial sampling of the ConeInjection (Sec. S1 S1.5). This is justified by two complementary observations. (a) The per-step fractional change in is (26%) for the largest droplets, but drops below 1% for , and below 0.1% for . Because sub-10 droplet nuclei dominate aerosol exposure risk in the clinically relevant size class, the static- approximation is essentially exact over that fraction. (b) For the larger end of the spectrum (10 to 50 ), the foam-extend trajectory targets already incorporate the inertia reduction caused by Wells’ law, so the GNN learns a position–velocity mapping that implicitly absorbs the diameter drift from the target data, without having to track it explicitly in the state vector; fidelity of that absorption is part of what is being measured by the MDE and Rg-err metrics in main-text Table III. Diameter dynamics, if required by a downstream infection-risk calculation, can be applied as an analytic post-processing step on the predicted trajectories.
Implementation in foam-extend 4.1
The CFD reference data are produced by the foam-extend 4.1 reactingParcelFoam solver,[12] which natively integrates Schiller–Naumann drag (SphereDrag), gravity, the DRW turbulent dispersion model (StochasticDispersionRAS), and Wells’ -law evaporation (LiquidEvaporation via the Spalding and the enthalpyDifference latent-heat sink) on the polydisperse ConeInjection cloud. The Cunningham slip-correction factor and the McLaughlin–Saffman shear-lift are not shipped as parcel sub-models in foam-extend 4.1; both are supplied analytically inside the GNN edge encoder from the local parcel diameter and the carrier-phase gradient exported alongside , , and at every CFD output step. Brownian diffusion is also applied analytically in the GNN node update for , since the reactingCloud class does not provide a built-in Brownian sub-model.
Particle injection and size distribution
A continuous ConeInjection of computational parcels emanates from the patient’s oral cavity throughout the full 30-second simulation. The injection is centred at , aimed horizontally toward the dentist (), and dispersed within a cone of half-angle . The injection rate is parcels/s, with volumetric flow rate (5 mL min), giving a total injected mass of mg over 30 s. Each parcel statistically represents a packet of physical droplets whose diameters are drawn from the Rosin–Rammler distribution.[13]
with diameters truncated to . Parcels are injected at body temperature and subject to standard wall interaction (escape on contact with no-slip surfaces).
Solver and numerical setup
Each case is integrated as a two-stage pipeline (Table S2). Stage 1 (steady-state RANS): simpleFoam (residual-controlled SIMPLE algorithm) is run until every component of falls below a relative residual. The wall treatment uses the OpenFOAM/foam-extend –
SST wall functions (kqRWallFunction, omegaWallFunction, nutkWallFunction), which automatically blend the linear viscous sublayer () and the logarithmic region () through Spalding’s law of the wall.3 For the present mesh the cell-centre wall distance gives in the range 3–15 on the dentist, patient, floor, ceiling and side-wall patches at the highest ventilation case (), well inside the validity window of the SST blended near-wall formulation.
Stage 12 transition: the kinematic pressure is overwritten with absolute pressure , wall functions are switched to compressible counterparts, and three thermophysical fields are initialised: temperature , ambient water-vapour mass fraction (50 % RH), and . Stage 2 (transient compressible reacting parcels): reactingParcelFoam advances the transient compressible Navier–Stokes equations together with species transport for of physical time using PIMPLE (merged PISO–SIMPLE) pressure–velocity coupling. The thermophysics template reads NASA-7 polynomial coefficients for and ; chemical reactions are switched off. The Lagrangian reactingCloud1 is integrated concurrently with the carrier phase (coupled true, cellValueSourceCorrection on). The time step is adaptive with and , giving a typical step of in the spray core. Solution snapshots and parcel state are written every via adjustableRunTime, providing frames per case.
One-way coupling regime. The peak particle volume fraction near the nozzle exit is , well below the onset of dynamically significant two-way coupling identified by Balachandar and Eaton.[14] The contribution of the parcel back-reaction on the carrier momentum is therefore small but is retained for completeness.
