🤖 machine learning

Di-BiLPS: Denoising induced Bidirectional Latent-PDE-Solver under Sparse Observations

The paper introduces Di-BiLPS, a unified neural framework that leverages variational autoencoders, latent diffusion, and contrastive learning to solve forward and inverse PDE problems with state-of-the-art accuracy and efficiency under extremely sparse observations (as low as 3%) while enabling zero-shot super-resolution.

Original authors: Zhonghao Li, Chaoyu Liu, Qian Zhang

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhonghao Li, Chaoyu Liu, Qian Zhang

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

The Big Problem: The "Blind" Puzzle

Imagine you are trying to solve a massive, complex jigsaw puzzle that represents how water flows through a pipe, how heat spreads through metal, or how air moves around a plane wing. In the real world, these are called Partial Differential Equations (PDEs).

Usually, to solve these puzzles, you need a complete picture of every single piece. But in real life, we often only have a few scattered pieces—maybe just 3% of the data. It's like trying to guess the entire weather pattern for a whole country by looking at only three thermometers scattered across a field.

Old computers (classical solvers) get stuck because they need a full grid of data to work.
Old AI models get confused because they try to guess the missing pieces based on patterns, but when the data is this sparse, they make wild guesses that are often wrong.

The Solution: Di-BiLPS

The authors created a new AI system called Di-BiLPS. Think of it as a super-smart detective who can solve the puzzle using only those three scattered thermometers, while also knowing the laws of physics (like "heat always moves from hot to cold").

Here is how Di-BiLPS works, broken down into three simple steps:

1. The "Compression Suit" (The VAE)

Imagine you have a giant, messy room full of furniture (the complex data). It's too big to carry around.

What Di-BiLPS does: It puts all that furniture into a tiny, magical suitcase (a Latent Space). It doesn't throw anything away; it just folds it up perfectly so it fits in a small box.
Why? Doing math on a tiny suitcase is much faster than doing math on a giant room. This allows the AI to work at high speeds without getting overwhelmed.

2. The "Translator" (Contrastive Learning)

Now, imagine you have a blurry, low-resolution photo of a face (your sparse data) and a crystal-clear photo of the same face (the full data you want to predict). They look very different.

What Di-BiLPS does: It uses a special "translator" to learn how to match the blurry photo to the clear one. It learns that "this one dot on the blurry photo corresponds to this specific curve on the clear photo."
Why? This helps the AI understand the relationship between the tiny bit of data it has and the full solution it needs to find.

3. The "Physics Detective" (Diffusion & Denoising)

This is the magic part. Imagine you have a noisy, static-filled TV screen. You want to see the clear image underneath.

Standard AI: Tries to guess the image by looking at the noise.
Di-BiLPS: It uses a process called Diffusion. It starts with pure static (random noise) and slowly "denoises" it, step-by-step, to reveal the image.
The Twist: Most AIs just guess what the image looks like. Di-BiLPS also checks the Laws of Physics at every step. If the AI guesses a solution that breaks the laws of physics (like water flowing uphill), the system says, "Nope, try again," and nudges the solution back into the right direction.
The Result: It cleans up the noise while strictly obeying the rules of nature, ensuring the final answer is both accurate and physically possible.

What Can It Do?

The paper claims Di-BiLPS is a "two-way street" solver:

Forward Problem: If you know the starting conditions (like the shape of a pipe), it can predict how the fluid will flow.
Inverse Problem: If you see the result (the fluid flow), it can figure out what the starting conditions were (what the pipe looked like).

It can do both of these even when it only has 3% of the data.

The "Superpower": Zero-Shot Super-Resolution

Here is the coolest trick. Imagine you trained the AI on a low-resolution map (like a pixelated video game). Usually, if you ask it to show you a high-definition version, it would just make the pixels bigger and blurrier.

Di-BiLPS is different. Because it learned the "rules" of the shape in its compressed suitcase, you can ask it: "Show me the solution at this exact tiny point I'm pointing to," even if that point is in a high-resolution area it never saw before. It can generate a crystal-clear, high-definition answer instantly without needing to be retrained.

Summary of Results

The authors tested this on five different complex physics problems (like fluid flow and heat).

Accuracy: It was much more accurate than previous AI methods when data was extremely scarce (down to 3%).
Speed: It was about 90% faster than the previous best AI method (DiffusionPDE) because it does the heavy math in the "compressed suitcase" rather than the full room.
Reliability: It works for both predicting the future (forward) and guessing the past (inverse).

In short: Di-BiLPS is a smart, physics-aware AI that can solve complex natural puzzles using very little data, work incredibly fast, and draw high-definition pictures from low-resolution clues.

Technical Summary: Di-BiLPS

Problem Statement

Partial differential equations (PDEs) are fundamental for modeling natural and physical phenomena, yet their practical application is often hindered by the scarcity of observational data. In real-world scenarios, measurements are frequently extremely sparse across both spatial and temporal domains (e.g., observation rates below 3%). Under such conditions, classical numerical solvers and existing neural approaches (including state-of-the-art neural PDE models) struggle to maintain accuracy and reliability. Furthermore, while neural methods have shown promise under moderately sparse conditions, they often suffer from limited inference efficiency at high resolutions and substantial accuracy degradation in extremely sparse regimes. There is a critical need for a unified framework capable of robustly handling both forward (simulating system states) and inverse (inferring parameters) PDE problems under severe data sparsity while maintaining computational efficiency.

