🤖 machine learning

Diffusion-corrected Autoregressive Fourier Neural Operator for Droplet Evolution Prediction

This paper introduces DiffARFNO, a two-stage framework combining an autoregressive Fourier-MIONet for coarse long-horizon forecasting with a conditional DDIM corrector for fine-detail refinement, which significantly improves the accuracy of droplet evolution predictions in inkjet printing compared to existing state-of-the-art models.

Original authors: Jinghao Cao, Minsung Kang, Hongyue Sun, Chi Zhou, Jihoon Chung, Xubo Yue, Sanchoy Das, Bo Shen

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jinghao Cao, Minsung Kang, Hongyue Sun, Chi Zhou, Jihoon Chung, Xubo Yue, Sanchoy Das, Bo Shen

Original paper licensed under CC BY 4.0 (http://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 trying to predict the future of a tiny, dancing water droplet. It's not just a simple splash; it's a chaotic, high-speed ballet where the droplet stretches into a thin thread, snaps apart, and sometimes sprouts tiny "satellite" droplets like a cosmic tree branching out. This happens in the world of inkjet printing, a technology used to build everything from flexible electronics to biological tissues. But here's the catch: these droplets move faster than the blink of an eye and are smaller than a human hair. To understand them, scientists usually have two choices. They can take super-fast photos, which is expensive and hard to do, or they can run massive computer simulations. These simulations are like digital wind tunnels that solve complex physics equations to see how the droplet behaves. The problem? Running these simulations is incredibly slow. It can take hours to simulate a split-second of a droplet's life, making it impossible to test thousands of different scenarios quickly.

Enter the world of artificial intelligence, where researchers are teaching computers to "guess" the physics instead of calculating every single step. Think of it like a student who has studied a textbook so well they can predict the next page of a story without reading it. However, there's a flaw in this approach: when you ask a computer to predict a long story (or a long sequence of droplet frames) one step at a time, it tends to get "blurry." Just like a game of "telephone" where the message gets distorted as it passes from person to person, the computer's predictions start to lose their sharp details. The droplet might look like a fuzzy blob instead of a crisp, sharp drop. This paper tackles that specific problem: how do we get a computer to predict the entire future of a droplet quickly, without it turning into a fuzzy mess?

The researchers behind this study, led by Jinghao Cao and colleagues, have built a new AI system called DiffARFNO. You can think of this system as a two-person team working together to predict the droplet's future. The first member is a "fast forecaster" (based on a model called Fourier-MIONet). This forecaster is incredibly quick and good at seeing the big picture. It can look at the first 50 frames of a droplet's life and guess what the next 550 frames will look like. But, as we mentioned, it's a bit sloppy with the fine details; its predictions are like a rough sketch.

The second member of the team is a "detail artist" (based on a diffusion model). This artist doesn't start from scratch. Instead, they take the rough sketch from the fast forecaster and use a special technique called "iterative denoising" to clean it up. Imagine the fast forecaster draws a blurry outline of a bird, and the detail artist comes in with a fine-tipped pen to sharpen the feathers and define the beak. In the paper's experiments, this two-stage process was tested on a dataset of 212 simulated droplet videos generated by ANSYS Fluent software. The results showed that this team approach was far superior to using just the fast forecaster or other existing AI models.

When the researchers measured how accurate the predictions were, DiffARFNO came out on top. It achieved a Mean Squared Error (MSE) of 0.0003 and a Peak Signal-to-Noise Ratio (PSNR) of 34.7104, which are numbers indicating very high precision and clarity. More importantly, it could generate a full 600-frame prediction of a droplet's evolution in just about 40 seconds. In contrast, the traditional computer simulation method that the AI is trying to replace takes about 4 hours to produce the same result. The paper explicitly rules out the idea that a single model can do this perfectly on its own; they argue that trying to use just the diffusion model for the whole long sequence would be too slow, and using just the fast forecaster leads to blurry, inaccurate results. By combining the speed of the forecaster with the precision of the corrector, they suggest this hybrid method offers a reliable way to predict complex droplet behaviors like "ligament thinning" (when the drop stretches into a string), "pinch-off" (when it snaps), and "satellite droplet formation" (when tiny drops break off).

The study confirms that this method works well across different types of fluid behaviors, maintaining sharp edges and accurate shapes where other models failed. While the paper focuses on simulated data and doesn't claim to have solved every problem in inkjet printing, it demonstrates a significant step forward. It suggests that by letting a fast AI do the heavy lifting and a smart AI do the touch-ups, we can finally predict the chaotic dance of ink droplets with both speed and high fidelity, potentially helping engineers design better printers and materials without waiting hours for a computer to catch up.

