[Purpose/significance] Aiming at the problems of single feature dimension and insufficient interpretability in the early identification of current breakthrough papers, this study constructs an interpretable machine learning meth⁃od to improve the identification accuracy and logical transparency, providing methodological support for scientific re⁃search management and innovation layout. [Method/process] Firstly, starting from the connotation of breakthrough pa⁃pers, on the basis of traditional features, innovation attribute measurement is introduced, knowledge innovation vectors are obtained by inducing large language Prompts, and a multi-dimensional feature system of breakthrough papers is constructed; secondly, an information expression system is built based on formal concept analysis (FCA), core features are screened by combining statistical correlation analysis and FCA attribute reduction algorithm, and various machine learning classifiers are used to predict the model identification effect; finally, a two-layer interpretation framework based on FCA concept lattice and SHAP analysis is constructed to form a visual interpretation chain from screening rules to prediction verification. [Result/conclusion] XGBoost model has an F1 value of 0.952 on multi-disciplinary da⁃tasets, which is significantly better than traditional methods; the two-layer interpretation system clarifies the feature combination rules of high innovation attributes and strong knowledge correlation, and quantifies the contribution of sin⁃gle features to the prediction results.