Accurately measuring the three-dimensional (3D) temperature field of turbulent flames, with a quantifiable measure of confidence, remains a critical yet unsolved challenge for combustion diagnostics. To bridge this gap, this paper introduces an uncertainty-aware Bayesian Generative Adversarial Network (B-GAN) framework. The core innovation lies in its synergistic fusion of low-fidelity 3D light-field data and high-fidelity two-dimensional (2D) temperature measurements within a probabilistic deep learning architecture, enabling simultaneous high-fidelity reconstruction and inherent uncertainty quantification of the 3D temperature field. Trained and validated on numerical simulation data of a hydrogen jet flame from a nine-slot burner, the proposed B-GAN framework demonstrates two key advances. It achieves superior reconstruction accuracy, with quantitative metrics reaching 35.22 dB for peak signal-to-noise ratio and 0.9788 for structural similarity index, outperforming conventional deterministic deep learning models. More importantly, it provides a per-voxel predictive uncertainty map alongside the 3D temperature field. Analysis shows a maximum temperature variation of 10.3 K within the 95% confidence interval across the flame region, explicitly quantifying the reliability of the reconstruction. These results verify that the B-GAN framework not only accurately infers the 3D temperature distribution but also pioneers a probabilistic paradigm for combustion diagnostics. By transforming neural network outputs from single-point predictions to trustworthy predictive distributions, this work establishes a fundamental basis for reliability analysis and risk-informed decision-making in 3D combustion reconstruction.
Su et al. (Fri,) studied this question.