Physics-aligned Schrödinger bridge

Zeyu Li, Hongkun Dou, Shen Fang, Wang Han, Yue Deng, Lijun YangarXiv 2024 · arXiv · OpenReview

TL;DR

This paper introduces PalSB, a novel deep learning framework for reconstructing complex physical fields from sparse data. By combining a diffusion Schrödinger bridge with physics-informed constraints and a boundary-aware sampling strategy, PalSB achieves higher accuracy and better adherence to physical laws than existing methods, making it promising for scientific and engineering applications.

What Problem Does This Paper Address?

Reconstructing detailed physical fields (like fluid flow or temperature distributions) from limited sensor data is crucial in many scientific and engineering fields. While deep learning models excel at learning from data, they often struggle with two key challenges: achieving high accuracy on complex, nonlinear systems and, more importantly, ensuring that the reconstructed fields obey fundamental physical laws (governing equations and boundary conditions). Traditional physics-based methods can be computationally expensive, and existing deep learning approaches often fail to enforce these physical constraints, leading to inaccurate or unrealistic results.

Methods

The core of the PalSB framework is a two-stage training process built upon the Diffusion Schrödinger Bridge (DSB) mechanism.

  1. Localized Pretraining: Instead of training on full, high-resolution fields (which is computationally demanding and often impossible due to data limitations), PalSB first trains a neural network on small, randomly cropped local patches. This makes the training more data-efficient and scalable. The DSB is used here to learn local, high-resolution field generation from sparse measurements.
  2. Physics-aligned Finetuning (PF): After pretraining, the model is further refined using a physics-informed loss function. This stage explicitly encourages the generated fields to satisfy the governing physical equations. A key innovation here is how they handle the challenges of optimizing physics-informed losses in generative models, drawing inspiration from reinforcement learning techniques. They also incorporate a regression loss to maintain stability.
  3. Boundary-aware Sampling (BA): To ensure that the reconstructed fields respect physical boundary conditions, PalSB employs a novel sampling strategy. This involves padding the input data according to the boundary conditions before interpolation and then trimming the padded region from the final output.
  4. Early-stop Sampling (ES): As an additional enhancement for efficiency, they found that using smaller step sizes and stopping the sampling process earlier (with fewer function evaluations) can yield comparable or even better results than traditional multi-step sampling.

Datasets

The paper evaluates PalSB on three distinct physical systems:

  • Cylinder flow (PIV): Data from particle image velocimetry experiments, focusing on velocity vectors.
  • 2D forced turbulence (Kolmogorov flow): A high-fidelity dataset of complex spatiotemporal patterns governed by Navier-Stokes equations.
  • Reaction-diffusion system (Gray-Scott type): A system commonly used in biological and chemical applications.

For each dataset, they consider two tasks: Fourier-space interpolation (FI) and real-space interpolation (RI).

Key Insights and Ideas

  • Patch-based pretraining is crucial for efficiency: Training on local patches significantly reduces computational burden and makes the approach scalable to high-resolution fields, circumventing the need for full-field data during initial training.
  • DSB as a foundation for reconstruction: Leveraging DSB for initial high-resolution field generation provides a strong starting point for subsequent physics-informed refinement.
  • Physics-informed finetuning needs careful handling: Directly applying physics-informed losses to generative models can lead to convergence issues. PalSB's approach of using a pretraining stage and inspiration from RLHF helps stabilize this process.
  • Boundary conditions matter and can be enforced: The boundary-aware sampling strategy is a practical way to ensure global coherence and adherence to physical boundaries, which is often overlooked in purely data-driven methods.
  • Early stopping can be surprisingly effective: For a fixed number of function evaluations, a carefully chosen early-stop strategy can achieve excellent results, balancing performance and computational cost.

Key Results

PalSB demonstrates superior performance across all tested systems and tasks compared to baseline methods like interpolation, End-to-End (E2E) models with physics-aligned finetuning, and PIDM.

  • Quantitative Improvements: PalSB consistently achieves lower normalized root mean squared error (nRMSE) for point-wise accuracy and significantly lower normalized equation residual (nER) for physical insatisfaction. For example, in the Kolmogorov flow FI task, PalSB achieves an nER of 0.240, while the next best method (SB) is 0.799.
  • Qualitative Improvements: Visual comparisons (Figures 2, 10-17) show that PalSB recovers finer spatial patterns and more accurate physical structures, especially in complex turbulent flows, compared to other methods.
  • Robustness to Noise: PalSB maintains strong performance even when the input low-fidelity data is perturbed with Gaussian noise (Table 3).

Limitations and Future Work

  • Soft Constraint Encoding: The paper acknowledges that constraints are encoded in a "soft manner," meaning some residual discrepancies from exact equations are still possible.
  • Intractable Manifolds: Directly generating complex physical systems on the exact solution manifold of governing equations (especially for highly nonlinear systems like turbulence) remains a significant challenge.
  • 2D Limitation: The current study is limited to two-dimensional examples due to computational constraints. Extending the framework to three-dimensional cases is a critical next step for broader applicability.

Summary updated

Abstract

The reconstruction of physical fields from sparse measurements is pivotal in both scientific research and engineering applications. Traditional methods are increasingly supplemented by deep learning models due to their efficacy in extracting features from data. However, except for the low accuracy on complex physical systems, these models often fail to comply with essential physical constraints, such as governing equations and boundary conditions. To overcome this limitation, we introduce a novel data-driven field reconstruction framework, termed the Physics-aligned Schrödinger Bridge (PalSB). This framework leverages a diffusion Schrödinger bridge mechanism that is specifically tailored to align with physical constraints. The PalSB approach incorporates a dual-stage training process designed to address both local reconstruction mapping and global physical principles. Additionally, a boundary-aware sampling technique is implemented to ensure adherence to physical boundary conditions. We demonstrate the effectiveness of PalSB through its application to three complex nonlinear systems: cylinder flow from Particle Image Velocimetry experiments, two-dimensional turbulence, and a reaction-diffusion system. The results reveal that PalSB not only achieves higher accuracy but also exhibits enhanced compliance with physical constraints compared to existing methods. This highlights PalSB's capability to generate high-quality representations of intricate physical interactions, showcasing its potential for advancing field reconstruction techniques.
BibTeX
@misc{li2024physics,
  author = {Zeyu Li and Hongkun Dou and Shen Fang and Wang Han and Yue Deng and Lijun Yang},
  title = {{Physics-aligned Schrödinger bridge}},
  year = {2024},
  eprint = {2409.17825},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.17825},
}
Figure