Physics-Informed Graph Neural Network Surrogates for Turbulent Nanoparticle Dispersion in Dental Clinical Environments

This paper introduces ELGIN, a physics-informed graph neural network surrogate that significantly accelerates and improves the accuracy of predicting turbulent nanoparticle dispersion in dental clinics compared to traditional CFD simulations, enabling near real-time infection-risk screening.

The Gist

The Problem: Invisible Clouds in the Dentist's Chair

Imagine you are sitting in a dentist's chair. When the dentist uses a high-speed drill or an ultrasonic cleaner, it creates a tiny, invisible mist of water droplets and saliva. These droplets are so small (some are smaller than a grain of sand) that they can float in the air for a long time, like dust motes dancing in a sunbeam.

If a patient has a virus, these floating droplets can carry it to the dentist, the hygienist, or anyone else in the room. To understand how these droplets move, scientists usually use powerful computer simulations (called CFD). Think of these simulations as a super-slow-motion movie that calculates the physics of every single air molecule and water drop.

The Catch: Making this "movie" takes a long time. Running one simulation for a single dental appointment scenario takes about 40 minutes on a fast computer. This is too slow to be useful in real life. If a dentist wants to know, "Is the air safe right now if I change the fan speed?" they can't wait 40 minutes for an answer. They need an answer in seconds.

The Solution: ELGIN (The "Smart Apprentice")

The authors created a new tool called ELGIN. Instead of calculating every physics equation from scratch every time (like the slow simulation), ELGIN is a smart apprentice that has studied thousands of hours of those slow movies.

ELGIN is a type of Artificial Intelligence called a Graph Neural Network.

• The Analogy: Imagine the dental room is a giant city. The slow simulation calculates the traffic flow for every single car and pedestrian individually. ELGIN, however, is like a traffic control system that looks at the whole city map (the "graph") and predicts where traffic will go based on patterns it learned previously.

How ELGIN Works (The Hybrid Approach)

The paper highlights that ELGIN is special because it uses a hybrid approach, combining two different ways of thinking:

1. The Air (The River): ELGIN predicts how the air moves (the "carrier flow"). It looks at the room's layout—the dentist, the patient, the walls, and the air vents—and predicts the wind currents.
2. The Droplets (The Leaves): ELGIN also tracks the floating droplets. It knows that some droplets are heavy and fall quickly, while others are light and float like leaves on a stream.

The Innovation: Previous AI models tried to guess the droplets' path just by looking at other droplets nearby. This is like trying to predict where a leaf will go by only looking at the leaves next to it, without knowing where the river is flowing. ELGIN fixes this by always checking the "river" (the air flow) to see where the wind is pushing the droplets. It also pays attention to the "walls" (obstacles like the dentist's head) to know where the air swirls around them.

The Training: Learning by Doing

To teach ELGIN, the authors didn't just show it pictures; they used a four-stage training curriculum, which is like a rigorous boot camp:

1. Stage 1: It learned to predict the wind patterns in the room.
2. Stage 2: It learned to predict how a single droplet moves in one second.
3. Stage 3: It learned to combine both, ensuring the wind and droplets obeyed the laws of physics (like conservation of energy).
4. Stage 4: It practiced predicting the entire 26-second movie of a dental procedure, learning to correct its own mistakes as it went along.

The Results: Fast and Accurate

The authors tested ELGIN on a specific dental room scenario and compared it to:

• The Slow Simulation (The Gold Standard): Takes 40 minutes.
• The Old AI Model (M0): A simpler AI that didn't look at the air flow.
• ELGIN (The New Model): The hybrid AI.

The Performance:

• Speed: ELGIN predicted the 26-second movie in about 64 seconds. That is roughly 37 times faster than the slow simulation.
• Accuracy: The old AI model (M0) made mistakes about where the droplets went, with an average error of nearly 20% of the room's width. ELGIN reduced this error to about 16%.
• Shape: The old AI model also got the "shape" of the cloud wrong (it spread out too much or too little). ELGIN got the shape of the cloud much closer to reality.

What This Means (According to the Paper)

The paper states that this is a proof-of-concept. They successfully showed that:

1. It is possible to train an AI to predict how dental aerosols move in a room.
2. By combining air-flow prediction with droplet tracking, the AI is much more accurate than models that only look at the droplets.
3. The system is fast enough to potentially be used for real-time infection risk screening in the future (e.g., telling a dentist if a specific ventilation setting is safe before they start a procedure).

Important Note from the Paper:
The authors are careful to say this is a single-case demonstration. They trained and tested it on one specific room setup. They are currently working on training it on 20 different scenarios to prove it works in all kinds of dental rooms, not just this one. They also note that before this can be used in real clinics, it needs to be tested against real-world measurements (not just computer simulations) and expanded to 3D rooms.

Summary Analogy

Think of the slow computer simulation as a master painter who takes 40 minutes to paint a perfect, detailed landscape.
The old AI was a student who tried to guess the landscape by looking at a blurry photo of the previous day's painting.
ELGIN is a smart apprentice who has studied the master's techniques, understands how wind and light work, and can paint a very good approximation of the landscape in just over a minute. It's not perfect yet, but it's fast enough to be useful.

