We develop a novel circuit-level stochastic local search (SLS) method D-CRSat for Boolean satisfiability by integrating a structure-based heuristic into the recent CRSat algorithm. D-CRSat significantly improves on CRSat on real-world application benchmarks on which other current CNF and circuit-level SLS methods tend to perform weakly. In some cases D-CRSat compares favourably even with a modern circuit-level conflict-driven clause learning solver. We also give an intricate proof of probabilistically approximate completeness for D-CRSat, highlighting key features of the method.