Parameter sweep and dataset
Twenty cases (Sweep_Case_01–Sweep_Case_20) span a full factorial grid over the two clinically dominant control parameters: the ceiling supply velocity , representing HVAC settings from minimal to high ventilation; and the dental handpiece nozzle injection speed , covering air-polishers to high-speed turbine drills. The spray cone half-angle is held at across the sweep, matching the median Rosin-
Rammler dental-handpiece aerosol cone reported by Harrel and Molinari.[13] Cases are split 80/10/10 (train/validation/test) by random shuffle with a fixed seed (16/2/2 for the present sweep). The full case list is given in Table S3. Dataset extraction statistics are summarised in Table S4.
| Case | (m s) | (m s) | (°) |
| Sweep_Case_01 | 0.10 | 10 | 20 |
| Sweep_Case_02 | 0.10 | 20 | 20 |
| Sweep_Case_03 | 0.10 | 30 | 20 |
| Sweep_Case_04 | 0.10 | 40 | 20 |
| Sweep_Case_05 | 0.10 | 50 | 20 |
| Sweep_Case_06 | 0.20 | 10 | 20 |
| Sweep_Case_07 | 0.20 | 20 | 20 |
| Sweep_Case_08 | 0.20 | 30 | 20 |
| Sweep_Case_09 | 0.20 | 40 | 20 |
| Sweep_Case_10 | 0.20 | 50 | 20 |
| Sweep_Case_11 | 0.35 | 10 | 20 |
| Sweep_Case_12 | 0.35 | 20 | 20 |
| Sweep_Case_13 | 0.35 | 30 | 20 |
| Sweep_Case_14 | 0.35 | 40 | 20 |
| Sweep_Case_15 | 0.35 | 50 | 20 |
| Sweep_Case_16 | 0.50 | 10 | 20 |
| Sweep_Case_17 | 0.50 | 20 | 20 |
| Sweep_Case_18 | 0.50 | 30 | 20 |
| Sweep_Case_19 | 0.50 | 40 | 20 |
| Sweep_Case_20 | 0.50 | 50 | 20 |
Table S3. Per-case (, , ) for the twenty foam-extend 4.1 reactingParcelFoam runs (Sweep_Case_01–Sweep_Case_20).
| Quantity | Value |
| Total cases | 20 (planned 16 / 2 / 2 split, fixed seed) |
| Cases used in present checkpoint | 1 (Sweep_Case_03) |
| Physical time per case | 30 s |
| Extraction window | |
| Save interval | |
| Frames per case () | 261 |
| Active parcels (mean) | at quasi-steady state |
| Tracked parcels per snapshot | 1000 (persistent origId) |
| Graph snapshots / case (full sweep) | ( for 16-case train split) |
| Lagrangian connectivity radius | 0.10 m (ELGIN); 0.30 m (M0) |
| Lagrangian edges per snapshot | (ELGIN); (M0); directed both senses |
| Node features per Lagrangian node | 15+ (LSTM history, SDF, drag, , TKE, , ) |
| Edge features per Lagrangian edge | 4 (rotation-invariant local frame plus Cartesian drag; Eq. (9) of main text) |
| Raw trajectory storage | GB |
| Processed graph tensors | GB |
| BZE range across CFD sweep | subject to BZ-rectangle definition; see § 6.D of main text |
| Deposition fraction range | 0.12 – 0.61 across the twenty CFD cases |
Table S4. Numerical summary of the graph-tensor training dataset extracted from the twenty-case reactingParcelFoam sweep. The CFD case grid, train/validation/test split.
Comprehensive GNN-CFD literature review
Table S5 presents a comprehensive comparison of GNN and machine-learning methods applied to CFD problems, summarising methodology, key findings, advantages, and limitations of each approach. The “present work” rows confirm the headline metrics reported in main text Table II.
Note. Key: EPD = Encode-Process-Decode; SPH = Smoothed Particle Hydrodynamics; MPM = Material Point Method; FV/FE = Finite Volume/Element; RANS = Reynolds-Averaged Navier–Stokes; LES = Large-Eddy Simulation; UQ = Uncertainty Quantification; DDP = Distributed Data Parallel. “+” = advantage; “−” = limitation. {sup}†`The manuscript emphasises comparisons between standalone GNS (M0) and Eulerian-conditioned ELGIN; M2 labels an archived GAT–GNN configuration retained in this table for literature context. M1 was an intermediate design superseded during development and is not reported separately.