Technical Summary

Technical Summary: Physics-Informed Graph Neural Network Surrogates for Turbulent Nanoparticle Dispersion in Dental Clinical Environments

Problem Statement
Dental procedures, such as high-speed drilling and air-polishing, generate polydisperse bioaerosols (0.5–50 µm) that remain airborne for extended periods in enclosed clinics, posing significant infection risks. While Reynolds-Averaged Navier–Stokes (RANS) simulations with Euler–Lagrange particle tracking accurately model this transport, they are computationally prohibitive for real-time clinical decision support, requiring approximately 40 minutes per scenario on a single core. Existing machine learning surrogates for Computational Fluid Dynamics (CFD) often rely on fixed-resolution structured grids or lack the ability to generalize to the unstructured, topology-varying meshes required for complex clinical geometries. Furthermore, purely data-driven Graph Neural Network (GNN) approaches often fail to conserve energy or momentum over long rollouts and struggle to capture the specific physics of turbulent bioaerosol dispersion, including the coupling between carrier flow and polydisperse particles.

Methodology
The authors propose ELGIN (Eulerian–Lagrangian Graph Interaction Network), a physics-informed GNN surrogate designed to jointly predict carrier-flow dynamics and the motion of a polydisperse spray cloud on an OpenFOAM polyhedral mesh.

• Dual-Graph Architecture: ELGIN operates on two coupled graphs:
  1. Eulerian Graph ($G_E$): Represents the static OpenFOAM polyMesh (7,704 cells). It utilizes a multi-head Graph Transformer processor with a learnable turbulence-closure head and a Jacobi-preconditioned learnable pressure projection to solve for the divergence-corrected RANS velocity field ($U, p, k, \omega$).
  2. Lagrangian Graph ($G_L$): Represents the time-varying particle cloud (1,000 tracked parcels). It employs an Interaction Network with sigmoid-gated attention, an LSTM velocity-history encoder, and a symplectic Störmer–Verlet integrator for position updates.
• Coupling Mechanism: The two graphs are coupled via a differentiable, inverse-distance-weighted (IDW) interpolation operator. This allows Lagrangian parcels to access local RANS velocity, turbulent kinetic energy (TKE), distance-to-wall, and wall-normal vectors interpolated from the Eulerian mesh.
• Physics Integration: The model explicitly incorporates physical laws rather than learning them from scratch. It includes analytic Cunningham-corrected Stokes drag, Saffman shear-lift, and Brownian diffusion terms as input features. The pressure projection enforces discrete mass conservation on the graph stencil.
• Training Curriculum: A four-stage curriculum stabilizes long-horizon autoregressive rollouts:
  1. Eulerian fluid pre-training.
  2. One-step particle supervision with input noise.
  3. PDE-informed joint training (minimizing continuity, momentum, and turbulence residuals).
  4. Back-propagation through time (BPTT) rollout fine-tuning.
• Conditioning: The model is conditioned on per-case air-inlet velocity vectors and semantic boundary embeddings (dentist, patient, walls, etc.), enabling potential generalization across ventilation rates and spray speeds.

Key Contributions

1. Hybrid Surrogate Framework: The first GNN surrogate for turbulent bioaerosol spatial transport in enclosed clinical spaces conditioned on a pre-computed physical carrier-phase flow field, jointly predicting the RANS velocity field and particle trajectories.
2. Robust Training Protocol: A four-stage curriculum that successfully stabilizes 260-step autoregressive rollouts for a dissipative multi-physics problem without gradient explosion.
3. Data Extraction and Handling: A full-timeline protocol with persistent origId tracking and an "alive-mask" objective to correctly handle variable parcel counts due to injection and deposition.
4. Explicit Boundary Encoding: Name-aware encoding of nine semantic boundary classes and per-case inlet velocity conditioning to facilitate generalization across the parameter sweep.

Results
The paper reports a single-case demonstration (Sweep_Case_03) comparing ELGIN against a Lagrangian-only GNS baseline (M0) trained on the same dataset.

• Accuracy: ELGIN significantly outperforms the M0 baseline. The mean parcel displacement error (MDE) decreases from 19.56% to 16.20% of the room width, and the cloud radius-of-gyration error drops from 9.85% to 6.58%.
• Speed: A 26-second rollout completes in approximately 64 seconds on a 4 GB GPU, representing a ~37× speedup compared to the reference foam-extend 4.1 CFD pipeline (which takes ~40 minutes).
• Clinical Metric: ELGIN tracks the Breathing Zone Exposure (BZE) fraction more faithfully than M0, though the authors note the BZE metric on the 1,000-parcel subset is subject to Poisson counting noise.
• Physical Fidelity: The model correctly captures the advection-dominated transport (high turbulent Péclet number) and the tracer-like behavior of the droplets (low Stokes number), as evidenced by the non-dimensional analysis of the underlying CFD data.

Significance and Limitations
The authors position this work as a proof-of-concept demonstrating that physics-informed hybrid Eulerian–Lagrangian GNNs can reproduce multiscale clinical aerosol dynamics at speeds compatible with per-appointment risk screening. The architecture and training protocol are presented as a foundation for future multi-case generalization and 3D extensions.

The paper explicitly acknowledges several limitations:

• Single-Case Validation: The reported metrics are from a single case (Sweep_Case_03); full generalization across the 20-case parameter sweep (16/2/2 train/validation/test split) is in progress.
• 2-D Geometry: The study is restricted to a 2-D cross-section; extension to full 3-D geometry is identified as a critical next step.
• Steady-State Carrier Flow: The carrier flow is based on steady-state RANS, excluding ventilation transients and occupant-induced unsteadiness.
• Experimental Validation: The surrogate has been validated against CFD references only; clinical deployment requires validation against optical particle counter or phase-Doppler anemometry measurements.

The authors conclude that while the current results are modest and specific to a single realization, the framework offers a viable path toward real-time infection-risk screening in healthcare settings once the multi-case retraining and 3-D extensions are completed.