Technical Summary: Diffusion-corrected Autoregressive Fourier Neural Operator for Droplet Evolution Prediction

Problem Statement
Predicting the evolution of inkjet printing (IJP) droplets is critical for maintaining printing quality in additive manufacturing. However, long-horizon forecasting of these dynamics remains challenging due to the complex coupling of process variables (viscosity, density, surface tension) and the highly transient, nonlinear nature of free-surface fluid dynamics. Existing data-driven approaches often struggle with error accumulation and the degradation of high-frequency structures (e.g., ligament thinning, pinch-off, satellite droplet formation) during autoregressive rollouts. Furthermore, while Computational Fluid Dynamics (CFD) offers high-fidelity simulations, it is computationally prohibitive for rapid exploration of parameter spaces or real-time control.

Methodology: DiffARFNO
The authors propose DiffARFNO (Diffusion-corrected Auto-Regressive Fourier Neural Operator), a two-stage framework designed to balance computational efficiency with high-fidelity long-horizon prediction.

Stage 1: Coarse Autoregressive Prediction (Fourier-MIONet)
- A Fourier-MIONet backbone serves as the primary predictor. It is a multiple-input neural operator that processes three inputs: an image sequence ( $X_{in}$ ), a static property vector ( $p$ ) containing material properties, and a normalized coordinate grid ( $c$ ).
- The model employs a branch-trunk architecture where video, property, and positional features are encoded and fused via a shared Fourier-based decoder (UFNO).
- Autoregressive Rollout: The model predicts future frames in fixed-length segments. To mitigate boundary mismatches and "droplet shaking" between consecutive segments, a linear blending strategy is applied to the overlapping frames of adjacent windows.
- Training Loss: The stage utilizes a combined loss function comprising an RGB pixel-wise loss and a binary mask loss (derived via Otsu thresholding) to enforce structural fidelity.
Stage 2: Conditional Diffusion Correction (DDIM)
- A conditional Denoising Diffusion Implicit Model (DDIM) acts as a corrector to refine the coarse predictions from Stage 1.
- The diffusion model is trained to recover high-frequency details lost during the autoregressive rollout. It conditions the denoising process on the coarse prediction ( $\hat{X}_{out}$ ) generated by the frozen Fourier-MIONet.
- Inference: During inference, the process starts with Gaussian noise and iteratively applies deterministic DDIM reverse steps to generate a refined output ( $\tilde{X}_{out}$ ) that corrects accumulated errors and restores fine-scale structures.
- The refined output is fed back as the context for the next autoregressive step, creating a closed-loop refinement system.

Key Contributions

Problem Formulation: The work frames inkjet droplet evolution as a long-horizon prediction task, aiming to predict the entire future sequence (e.g., 550 frames) given an initial short sequence (e.g., 50 frames) and material properties.
Hybrid Framework: The introduction of DiffARFNO, which decouples coarse prediction from fine-detail correction. This approach leverages the efficiency of neural operators for stable rollouts and the generative capability of diffusion models for error correction and detail recovery.
Novel Loss Function: The incorporation of a binary mask-based training loss alongside standard pixel-wise loss to improve prediction accuracy across multiple evaluation metrics, particularly for segmentation tasks.
Comprehensive Evaluation: The method is evaluated across three distinct physical regimes: ligament thinning, pinch-off dynamics, and satellite droplet formation.

Results
The proposed method was evaluated on a dataset of 212 simulated droplet videos generated via ANSYS Fluent (using the Volume of Fluid method). The dataset covers varying densities, viscosities, and surface tensions.

Quantitative Performance: DiffARFNO outperformed state-of-the-art benchmarks (including Fourier-MIONet, PIANO, UFNO, FNO, and GeoFNO) across all six metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), $R^2$ $R^{2}$ , and Intersection over Union (IoU).
- Notably, DiffARFNO achieved an MSE of 0.0003 and an IoU of 0.8711, significantly surpassing the second-best model (Fourier-MIONet) in structural fidelity and segmentation overlap.
Computational Efficiency: While high-fidelity CFD simulations require approximately 4 hours to generate a single droplet evolution sequence, DiffARFNO produces the corresponding prediction in approximately 40 seconds.
Qualitative Results: Visual inspections confirmed that DiffARFNO better preserves fine-scale structures (such as thin ligaments and satellite droplets) and maintains temporal consistency compared to benchmarks, which tended to blur high-frequency features over long horizons.

Significance
The paper claims that DiffARFNO addresses the specific limitations of existing neural operators (error accumulation and blurring) and diffusion models (computational expense when used as standalone long-horizon generators). By synthesizing a fast, stable coarse operator with a lightweight diffusion corrector, the framework enables high-fidelity, long-horizon predictions of complex fluid dynamics at a fraction of the computational cost of traditional CFD. This capability is presented as essential for advancing process monitoring, control, and optimization in inkjet printing and additive manufacturing applications.

Technical Summary: Diffusion-corrected Autoregressive Fourier Neural Operator for Droplet Evolution Prediction

More like this