⚡ electrical engineering

Geometry-conditioned diffusion generation of Lagrangian particle trajectories in human airways

This paper proposes a geometry-conditioned guided denoising diffusion probabilistic model that generates physically plausible, subject-specific Lagrangian particle trajectories in human airways by encoding airway anatomy via a variational autoencoder and enforcing geometric feasibility through signed distance function guidance, thereby overcoming the computational and preprocessing limitations of traditional simulations despite being trained on a limited dataset.

Original authors: Ignacio Rodrigo Bartol, Martin Graffigna, Mauricio Tano, Shaheen Dewji

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ignacio Rodrigo Bartol, Martin Graffigna, Mauricio Tano, Shaheen Dewji

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict how a swarm of tiny, invisible drones (like microscopic dust or medicine droplets) will fly through the twisting, turning tunnels of a human lung. This isn't just a game of chance; it's a life-or-death calculation for doctors trying to deliver asthma inhalers to the right spot or for scientists figuring out how much radiation a worker might absorb from a radioactive cloud. To do this accurately, scientists usually run super-complex computer simulations that act like a high-speed video camera, tracking every single particle as it bumps into walls, gets pushed by wind, and settles down. But these simulations are like trying to solve a million-piece puzzle while running a marathon: they take forever, require massive supercomputers, and need the lung's shape to be perfect down to the last pixel. If the digital lung has even a tiny glitch, the whole simulation crashes or gives nonsense results.

The big question is: Can we teach a computer to "dream" these flight paths instead of calculating them step-by-step? This paper dives into that idea using a type of artificial intelligence called a "diffusion model." Think of this model like an artist who starts with a canvas covered in static noise (like TV snow) and slowly, step-by-step, removes the noise to reveal a clear picture. In this case, the "picture" isn't a face or a landscape, but a 3D path that a particle takes through a lung. The researchers wanted to see if they could train this AI to look at a specific person's lung shape and instantly generate thousands of realistic particle paths, skipping the hours of heavy calculation. They also had to make sure the AI didn't get confused and draw particles flying through the lung walls, which would be physically impossible.

The Paper's Story: Teaching AI to Dream Lung Paths

The authors, a team of engineers and scientists, built a new kind of AI system to solve this "dreaming" problem. They call it a "geometry-conditioned guided diffusion model." That's a mouthful, so let's break it down with a simple analogy.

Imagine you are trying to teach a robot to draw a path for a marble rolling through a specific, unique maze.

The "Geometry Conditioning" (The Map): First, the robot needs to know what the maze looks like. The researchers fed the AI a digital 3D map of the airways (the maze). But instead of just showing the robot a picture, they compressed the entire 3D shape into a tiny, secret "code" (a latent vector) using a special tool called a Variational Autoencoder. This code acts like a fingerprint for that specific person's lung. When the AI starts its work, it looks at this fingerprint to understand the rules of the maze it's about to draw.
The "Diffusion" (The Dreaming): The AI starts with a chaotic mess of random dots in 3D space—like a cloud of dust. It then slowly cleans up this mess, step by step, turning the random noise into a smooth, flowing line. It's like watching a sculptor chip away at a block of marble to reveal a statue, but in reverse: the AI chips away the "noise" to reveal the "trajectory."
The "Guidance" (The Bouncer): Here is the tricky part. If the AI just dreams, it might accidentally draw a path that goes straight through a solid wall of the lung. To stop this, the researchers added a "bouncer" called a Signed Distance Function (SDF). Think of the SDF as an invisible force field. If the AI tries to draw a particle outside the lung (where the air isn't), the bouncer gives it a gentle but firm push back inside. This ensures the paths stay physically possible without needing to run the slow, heavy physics simulations again.

What They Found

The team tested this system using data from nine different simulated human airway geometries (specifically, nine male anatomies). They trained the AI on particle tracks generated by the slow, traditional super-computer simulations and then asked the AI to generate new paths on its own.

The results were promising, but with some important caveats:

It Works (Mostly): The AI successfully generated thousands of particle paths in seconds that looked very similar to the slow, expensive simulations. The paths followed the main branches of the lungs correctly, and the speed of the particles matched the real physics quite well.
The "Bouncer" Helps: Without the SDF guidance, the AI's "dreams" often drifted outside the lung walls. With the guidance, the paths stayed inside the airways much better, though it wasn't a perfect 100% guarantee.
The Limits of the Dream: The AI is only as good as the examples it was trained on. Since it only saw nine different lung shapes, it struggled a bit when shown a lung that was very different from those nine. In some cases, it smoothed out the paths too much, missing the wild, fast, or erratic movements that happen in the tiny, deep branches of the lung. It also sometimes missed the "extreme" events, like the very fastest particles, because it was trying to find the "average" safe path.
Compared to Old Tricks: The researchers also compared their new AI to an older, simpler method called Dynamic Mode Decomposition (DMD). The new diffusion model was much better at capturing the complex, wiggly details of how particles move. The older method was too rigid and missed the chaotic nature of the airflow, often creating weird, broken paths that didn't make sense.

Why This Matters (and What's Next)

This paper suggests that we might soon be able to replace hours of super-computer crunching with a few seconds of AI "dreaming" to predict how medicine or pollutants move through a patient's lungs. This could be huge for personalized medicine, where a doctor could instantly see how a specific inhaler would work for a specific person's unique lung shape.

However, the authors are careful not to call this a finished product. They admit that because they only trained on nine lung shapes, the AI isn't ready for every human yet. It needs to see more examples to learn the full variety of human anatomy. They also note that the current system doesn't yet account for different particle sizes or breathing patterns, which are crucial for real-world use. But the core idea—that we can teach an AI to generate realistic, physics-compliant particle paths based on a lung's shape—is a solid, exciting step forward. It's not a magic wand yet, but it's a very powerful new tool in the toolbox.

