Abstract: Abstract：Disruptive technologies are regarded as a revolutionary force to "change the rules of the game" and "reshape the future pattern". All previous industrial revolutions have shown that whoever has mastered disruptive technologies first will take priority to occupy the technological commanding heights. With the prominence of the leading characteristics of disruptive technologies, they have increasingly risen to the fields of science and technology, national defense and military, and gradually developed to the height of national strategic leadership. Identifying disruptive technologies in advance is of great significance to the cultivation of disruptive technologies and policy formulation. Disruptive technologies are a kind of technology that "blazes a new trail", and "blazes a new trail" means differences from mainstream technologies, so identifying disruptive technologies from the perspective of heterogeneity is more consistent with the law of technological development. Similarity technologies tend to optimize existing technologies and extend the technology track and technology life cycle; Long distance and heterogeneous technologies are more likely to realize technological trajectory transition and produce disruptive technologies. Heterogeneous technology can lead to the variation of technological structure and promote technological innovation and development. From the perspective of heterogeneous patents, this paper proposes an early identification method of potentially disruptive technologies based on machine learning. Firstly, the BERT semantic vector and IPC weight vector are combined to obtain patent features; then a variety of anomaly detection algorithms are used to identify heterogeneous patents, and the patent influence is calculated through the expanded patent co-citation network; then construct heterogeneous patent indicators and technical influence relationship data; Finally, use machine learning algorithms to predict potential disruptive technologies from a large number of heterogeneous patents, and predict the future direction of technology research and development. This paper uses patents in the field of 3D printing (Additive Manufacturing, AM) to verify the effectiveness of this method. The research conclusions mainly include: (1) There are a lot of heterogeneous technologies in the field of 3D printing, but there are few heterogeneous technologies that really have an important impact on this field. It is unrealistic to use traditional expert evaluation methods (2) Among machine learning algorithms, random forest algorithm has the best performance in identifying potentially subversive technologies in 3D printing field. (3) The research and development directions of potential disruptive technologies in 3D printing field in the future are: ① metal 3D printing and preparation of alloy, concrete and ceramic materials; ② The application of 3D printing technology in the fields of medicine, electronic instruments, electrical engineering and electronic energy, as well as the research and development of polymer chemistry, polymers and other materials; ③ Computer, control and communication technology; ④ Chemical engineering, biotechnology and biomaterial analysis; ⑤ The application of 3D printing technology in engines, pumps, turbines, textile and paper machines, mechanical parts, civil engineering, furniture, games, transportation, other consumer goods and other fields. The potential subversive technologies identified in the 3D printing field are in line with the reality and have reference value for policy formulation, enterprise research and development and subsequent academic research.