Learning Interface Breakup: A Geometry-Conditioned Latent Surrogate for Spray Formation

This paper introduces a geometry-conditioned latent surrogate model that achieves a $6\times10^4$ speed-up over high-fidelity CFD simulations by encoding adaptive mesh refinement cell-density fields as a compact proxy to accurately reconstruct transient two-phase spray breakup dynamics for efficient nozzle design.

The Gist

Imagine you are an engineer trying to design the perfect nozzle for a spray bottle, a fuel injector, or a perfume sprayer. Your goal is to figure out how the shape of the nozzle affects the way liquid breaks apart into tiny droplets.

In the real world, to test a new shape, you have to run a super-complex computer simulation. Think of this simulation like a high-definition movie camera filming a drop of water shattering in slow motion. But here's the catch: to get a clear picture, the computer has to zoom in incredibly close on the edges where the water meets the air. This "zooming" (called Adaptive Mesh Refinement) makes the simulation incredibly accurate but also incredibly slow. Running one simulation can take hours of computer time. If you want to test 1,000 different nozzle shapes, you'd be waiting for months.

The Problem
The authors of this paper wanted to speed this up. They tried using "surrogate models" (AI shortcuts) to predict the results instantly. However, standard AI models struggle here because the "camera" (the computer grid) keeps changing its zoom level and shape as the water moves. It's like trying to teach a student to recognize a face when the photo keeps changing resolution and angle every second.

The Solution: The "Zoom Map" Trick
The team came up with a clever new way to teach the AI. Instead of feeding the AI the entire, messy movie of the water and air, they decided to feed it a "Zoom Map."

• The Analogy: Imagine the computer simulation is a city. The "Zoom Map" doesn't show every building or street; it just shows a heat map of where the computer is looking closely. It tells the AI: "Hey, the computer is zooming in heavily right here on the water's edge, and less here in the empty air."
• The Innovation: They realized this "Zoom Map" (which they call the AMR cell-density field) is actually a perfect, compact summary of where the action is happening. By training the AI on this map, they could teach it the physics of the spray without getting bogged down by the massive amount of data.

How the AI Works (The Two-Stage Process)
The system they built works like a two-step magic trick:

1. Stage 1: The "Memory" (The Latent Surrogate)
   The AI looks at the "Zoom Map" and compresses the entire story of the spray into a tiny, 48-number "memory code."

   • From this single code, the AI can reconstruct the nozzle shape (the walls of the pipe).
   • It can also reconstruct the "Zoom Map" at any point in time, showing where the water is breaking apart.
   • Key feat: Even if you only give the AI the shape of the nozzle (without the water data), it can mathematically "guess" the right memory code to recreate the spray.

2. Stage 2: The "Fill-in"
   Once Stage 1 has the "Zoom Map" and the shape, a second, smaller AI (a U-Net) fills in the rest of the details: the speed of the water and the exact amount of liquid in every spot.

The Results
The results were dramatic:

• Speed: The old way took hours (specifically, 0.75 to 1.5 hours on a powerful computer). The new AI method takes 0.045 seconds. That is a speed-up of over 60,000 times.
• Accuracy: Despite being so fast, the AI correctly predicted how the water breaks apart, capturing the complex swirls and droplets with high fidelity.
• Generalization: The AI didn't just memorize the training examples. When shown a brand-new nozzle shape it had never seen before, it could still predict the spray pattern accurately, proving it learned the underlying physics, not just the shapes.

Why This Matters
This paper proves that you don't need to feed an AI every single pixel of a complex simulation. Instead, you can feed it the "skeleton" of the simulation (the zoom map), and the AI can learn to fill in the rest. This turns a process that used to take days of computer time into a split-second calculation, opening the door for engineers to rapidly design better sprays, fuels, and medical delivery systems by testing thousands of ideas instantly.

Limitations Mentioned
The authors are careful to note that this was tested on a specific type of nozzle and a specific range of conditions. While the method is powerful, it currently relies on the specific "rules" of the computer solver they used. It's a huge step forward, but it's not a universal magic wand for every fluid problem in the universe yet.

Technical Summary

Technical Summary: Learning Interface Breakup

Problem Statement

Designing spray nozzles requires predicting how nozzle geometry influences transient two-phase flow breakup. While high-fidelity Volume-of-Fluid (VOF) simulations with Adaptive Mesh Refinement (AMR) can accurately model these phenomena, they are computationally prohibitive for iterative design exploration. A single forward solution can take hours on multi-core CPUs due to the need for small timesteps to resolve thin, rapidly evolving interfaces and the stiff coupling between Navier-Stokes equations, VOF advection, and surface tension.

Standard neural surrogate models face unique challenges in this setting:

1. Dynamic Discretization: The underlying AMR mesh evolves over time and changes with geometry, making fixed-grid representations insufficient.
2. Complex Dynamics: Multiphase flows involve discontinuities, topology changes, and highly nonlinear interface dynamics.
3. Geometry-Flow Coupling: Existing approaches typically either condition on geometry in simplified, steady-state settings or learn latent representations of complex dynamics without explicit geometric conditioning. There is a lack of joint representations that simultaneously capture nozzle geometry and transient multiphase evolution.