Methodology

The authors propose Di-BiLPS (Denoising induced Bidirectional Latent-PDE-Solver), a unified neural framework designed to operate effectively under extremely sparse observations. The architecture integrates three core components:

Variational Autoencoder (VAE) for Irregular Data:
To handle inputs defined on irregular grids and continuous solution spaces, the framework employs a VAE built upon a U-Net architecture augmented with GINO-based mesh encoders and decoders. This module compresses high-dimensional inputs (coefficients $A$ and solutions $U$ ) into a compact latent space. By performing learning in this compressed latent space, the model accelerates convergence and enhances generalization capabilities.
Contrastive Learning for Sparse Observations:
To bridge the gap between sparse observations and full solution spaces, a contrastive learning module (GiT) is introduced. This module utilizes a hybrid encoder combining GINO-based mesh encoders and Vision Transformers (ViT). It aligns representations between sparse point clouds (partial observations) and their corresponding full observation spaces. By optimizing matching scores via cross-entropy loss, the framework learns robust conditioning embeddings ( $C_A, C_U$ ) that facilitate efficient mappings from sparse inputs to full solutions.
Latent Diffusion with PDE-Guided Denoising:
The core inference engine is a latent diffusion model that enables bidirectional inference within the compressed latent space. A novel PDE-informed denoising algorithm is introduced, based on a variance-preserving diffusion process. This algorithm integrates physical constraints directly into the sampling process. Specifically, it employs a guided diffusion approach where the noise estimation is adjusted by two gradient terms:
- PDE Guidance ( $N_{pde}$ ): A term derived from the governing equations (measurement operator $M$ ) to ensure physical consistency.
- Observation Guidance ( $N_{obs}$ ): An $\ell_2$ penalty term that minimizes the discrepancy between the reconstructed solution at observed points and the actual sparse measurements.
  This dual-guidance mechanism allows the model to perform accurate bidirectional inference (forward and inverse) while adhering to physical laws and sparse data constraints.

Key Contributions

The paper outlines four primary contributions:

Unified Framework: The proposal of Di-BiLPS, a scalable neural framework that jointly addresses forward and inverse PDE problems under extremely sparse observations (as low as 3%). It reformulates PDE learning in a compressed latent space to achieve improved computational efficiency and flexible input-output mapping compared to existing methods.
PDE-Informed Denoising: The development of a denoising algorithm based on variance-preserving diffusion that effectively integrates sparse observations with physical constraints, enabling accurate bidirectional inference.
Performance and Efficiency: Extensive experiments on five PDE benchmark datasets (Darcy Flow, Helmholtz, Poisson, Bounded/Non-Bounded Navier-Stokes) demonstrate that Di-BiLPS consistently achieves state-of-the-art (SOTA) performance in both prediction accuracy and computational efficiency. Notably, it reduces inference time by approximately 90% compared to DiffusionPDE while significantly lowering prediction errors.
Zero-Shot Super-Resolution: Leveraging the continuous nature of the GINO-based decoder, Di-BiLPS supports zero-shot resolution generalization, allowing predictions over continuous spatial-temporal domains and on unseen spatial resolutions without retraining.

Experimental Results

The authors evaluated Di-BiLPS on five standard PDE benchmarks under sparse observational settings (3% observation rate).

Accuracy: The model consistently outperformed baselines including PINO, DeepONet, PINNs, FNO, DiffusionPDE, and GraphPDE. For instance, in the forward Darcy Flow task, Di-BiLPS achieved a relative $\ell_2$ error of 0.021 compared to 0.025 for DiffusionPDE and significantly higher errors for other baselines. In inverse problems, such as the Helmholtz equation, it reduced errors by 47% compared to DiffusionPDE.
Efficiency: Di-BiLPS demonstrated a substantial reduction in computational cost, with inference times roughly 90% lower than DiffusionPDE (e.g., ~20 seconds vs. ~255 seconds).
Ablation Studies: Removing the observation guidance led to significant performance degradation across all tasks, highlighting its critical role. Removing the condition guidance caused severe deterioration in complex scenarios. Interestingly, the PDE guidance term showed a negligible impact on performance in the ablation study, suggesting its role is auxiliary in the current latent space implementation.
Generalization: The model maintained stable performance as sparsity increased from 10% to 1%, though performance degraded gracefully at 0.1%.

Significance and Claims

The paper claims that Di-BiLPS offers a robust solution for scientific computing problems where data is severely limited, a common constraint in industrial and real-world applications. By leveraging carefully constructed low-rank latent representations and diffusion-based models, the framework demonstrates that accurate and efficient PDE solving is possible even with extremely sparse inputs. The authors posit that the proposed PDE denoising algorithm and the unified bidirectional framework provide valuable insights for future research on complex PDE challenges. They emphasize that the method supports zero-shot super-resolution, enabling inference on continuous domains without the need for retraining on specific grid resolutions.

The authors acknowledge limitations, noting that the PDE guidance term currently offers only marginal improvements in their specific latent space implementation and that performance deteriorates drastically when observation rates drop below 0.1%. They suggest that exploring the minimal sparsity threshold remains a direction for future work.