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.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the problem of reconstructing physical fields from sparse measurement data. Specifically, it focuses on how to recover the spatio - temporal information of physical systems from limited observational data in scientific and engineering applications. Although traditional methods can extract features, they often have low precision when dealing with complex physical systems and are difficult to meet basic physical constraints (such as governing equations and boundary conditions). To this end, the authors propose a new data - driven physical field reconstruction framework - **Physics - aligned Schrödinger Bridge (PalSB)**. #### Main problems and challenges 1. **Low precision and inconsistent physical constraints**: - Existing deep - learning models have low precision when dealing with complex physical systems. - These models are usually unable to strictly abide by physical laws, such as governing equations and boundary conditions. 2. **Ill - posed nature of inverse problems**: - Physical field reconstruction is essentially an inverse problem with an ill - posed nature, making reconstruction particularly difficult. - Complex spatio - temporal interactions (such as turbulence) further increase the difficulty of the problem. 3. **High computational cost**: - Traditional physics - based methods need to be repeatedly calibrated to balance observational data and physical laws, resulting in high computational costs. 4. **Non - convexity and infeasible regions**: - The domains defined by physical laws are usually non - convex and intractable, making direct projection methods inapplicable. - Using physics - informed neural networks (PINNs) to embed physical laws as soft constraints may lead to convergence problems. ### Solutions To solve the above problems, the authors propose the **Physics - aligned Schrödinger Bridge (PalSB)** framework, which has the following main features: - **Two - stage training strategy**: - **Pre - training stage**: Use the diffusion Schrödinger bridge (DSB) to generate local high - resolution fields from sparse measurements. - **Fine - tuning stage**: Further optimize the model through a physics - informed loss function to make it more in line with physical principles. - **Boundary - aware sampling technique**: - Ensure that the generated physical fields meet the boundary conditions and improve global consistency. - **Innovative sampling process**: - Propose techniques such as boundary - aware sampling (BA) and early - stopping sampling (ES) to improve sampling efficiency and ensure the consistency of boundary conditions. Through these improvements, PalSB not only improves the reconstruction accuracy but also significantly enhances the ability to follow physical constraints, and is suitable for complex non - linear systems, such as cylinder flow, two - dimensional turbulence and reaction - diffusion systems. ### Experimental verification The authors verify the effectiveness of PalSB through three complex non - linear systems (cylinder flow, two - dimensional turbulence and reaction - diffusion systems). The experimental results show that PalSB is superior to existing methods in both point accuracy and physical constraint consistency, demonstrating its potential in high - fidelity physical field reconstruction.

Physics-aligned Schrödinger bridge

Flow Reconstruction with Uncertainty Quantification from Noisy Measurements Based on Bayesian Physics-Informed Neural Networks

DPO-SDF: Differentiable Physical Optics Method with Signed Distance Field for 3-D Reconstruction of Perfect Electric Conductors

Aligned Diffusion Schrödinger Bridges

Physical Constraint Flow Field Reconstruction Method for Immersion Scan-Step Flow

Physics informed neural fields for smoke reconstruction with sparse data

Physiome Model Based State-Space Framework for Cardiac Kinematics Recovery

PhyRecon: Physically Plausible Neural Scene Reconstruction

DeepPhysics: a physics aware deep learning framework for real-time simulation

Physics-constrained bayesian neural network for fluid flow reconstruction with sparse and noisy data

Discovering Nonlinear PDEs from Scarce Data with Physics-encoded Learning

PhySR: Physics-informed Deep Super-resolution for Spatiotemporal Data

Guided MRI Reconstruction via Schrödinger Bridge

Physics-enhanced deep surrogates for partial differential equations

Physics-guided machine learning from simulated data with different physical parameters

Provably Convergent Schrödinger Bridge with Applications to Probabilistic Time Series Imputation

A novel physics-informed framework for reconstruction of structural defects

PARCv2: Physics-aware Recurrent Convolutional Neural Networks for Spatiotemporal Dynamics Modeling

Bridging deep learning force fields and electronic structures with a physics-informed approach

Surrogate Modeling for Porous Flow Using Deep Neural Networks

PiRD: Physics-informed Residual Diffusion for Flow Field Reconstruction