Abstract: To address the mud film quality issues in highly permeable sand-gravel strata, mud penetration simulation tests were conducted under varying formation and mud parameter conditions. A high-precision prediction model was developed based on a Transformer deep learning architecture, and the key factors affecting mud film quality were further analyzed. The results indicate that filtrate volume, mud cake type, and mud cake quality differ markedly among muds in different formations, suggesting that mud formulation should be tailored to formation characteristics. As the formation permeability coefficient increases, the effective particle size of the mud (i.e., particle gradation) becomes the dominant factor influencing filtrate volume, whereas the effect of mud viscosity gradually weakens and eventually becomes negligible. The relationship graphs among filtrate volume, formation permeability coefficient, and mud parameters derived from the model can optimize mud parameter design and provide scientific guidance for mud preparation. Moreover, a mud film quality evaluation method based on the concept of an ideal mud cake is proposed. By introducing the filtrate volume ratio, mud film quality can be effectively assessed, with a reasonable range of 0.4-0.6.