Technical Summary: Geometry-Conditioned Diffusion Generation of Lagrangian Particle Trajectories in Human Airways

Problem Statement
Accurate prediction of particle transport in human airways is critical for assessing aerosol deposition, optimizing inhaled drug delivery, and conducting radiation dose research. While high-fidelity Computational Fluid and Particle Dynamics (CFPD) simulations provide detailed Lagrangian trajectory data, their application in clinical or large-cohort settings is hindered by stringent geometric preprocessing requirements (segmentation, mesh repair, smoothing) and prohibitive computational costs. Existing data-driven surrogates primarily focus on Eulerian flow fields or simplified 2D settings and struggle to generalize to complex, subject-specific 3D anatomies. Furthermore, purely deterministic regressors fail to capture the intrinsic multimodality and stochastic nature of particle trajectories, which are sensitive to local geometry, turbulence, and stochastic forces.

Methodology
The authors propose a geometry-conditioned guided denoising diffusion probabilistic model (DDPM) to generate physically plausible particle trajectories directly in a Lagrangian representation. The framework integrates three core components:

Variational Geometry Encoding: Airway anatomy is represented as a 3D voxelized occupancy grid. A Variational Autoencoder (VAE) encodes this high-dimensional geometry into a compact, stochastic latent vector ( $z_G$ ). This latent vector serves as a global conditioning input for the diffusion model, allowing the network to learn a mapping between anatomical features and trajectory statistics.
Conditional Diffusion Model: The core generator is a 1D temporal U-Net trained to predict noise added to a sequence of 3D particle positions ( $x \in \mathbb{R}^{K \times 3}$ , where $K=1024$ time steps). The model is conditioned on the encoded geometry latent and the diffusion timestep. The training objective minimizes the noise prediction error ( $L_{ddpm}$ ) alongside a Kullback-Leibler (KL) regularization term to prevent the encoder from overfitting to discrete training geometries.
Signed Distance Function (SDF) Guidance: To enforce geometric feasibility during inference, a Signed Distance Function (SDF) is constructed from the airway geometry. During the reverse diffusion process, a soft geometric penalty is applied. The SDF acts as a repulsive energy field where values outside the airway lumen are positive. The gradient of this energy is used to shift the mean of the reverse transition, biasing the generated samples away from the airway walls and toward the lumen. This guidance is applied at inference time without retraining the base model.

Key Contributions
The paper outlines four primary contributions:

A Diffusion-Based Generative Model: The first application of a conditional diffusion model to generate ensembles of Lagrangian particle trajectories conditioned on 3D patient-specific airway geometry.
Variational Encoding Strategy: A method to condition the generative model on anatomies not seen during training by compressing voxelized geometries into a latent space.
Inference-Time Guidance: A mechanism using SDF gradients to improve geometric plausibility and reduce non-physical wall-penetrating outliers without enforcing hard constraints that might disrupt the generative process.
Comprehensive Evaluation: An analysis demonstrating the model's ability to recover velocity magnitude statistics and produce visually consistent trajectory bundles, while explicitly acknowledging limitations in generalization to out-of-distribution geometries due to the limited training set.

Results
The model was trained and evaluated on particle track data derived from CFPD simulations of nine representative male airway geometries (selected from a larger cohort of 412).

Qualitative Performance: SDF-guided samples showed significantly fewer out-of-lumen segments compared to unguided samples. The generated trajectories aligned well with major airway pathways and bifurcation patterns in both training and held-out geometries. However, discrepancies persisted in distal branches, where generated tracks sometimes appeared smoother or less spatially dispersed than ground truth.
Quantitative Performance: The model recovered key kinematic statistics with reasonable accuracy. For the evaluation set, the mean speed differed by approximately -4.3% and the 99th percentile speed by -12.0% compared to reference data. Wasserstein-1 distances for pooled speed distributions were low (ranging from $10^{-3}$ to $10^{-2}$ m/s), indicating good distributional recovery.
Comparison with DMD: In an exploratory comparison with a Dynamic Mode Decomposition (DMD) reduced-order model on a single geometry, the guided DDPM outperformed DMD in reproducing speed distributions, component velocities, multi-lag increment statistics, and power spectral densities. The DMD reconstruction exhibited non-physical discontinuities and failed to capture the full conditional distribution of plausible tracks, whereas the DDPM provided a stochastic generator capable of sampling diverse realizations.

Significance and Claims
The paper positions this work as a proof-of-concept demonstrating that diffusion models can serve as fast, geometry-aware surrogates for Lagrangian particle tracking. The authors claim that the method bridges the gap between high-fidelity simulation and practical deployment by generating plausible trajectory ensembles in seconds to minutes, retaining key statistical-physical features of the underlying dynamics.

The authors are modest regarding the current scope, explicitly stating that:

Generalization to unseen anatomies is fundamentally limited by the diversity of the available simulated geometries (only nine used for training).
The model does not yet output carrier-flow variables or explicit deposition/escape labels; it focuses on position sequences.
Boundary violations still occur in challenging cases (e.g., thin distal branches), and strong guidance can reduce trajectory diversity in peripheral regions.
Future work requires expanding the dataset to include more diverse geometries and incorporating explicit conditioning variables such as particle size and breathing waveforms.

The study concludes that while the current framework is not a complete replacement for CFPD, it establishes a viable pathway toward rapid, subject-specific trajectory generation, provided that the limitations regarding dataset diversity and boundary constraints are addressed in future iterations.

More like this