A Physics-Informed Multiscale Diffusion Model Framework for Physical Field Imputation From Randomly Missing Data
Journal of Turbomachinery · 01.12.2026 · 148 (12)
DOI: 10.1115/1.4072581 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Turbomachinery, online erschienen am 03.09.2026. Verfasst von Bo Yu und 2 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben.
Abstract (Originalsprache)
Nonintrusive optical measurement techniques are widely employed to obtain high-resolution distributions of physical fields, such as pressure, temperature, or velocity, but frequently suffer from random data loss due to factors like geometric obstructions, surface reflections, or illumination issues. Traditional reconstruction methods often rely heavily on high-fidelity computational fluid dynamics (CFD) priors or require large-scale paired datasets, limiting their flexibility in handling arbitrary missing patterns with limited experimental samples. To address these challenges, this study proposes a physics-informed multiscale resampled denoising diffusion probabilistic model (MSR-DDPM) framework that shifts the reconstruction paradigm from traditional deterministic mapping to probabilistic distribution modeling. The framework integrates three key innovations: (1) a multiscale hierarchical reconstruction strategy that reduces computational cost by 56%, while enhancing stability; (2) embedded physics-informed constraints, such as divergence and gradient continuity terms, to ensure physical consistency and reduce dependency on large-scale training samples; and (3) an optional conditional diffusion framework that leverages auxiliary low-fidelity data to break through performance bottlenecks in extreme missing scenarios. In this work, the framework is validated for incompressible velocity field imputation using data from a turbine cascade passage to demonstrate its capability in capturing complex unsteady flow features. Results show excellent performance: for missing ratios of 10–45%, the model achieves high-precision imputation with symmetric mean absolute percentage error (SMAPE) maintained below 6% without auxiliary data. Furthermore, under extreme conditions (45–65% missing ratios), the integrated auxiliary guidance significantly recovers flow details, reducing errors by approximately 60%. The MSR-DDPM provides an innovative, data-efficient solution for missing data imputation in complex experimental flow measurements.
Zitieren
APA
Yu, B., Chen, P., & Mao, J. (2026). A Physics-Informed Multiscale Diffusion Model Framework for Physical Field Imputation From Randomly Missing Data. Journal of Turbomachinery, 148(12). https://doi.org/10.1115/1.4072581
Vancouver
Yu B, Chen P, Mao J. A Physics-Informed Multiscale Diffusion Model Framework for Physical Field Imputation From Randomly Missing Data. Journal of Turbomachinery. 2026;148(12). doi:10.1115/1.4072581
BibTeX
@article{Yu2026,
author = {Yu, Bo and Chen, Pingting and Mao, JunKui},
title = {A Physics-Informed Multiscale Diffusion Model Framework for Physical Field Imputation From Randomly Missing Data},
journal = {Journal of Turbomachinery},
year = {2026},
volume = {148},
number = {12},
doi = {10.1115/1.4072581}
}
EUVIMED