Termination is a fundamental liveness property for program verification. In this paper, we revisit the problem of non-termination analysis and propose the first black-box learning algorithm for synthesizing recurrent sets, where the non-terminating samples are effectively speculated by a novel method. To ensure convergence of learning, we develop a learning algorithm which is guaranteed to converge to a valid recurrent set if one exists, and thus establish its relative completeness. The methods are implemented in a prototype tool and experimental results on public benchmarks show its efficacy in proving non-termination as it outperforms state-of-the-art tools, both in terms of cases solved and performance. Evaluation on non-linear programs also demonstrates its ability to handle complex programs.