VLM-constrained three-dimensional reconstruction of a hydrogen flame under highly ill-posed imaging conditions

Three-dimensional (3D) reconstruction of turbulent flames under sparse-view tomography remains a big challenge in combustion diagnostics. In this work, we present a vision-language model (VLM)-constrained semantic-feedback framework that reconstructs a 128 × 128 × 128 hydrogen flame field from only three 128 × 128 projections acquired by a compact four-prism imaging system. By leveraging the VLM, Qwen-3.5, as an external evaluator of flame morphology, reaction-zone location, and cross-view consistency, high-level semantic constraints are imposed on this severely underdetermined inverse problem. Compared with a pixel-only self-supervised baseline, our approach improves PSNR from 17.37 to 27.95 dB and SSIM from 0.5436 to 0.7803, while substantially suppressing artifacts. These results demonstrate that VLM-guided semantic constraints significantly enhance quantitative fidelity in sparse-view 3D flame tomography, offering a promising pathway for high-dimensional optical reconstructions from limited measurements.