An explainable and transferable framework for multimodal flame image restoration with physics-smoothed stochastic interpolant

Abstract To address the severe degradation of imaging quality in optical combustion diagnostics under extreme operational environments characterized by high background noise and strong aerodynamic turbulence, this paper proposes a physically explainable and general framework for multimodal flame image restoration featuring exceptional cross-domain transferability. Existing deep learning models for flame enhancement frequently suffer from a critical lack of physical interpretability and poor generalization when deployed across different experimental setups. To overcome these limitations, this study innovatively leverages stochastic interpolant theory to construct a rigorous probability flow path that directly connects degraded, low-quality flame inputs to their high-fidelity target states. Furthermore, to bridge the gap between pure data-driven learning and underlying physical laws, we design a novel Physics-Smoothed Diffusion Block. This dedicated module embeds fundamental fluid dynamic transport and diffusion mechanisms directly into the network architecture as structural priors, thereby strictly enforcing physical conservation principles throughout the generative restoration process. Extensive quantitative and qualitative experimental results demonstrate that the proposed method not only achieves restoration accuracy significantly superior to existing state-of-the-art techniques across diverse combustion modalities—including complex swirl flames, ramjet combustors, and advanced laser diagnostics—but also exhibits robust cross-domain generalization capabilities when transferring smoothly from synthetic training data to real-world experimental scenarios. This work establishes a powerful and reliable solution for high-precision, artifact-free imaging analysis in complex and harsh combustion environments.