Training loss formulation and curriculum
Total loss components
The four-stage training curriculum (main text Sec. IV) uses the following total loss in Stage 3. The continuity and momentum terms are physics-informed residuals, penalising discrete mass and momentum imbalance in the objective in the same spirit as PDE-constrained neural surrogates[17] and residual-corrected finite-volume formulations.[16]
with default weights . Each component is defined below. No evaporation loss is included, since the GNN does not predict diameter changes (§ 16 § 16.3).
Particle MSE loss.
where is the per-component variance of ground-truth accelerations computed over the training set.
Continuity residual. , evaluated by face-flux summation on the Eulerian mesh graph, i.e. a graph discretisation of the divergence penalty consistent with mesh-native learned discretisations such as MeshGraphNets.[18]
Momentum residual. is the discrete L2 residual of the RANS momentum equation (Eq. (28)) on the Eulerian mesh graph. It is assembled in the finite-volume manner (control volumes, face fluxes, and volume-weighted cell averages as in the standard FV method27) and evaluated on the mesh graph in the same residual-on-graph spirit as mesh-native GNN simulators.[16, 18] The loss is normalised to be scale-independent:
where is the pointwise finite-volume residual of Eq. (28) on the mesh graph, is the reference shear rate, and . The time derivative is discretised with , consistent with the CFD snapshot interval. The angle brackets denote a volume-weighted average over mesh cells. The convective flux uses a symmetric face average rather than a first-order upwind flux, limiting spurious numerical dissipation in the residual (cf. centred versus upwind convective differencing in FV methods27).
Turbulence-closure residual. enforces the SST – algebraic eddy-viscosity relation between the predicted , , and the inferred at every cell, as in the baseline two-equation closure of Menter.[1]
Angular-momentum penalty. A weak penalty on the residual change in the cloud-averaged angular momentum about the cloud centroid, , regularises the network towards rotationally consistent step-to-step displacements, where is the change in the depth-averaged out-of-plane () component of angular momentum of the particle cloud between successive prediction steps. This term is small in magnitude () and serves primarily as a numerical regulariser preventing solid-body spin-up of the cloud during long-horizon BPTT rollouts, where small forcing biases can accumulate across autoregressive steps.[21]
BPTT rollout loss (Stage 4)
Stage 4 uses back-propagation through time (BPTT) over autoregressive unrolling, following the rollout-training recipe introduced for graph network simulators.[21] In Stage 4 the model is unrolled for steps with the predicted state of step used as input to step :
The Stage 4 total loss is a convex combination of the one-step and the rollout objectives: , with and rollout steps. Gradient checkpointing every five rollout steps keeps GPU memory bounded. An additional per-step positional noise of amplitude is injected between BPTT steps to simulate the long-horizon covariate shift that would otherwise prevent the model from learning to recover from its own prediction errors.[21]
Training loss curves
Training and validation mean-squared-error (MSE) traces for the production Sweep_Case_03 checkpoints are produced by the respective trainers: both ELGIN and the M0 baseline follow the four-stage CFD–GNN curriculum of main text Sec. IV with a 300-epoch budget. For ELGIN the per-stage allocation is 60 / 60 / 120 / 60; for M0, Stage 1 (Eulerian pre-training) is skipped and the same total is applied to Stages 2–4 as 75 / 150 / 75 (Sec. IV). Within each stage the loader holds out a random subset of graph snapshots for validation (a random_split of the in-memory dataset), i.e. case-internal batch hold-out rather than a separate CFD case. The runs quoted in main text § 6 (Table III) use this schedule; M0 still omits the Eulerian sub-network and carrier-phase losses in Stages 3–4 while sharing the staged optimisation loop. When only a single trajectory file is supplied, the baseline GNS trainer’s default split reuses that file for both training and validation (pilot-style configuration intended for pipeline checks), so the printed validation MSE is computed on the same case as training and is not a temporally or case-disjoint hold-out.