Methodology

The authors propose a geometry-conditioned latent surrogate trained on 797 two-phase nozzle simulations. The core innovation is encoding the AMR cell-density field (a proxy for where the solver concentrates resolution) rather than the full multi-channel flow state. This transforms the adaptive discretization from a computational by-product into a learnable signal.

The architecture consists of two stages:

Stage 1: Shared Latent Representation (VAE)

This stage learns a compact, time-independent latent vector $z$ that couples nozzle geometry and transient flow evolution.

• Encoder: A Fourier Neural Operator (FNO) maps a single density snapshot ($\kappa$) to latent parameters $(\mu, \log \sigma^2)$. The encoder is time-blind.
• Decoder (Flow): A time-conditioned SIREN (Sinusoidal Representation Network) decoder reconstructs the transient density evolution $\hat{\kappa}(x, y; t)$ from the latent code $z$, time $t$, and spatial coordinates. It uses FiLM conditioning with multi-scale spatial Fourier features and sinusoidal time features.
• Geometry Head: A 1D Implicit Neural Representation (INR) decodes the nozzle wall boundaries ($y_{upper}(x), y_{lower}(x)$) from the same latent code $\mu$.
• Training: The model is trained end-to-end on density and nozzle-wall supervision alone, using a loss function that includes reconstruction error, geometry error, and a free-bits KL divergence penalty to prevent posterior collapse.

Stage 2: Field Predictor

A lightweight U-Net is trained separately to map the reconstructed density proxy $\hat{\kappa}$ to the remaining flow variables: volume fraction ($f$) and velocities ($u, v$). This stage operates on the assumption that density acts as a sparse spatial signal conditioning the globally defined velocity and volume fraction fields.

Inference Strategies

The framework supports two inference paths:

1. Snapshot-based: Given a density snapshot, the encoder produces $z$, which is then swept through time to reconstruct the full trajectory.
2. Geometry-only: Given a target nozzle geometry $g^*$ without a flow snapshot, the latent $z^*$ is recovered via constrained optimization in the latent space to minimize the distance between the decoded geometry and $g^*$.

Key Contributions

1. AMR Cell-Density as a Representation: The paper introduces the AMR cell-density field as a solver-informed representation for learning transient two-phase flows, leveraging the adaptive mesh structure as a compact proxy for flow complexity.
2. Shared Latent Coupling: A shared latent representation is proposed that couples nozzle geometry and transient flow evolution. This enables geometric reconstruction and time-dependent flow decoding from a single compact code. Crucially, it allows for the recovery of a usable latent from geometry alone via constrained optimization.
3. Efficient One-Shot Inference: The method demonstrates fast, one-shot surrogate inference (no autoregressive rollout), reducing trajectory prediction time from CPU-hours to milliseconds while preserving key interface dynamics.

Results

The model was evaluated on 81 held-out test simulations (approximately 3,155 snapshots) with a strict simulation-level split.

• Reconstruction Accuracy:
  • Density: Achieved a test Mean Squared Error (MSE) of $2.74 \times 10^{-3}$ (PSNR 30.2 dB).
  • Geometry: Decoded nozzle walls with a Mean Absolute Error (MAE) of $1.13 \times 10^{-2}$ in physical coordinates.
  • Flow Fields: The Stage-2 U-Net achieved test MSEs of $1.33 \times 10^{-3}$ for volume fraction, $2.50 \times 10^{-3}$ for streamwise velocity ($u$), and $1.60 \times 10^{-3}$ for cross-stream velocity ($v$).
• Speed-up: Inference time is approximately 0.045 seconds per trajectory on an Apple M4 GPU, representing a speed-up of more than $6 \times 10^4$ relative to the reference Basilisk CFD solver (which takes 45–90 minutes on 12 CPU cores).
• Generalization vs. Nearest Neighbor: The surrogate outperformed a nearest-neighbor baseline (based on geometric distance) by approximately 2x in mean MSE across all channels, indicating it learned non-trivial geometry-to-flow mappings rather than simple shape retrieval.
• Geometry-Only Inversion: When recovering the latent solely from geometry (without a flow snapshot), the method closed 55% of the performance gap compared to an encoder oracle that had access to the ground-truth density snapshot.

Significance and Claims

The paper claims that AMR refinement structure can serve as a compact and learnable representation for geometry-conditioned surrogate modeling of transient two-phase flows. By encoding the solver's own resolution strategy, the model bypasses the need to learn the full high-dimensional state directly.

The authors position this work as a foundational step toward geometry-driven design workflows in multiphase systems. They argue that such shared latent representations are a natural building block for inverse design, where geometry and dynamics are intrinsically coupled.

Limitations and Scope:
The authors are modest regarding the scope of their claims:

• The training data is limited to 797 simulations within a specific three-parameter NURBS nozzle family and a single Reynolds–Weber regime.
• The Stage-2 predictor is trained independently, meaning errors in density reconstruction propagate to velocity and volume fraction predictions.
• The representation is tied to the specific solver (Basilisk) and refinement strategy used; transferability to other solvers is not yet established.
• Validation against independent CFD reruns or experimental measurements is noted as necessary for operational use but lies beyond the current scope.

Future work is identified as extending to higher-dimensional design spaces, adding heads for droplet distribution observables, and validating optimized designs through independent high-fidelity simulations to close the loop between learned design and physical reality.