Abstract: Breakthrough research is groundbreaking, forward-looking and leading scientific research that may lead to profound changes in scientific paradigms. Breakthrough papers are an important carrier and manifestation of breakthrough research. Early identification of breakthrough papers plays an important role in leading the frontier exploration of scientific research, guiding the efficient allocation of corporate R&D resources, and supporting the forward-looking layout decision-making of the government's innovation strategy. Therefore, how to efficiently and accurately identify breakthrough papers in the process of scientific and technological innovation has become a research hotspot in academic community. However, most of the current research on the identification of breakthrough papers focuses on some indicator characteristic data of the papers themselves. The selected indicators lack a close connection with the essential characteristics of scientific breakthroughs, and there is still a lack of an indicator system for identifying breakthrough papers from the perspective of knowledge breakthrough, knowledge innovation, and knowledge interdisciplinary. In addition, when using bibliometric methods to identify breakthrough papers, most of the existing identification methods are based on citation analysis methods, which has a time lag in early identifying breakthrough papers. Therefore, in view of the shortcomings in the current research on early identifying breakthrough papers, in this paper, we proposed a method for early identification of breakthrough papers based on machine learning. The research ideas of this method are as follows: firstly, according to the nature of breakthrough papers, an indicator system for identifying breakthrough papers was constructed from three dimensions: knowledge breakthrough, knowledge innovation, and knowledge interdisciplinary. At the same time, the characteristics of the subject of knowledge discovery or creation will also have an important impact on whether the knowledge is breakthrough. Thus, in order to be able to identify breakthrough papers more comprehensively and accurately, we added the characteristics of the authors of the papers when constructing the indicator system for identifying breakthrough papers, and constructed a comprehensive identification indicator system including the text characteristics, quantitative characteristics, and author characteristics of breakthrough papers. Secondly, we introduced machine learning methods into the research on the identification of breakthrough papers. By using the nonlinear relationship pattern acquisition ability of machine learning, we extracted the correlation relationship pattern between the characteristics of breakthrough papers and their breakthrough, and used this pattern to identify breakthrough papers at an early stage. This method solves the time lag problem of identifying breakthrough papers. Finally, taking the biomedical field as an example, we verified the feasibility and effectiveness of this method. The results of the case study in the biomedical field show that starting from the essence of breakthrough research, integrating the text characteristics, quantitative characteristics and author characteristics of the paper can more comprehensively and systematically characterize the breakthrough information of the paper; the use of machine learning algorithms can effectively solve the time lag problem in the identification of breakthrough papers and can achieve early identification of breakthrough papers. The breakthrough papers identification method based on machine learning provides a new research method for the early identification of breakthrough papers.