Geometry features and inlet conditioning
A defining feature of the ELGIN relative to the M0-only baseline is the explicit, name-aware injection of geometry and inlet information into both the Eulerian and the Lagrangian feature streams. Three sources of static geometric information are used.
a. True polyMesh distance-to-wall and wall-normal. For every owner cell in the polyMesh we compute the geometric distance to the nearest wall-type face by looping over all patch faces classified as wall, floor, ceiling, dentist or patient, taking the minimum face-midpoint distance, and storing the corresponding unit wall-normal . The 2-D extrusion frontAndBack empty patch is skipped during this loop so that the “every cell touches a wall” artefact of y-extruded meshes does not contaminate . Both fields are exported as additional channels of the mesh_graph.npz file and consumed by the Eulerian sub-network through Eq. (8). At every rollout step they are interpolated to the parcel positions by the same inverse-distance-weighted operator (Eq. (12)) used for the velocity field, and concatenated to the Lagrangian node features in Eq. (10).
b. Name-aware boundary-condition encoding. The OpenFOAM boundary dictionary is parsed by a regex-based classifier mapping patch names (e.g. ceilingInlet, leftWall, dentist) to nine semantic classes: interior, inlet, outlet, wall, floor, ceiling, dentist, patient and symmetryPlane/empty. Each cell adjacent to a non-interior, non-empty face inherits the corresponding integer identifier (with a first-wins rule that prevents the empty 2-D extrusion patch from overwriting a physical wall, dentist, patient, inlet or outlet on the same cell). A learned 8-dimensional embedding then converts these identifiers into Eulerian node features.
c. Per-case airInlet velocity vector. The per-case airInlet velocity is parsed directly from the foam-extend 0/U dictionary (the ceilingInlet patch’s value uniform (vx vy vz) entry) during data extraction, normalised by and broadcast to every Eulerian node as the global conditioning vector in Eq. (8). This single global feature distinguishes the twenty training cases from one another and lets a single trained model interpolate between ventilation rates without retraining.
FORCE MAGNITUDE ANALYSIS AND REGIME MAP
Beyond trajectory fidelity, it is important to verify that the GNN correctly represents the relative importance of each physical force across the dental aerosol size spectrum. Figure S1 presents the magnitude of each physical force per unit particle mass as a function of diameter , evaluated at representative near-wall dental surgery conditions: slip velocity , near-wall shear , room temperature , and relative humidity 50 %.
Figure S1. Force magnitudes per unit mass vs. particle diameter at typical dental room conditions (, ). Vertical dashed lines mark the Brownian–drag crossover () and Saffman activation ().
Drag and gravity. Cunningham-corrected Stokes drag and gravity are the dominant forces across the full clinical size range . The drag-to-gravity force ratio increases sharply as decreases: for , drag is approximately stronger than gravity; for , gravity and drag are comparable, explaining why large droplets settle to the floor rapidly while nuclei remain suspended for the full 30-second simulation.
Saffman shear-lift. The Saffman lift force becomes non-negligible (exceeding 5% of the drag magnitude) above and reaches approximately 25% of drag for in the near-wall shear layer. In the present production checkpoint the Saffman lift term is switched off (Sec. III C) because the single-case Sweep_Case_03 polydisperse spray is dominated by sub-30 droplets where drag and DRW dominate; enabling the term is part of the planned multi-case retraining, and the architectural hook is preserved through the carrier-field velocity gradient supplied by the Eulerian sub-network to each parcel node.
Brownian diffusion. Brownian diffusion dominates over drag below . At the Brownian random-walk step is to per parcel sub-step (the exact value depends on the local used by the reactingParcelFoam tracker; see Sec. S1 S1.3), which exceeds the deterministic drag displacement over the same sub-step. Sub-micron nuclei produced by complete droplet evaporation ( to ) therefore undergo diffusive rather than advective transport, enabling them to penetrate boundary layers and deposit on mucosal surfaces by diffusiophoresis. Like the Saffman lift, the Brownian model is switched off in the present single-case production checkpoint (the polydisperse cloud is dominated by droplets, for which Brownian motion is negligible relative to drag and DRW); it is essential for correctly predicting the long-time fate of sub-micron residues and will be re-enabled in the multi-case retraining and for future sub-micrometre nuclei studies.
