Karst caves play a vital role in hydrocarbon accumulation and reservoir formation, and accurately delineating cave features is of great significance for the exploration and development of carbonate reservoirs. Traditional seismic P-wave velocity inversion identifies karst caves primarily based on bead-like anomalous reflections, but its accuracy is limited by seismic data quality and structural discontinuities. Although recent deep learning approaches have improved automatic identification capabilities, they often suffer from blurred boundaries and inadequate resolution. To address these issues, we propose a method for predicting cave-related attribute parameters by integrating deep learning with seismic facies-controlled chaotic inversion. This approach combines the cave feature extraction capability of deep learning with the high-resolution advantage of chaotic inversion guided by seismic facies. It effectively overcomes the limitations of deep learning in delineating fine-scale cave geometries and reduces the sensitivity of inversion results to seismic data quality. Furthermore, we establish a porosity calibration model by linking seismic rock physics forward modeling with measured data, enabling the conversion of cave velocities into porosity estimates. This yields high-precision quantitative descriptions of cave attributes. Both theoretical modeling and field application results demonstrate that the proposed method can clearly delineate cave positions and morphologies, enhance vertical and lateral resolution, and provide reliable porosity distribution within caves. These capabilities offer strong technical support for the efficient exploration and development of carbonate reservoirs.