In the era of big-data, large-scale storage systems use NAND Flash-based solid-state disks (SSDs). Some upper-level applications put higher requirements on the performance of SSD-based storage systems. SSDs typically exploit a small amount of DRAM as device side cache, yet the limitation of the DRAM inside an SSD makes a better performance difficult to achieve. The wide application of the existing cache management schemes (e.g., LRU, CFLRU) provides a solution to this problem. With the popularity of near-data processing paradigm in storage systems, the near-data processing-based SSDs are designed to improve the performance of the overall system. In this work, a new cache management strategy named LCache is proposed based on NDP-enabled SSD using a machine learning algorithm. LCache determines whether I/O requests will be accessed in a period by trained machine learning model (e.g., decision tree algorithm model) based on characteristics of I/O requests. When the infrequently accessed I/Os that are not intensive are directly flushed into the flash memory, LCahe enables to update the dirty data that has not been accessed in the cache to the flash memory. Thus, LCache can generate clean data while replace cached data with priority to minimize the cost of data evicting. LCache also can effectively achieve the threefold benefits: (1) reducing the data access frequency to frequently access data pages in flash memory, (2) improving the response time by 59.8%, 60% and 14.81% compared with LRU, CFLRU, and MQSim, respectively, and (3) optimizing the performance cliff by 68.2%, 68%, and 30.2%, respectively.