Turbulent dispersion. The DRW stochastic force is not shown in Figure S1 because its magnitude depends on the local turbulent kinetic energy , which varies spatially. In the room interior (), the root-mean-square turbulent velocity fluctuation provides a forcing comparable in magnitude to Brownian diffusion for , while being the second most important force after drag for all larger particles.
Comparison with literature. The force hierarchy identified here (drag-dominated for , Brownian-dominated below , Saffman-relevant when lies in the 3 to 20 interval) is consistent with the theoretical predictions of Balan and Eaton[14] for dilute turbulent dispersed flows. The dental aerosol size distribution measured by Micik et al.28 and characterised by Harrel and Molinari[13] populates exactly the 3 to 20 interval where Saffman lift becomes a non-negligible correction, motivating its inclusion in the optional Lagrangian force terms (Sec. S1 S1.3) and its re-activation alongside the Brownian model in the planned twenty-case retraining.
NON-DIMENSIONAL AEROSOL TRANSPORT ANALYSIS
The twenty OpenFOAM cases provide a rich dataset for examining dental aerosol transport through classical non-dimensional fluid mechanics. Four non-dimensional groups are evaluated for each case: the particle Stokes number , the gravitational settling ratio , the air changes per hour ACH, and the jet Reynolds number . The Stokes number uses the mean Lagrangian diameter and may be written equivalently as , where is the Stokes relaxation time at that diameter and is the room convective time. Their definitions and computed ranges are collected in Table S6.
| Group | Definition | Range | Units |
| Stokes number | – | – | |
| Settling ratio | – | – | |
| Air changes / hr | – | ||
| Jet Reynolds number | – | – |
Table S6. Non-dimensional groups computed across the 20-case CFD sweep.
Stokes–settling regime map
Figure S2 presents the two-dimensional – regime map for all twenty cases, coloured by the peak Breathing Zone Exposure. All cases fall in the region , specifically , confirming that the dental aerosol particles respond to velocity fluctuations much faster than the flow time scale and behave as aerodynamically passive tracers of the carrier fluid. This is consistent with the classical criterion for tracer-like behaviour established by Balachandar and Eaton [14]. The settling ratio ranges from 0.029 at to 0.143 at , indicating that gravitational settling is secondary to ventilation-driven advection at high air-supply velocities but represents a non-negligible 14% of the air-speed at the lowest ventilation level. Wells [11] showed theoretically that droplets for which remain airborne long enough to complete full evaporation; the present confirms that all investigated size classes are in the airborne-transmission regime. The peak BZE correlates most strongly with rather than : the highest-BZE cases (BZE ) are those with largest (lowest ), because slow settling allows greater particle accumulation at breathing-zone height before ventilation clears the room.
Figure S2. Stokes number–settling ratio regime map for the 20-case CFD parameter sweep. Each point is one case coloured by peak Breathing Zone Exposure (BZE fraction). All cases fall in the tracer-like regime (). Marker shape distinguishes ventilation levels.
Note. (air), , (water droplet), (room height), (ceiling inlet area; 0.2-m-wide inlet with the 0.01 m 2-D extrusion depth used in the frontAndBack empty patch), (dental handpiece outlet), (pseudo-2D room volume after subtracting dentist and patient obstacle blocks at the same extrusion depth), the gravitational acceleration, and mean particle diameter (empirical parcel-mass-weighted mean over the first five frames of all 20 cases; the Rosin–Rammler number-mean is recovered under number weighting).
