A TabPFN- and CIEDE2000-Optimized Multi-Stage Color Clustering Framework for “Tang Yu”

Quality evaluation of gemstones is a core issue in gemological research, and color is one of the most decisive variables in gemstone grading. However, gemstone color assessment is often affected by perceptual subjectivity and by differences in viewing and measurement conditions. To improve the objectivity and repeatability of color grading for “Tang Yu” nephrite, this study developed a task-oriented hybrid workflow that integrates K-means geometric initialization, TabPFN-based boundary refinement, and CIEDE2000-based perceptual adjustment. Color measurements were performed on 120 “Tang Yu” samples using an X-Rite SP62 spectrophotometer under D65 illumination and an N9 background. Although L*, a*, b*, C*, and h° were reported for interpretation, only the original L*, a*, and b* coordinates were used as algorithmic inputs because C* and h° are derived from a* and b*. Ablation analysis showed that TabPFN acted as a conservative boundary-refinement module, reassigning only low-confidence boundary samples, whereas CIEDE2000 refinement contributed primarily to improved perceptual compactness. Bootstrap resampling with 1000 repetitions supported a statistically robust reduction in intra-cluster mean ΔE00 relative to K-means. The VDR increased on average but was interpreted conservatively because its bootstrap confidence interval crossed zero. These results suggest that the proposed workflow provides a re-producible and perception-oriented reference for small-sample gemstone color grading, while further external and market-based validation remains necessary.