Generative Cross-Scale Refinement Transformer for Supersonic Combustion Flow Prediction in Scramjets

Supersonic combustion flow prediction using scramjet wall pressure data shows promise, but fully supervised models are hindered by scarce paired flame-pressure measurements. To leverage abundant unlabeled flame images, we propose the Generative Cross-Scale Refinement Transformer (GCSRT). GCSRT integrates masked image modeling (MIM) for self-supervised pretraining on unpaired flame images, enabling efficient fine-tuning to predict future flame states solely from current wall pressure. The architecture combines a pressure encoding network (PEN) with a masked pyramid decoder (MPD). The MPD first learns visual representations via MIM pretraining without pressure labels. It then cascades with the PEN for end-to-end fine-tuning on limited pressure–flame pairs. Evaluated against Vision Transformer (ViT), EfficientMod, and a fully supervised variant (GCSRT-SL) using short- and long-term metrics, GCSRT achieves superior performance on independent tests. Results confirm that MIM pretraining effectively embeds implicit physical priors. This work demonstrates self-supervised learning’s value for physical flowfield modeling under data scarcity; GCSRT’s ability to bridge pressure sensors and flame diagnostics; and new pathways for reconstructing incomplete flowfields and advancing intelligent flow perception. Our approach establishes a paradigm for exploiting unlabeled data in high-speed combustion diagnostics.