Longitudinal and transverse dispersion analysis
Taylor dispersion framework and turbulent Péclet number (extended derivation)
The particle-trajectory dataset provides a direct route to quantifying the dispersion tensor of the aerosol cloud through Taylor’s classical framework.[10, 14] [29, 30] The analysis is performed on the complete OpenFOAM Lagrangian dataset: for each case, all particles present at (5,300–7,220 parcels per case) are tracked individually across 28 subsequent output times using the unique parcel identifier origId. After removing the ensemble mean drift, the longitudinal (horizontal, ) and transverse (vertical, ) mean-squared displacements at lag are
where and are the - and -displacement components of parcel over lag relative to reference time, denotes averaging over parcels at fixed , and the subscript indicates the instantaneous ensemble size. The effective dispersion coefficients are estimated from the slope of the MSD in the intermediate-time window , where is the largest lag included in the MSD curve. Main text Figure 8 shows the resulting and versus the non-dimensional lag time for all twenty cases; the ballistic-to-diffusive transition at – is consistent with the Lagrangian integral time scale to .
Constancy of the turbulent Péclet number. The turbulent Péclet number, with (, ), is remarkably constant across all twenty cases: (coefficient of variation ). This invariance arises because the – SST boundary conditions are prescribed at constant turbulence intensity , so and yields and therefore . At , advective transport dominates turbulent diffusion by four orders of magnitude, confirming that the aerosol fate is controlled by the mean flow pattern, a prerequisite for the GNN’s flow-feature-based prediction approach.
Effective dispersion coefficients and ventilation scaling
Figure S3 presents the non-dimensional effective diffusion coefficients and . Table S7 summarises the mean values by ventilation level. The longitudinal diffusivity decreases as , with values overlapping the Li et al.31 reference range (– for indoor spaces with obstacles) at intermediate ventilation levels (–22).
| (m/s) | ACH (h) | (mm/s) | ||||
| 0.10 | 6.2 | 0.0584 | 0.0219 | 3.67 | 14.0 | 2,164 |
| 0.20 | 12.5 | 0.0269 | 0.0079 | 3.77 | 15.8 | 2,603 |
| 0.35 | 21.8 | 0.0104 | 0.0023 | 9.89 | 12.7 | 2,962 |
| 0.50 | 31.1 | 0.0059 | 0.0023 | 4.43 | 12.7 | 3,031 |
Table S7. Mean non-dimensional effective dispersion coefficients and anisotropy ratio by ventilation level.
Figure S3. Non-dimensional effective dispersion coefficients with panel labels (a)–(d): longitudinal vs. ACH and (left column); transverse vs. the same abscissae (right column; log-scaled axes on the bottom row). Green bands reproduce the Li et al.31 indoor reference range.
Dispersion anisotropy and vertical drift decomposition
Figure S4 examines the anisotropy of particle spreading and the physical origin of the vertical drift. The dispersion anisotropy ratio equals 3.7–9.9 at the three lower ventilation levels and falls to at the highest ACH = 31.1, indicating near-isotropic spreading. This counter-intuitive result arises because the supply jet at is sufficiently strong to establish a quasi-two-dimensional recirculation pattern that enhances vertical mixing. The net vertical drift velocity – is within 10% of the pure-Stokes settling velocity across all twenty cases, confirming that Stokes settling is the dominant vertical-transport mechanism and validating the Stokes-based gravity model in the graph surrogate to better than 10%.
Figure S4. Panel (a): dispersion anisotropy vs. ACH for the full Lagrangian dataset, coloured by injection speed . Panel (b): transverse MSD decomposition into turbulent diffusion and drift contributions for representative ventilation levels.
Computational benchmarks
A single foam-extend CFD case requires approximately 40 min of wall time to produce 261 saved frames at , dominated by the transient reactingParcelFoam stage. GNN inference runs on a single NVIDIA Quadro P1000 GPU (4 GB VRAM) and completes a full 26-second rollout in for the paired ELGIN checkpoint used in the headline comparison (M0 completes in on the same GPU owing to the absent Eulerian sub-network), giving an end-to-end speed-up of relative to the simpleFoam+reactingParcelFoam reference pipeline. The dominant cost of the ELGIN run relative to M0 is the Eulerian Graph Transformer pass on the 7704-cell mesh and the 50 inner Jacobi-PCG iterations of the pressure projection, both of which contribute a few tens of seconds in total over the 261 frames; the ELGIN wall-clock for each archived rollout is recorded alongside the rollout output.
The speed-up compares favourably with published GNN/ML-CFD surrogates: Sanchez-González et al. [21] reported speed-up for granular GNS relative to MPM/SPH; Kumar and Vantassel [22] reported for landslide GNS; Pfaff et al. [18] demonstrated for MeshGraphNets relative to FEM on cylinder wakes; Hanke et al. [26] reported for ocean GNS. GPU memory peaks at GB during a single-case ELGIN rollout (1000 tracked parcels, mean Lagrangian edges and Eulerian edges), comfortably within the 4 GB budget. Training wall time is approximately 12 h per published surrogate (M0 baseline or ELGIN) from random initialisation on the 16-case training set.
References
- [1]S. K. Harrel and J. Molinari, Journal of the American Dental Association 135, 429 (2004).
- [2]R. E. Micik, R. L. Miller, M. A. Mazzarella, and G. Ryge, Journal of Dental Research 48, 49 (1969).
- [3]Y. Li, H. Qian, J. Hang, X. Chen, P. Cheng, H. Ling, S. Wang, P. Liang, J. Li, S. Xiao, J. Wei, L. Liu, B. J. Cowling, and M. Kang, Building and Environment 196, 107788 (2021).
- [4]C. Zemouri, H. de Soet, W. Crielaard, and A. Laheij, PLoS ONE 12, e0178007 (2017).
- [5]L. Morawska and D. K. Milton, Clinical Infectious Diseases 71, 2311 (2020).
- [6]C.-C. Wang, K. A. Prather, J. Sznitman, J. L. Jimenez, S. S. Lakdawala, Z. Tufekci, and L. C. Marr, Science 373, eabd9149 (2021).
- [7]Ferziger, J. H., & Peric, M. (2002). Computational Methods for Fluid Dynamics: 3rd ed..
- [8]F. R. Menter, AIAA Journal 32, 1598 (1994).
- [9]E. Cunningham, Proceedings of the Royal Society of London A 83, 357 (1910).
- [10]S. B. Pope, Turbulent Flows (Cambridge University Press, Cambridge, 2000).DOI
- [11]A. Einstein, Annalen der Physik 322, 549 (1905).
- [12]J. K. Gupta, C.-H. Lin, and Q. Chen, Indoor Air 20, 31 (2010).
- [13]X. Guo, W. Li, and F. Iorio, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2016) pp. 481–490.
- [14]D. Kochkov, J. A. Smith, A. Alieva, Q. Wang, M. P. Brenner, and S. Hoyer, Proceedings of the National Academy of Sciences 118, e2101784118 (2021).
- [15]Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, arXiv preprint arXiv:2101.08895 (2021).
- [16]L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis, Nature Machine Intelligence 3, 218 (2021).
- [17]M. Raissi, P. Perdikaris, and G. E. Karniadakis, Journal of Computational Physics 378, 686 (2019).
- [18]J. Gilmer, K. T. Schütt, O. A. von Lilienfeld, A. Tkatchenko, and K.-R. Müller, in Proceedings of the 34th International Conference on Machine Learning (ICML), Vol. 70 (2017) pp. 1263–1272.
- [19]P. W. Battaglia, J. B. Hamrick, B. Bapst, et al., arXiv preprint arXiv:1806.01261 (2018).
- [21]T. F. Pfaff, M. A. Fortunato, A. Sanchez-Gonzalez, and P. W. Battaglia, in International Conference on Learning Representations (ICLR) (2021).
- [22]N. Hanke, T. Alber, and D. M. Kochmann, Applied Ocean Research 154, 104424 (2025).
- [23]A. Aminimajd, J. Maia, and A. Singh, Soft Matter 21, 2826 (2025).
- [24]M. Milnarević, G. K. Holt, and A. Agnello, arXiv preprint arXiv:2503.00274 (2025).
- [25]V. Sharma and O. Fink, arXiv preprint arXiv:2501.07373 (2025).arXiv










