The objective of infill well placement optimization is to determine the optimal well locations that maximize the net present value (NPV). The most common method of well infilling in oil field is based on the engineer’s knowledge, which is risky. Additionally, numerous optimization techniques have been proposed to address the issues. However, locating the global optimum in a large-scale practical reservoir model is computationally expensive, even more so in the realistic extra-low permeability reservoir, where fractures are generated and underground conditions are complex. Thus, both determining well locations solely through human experience and obtaining them through traditional optimization methods have disadvantages in actual engineering applications. In this paper, we propose an infill well optimization strategy based on the divide-and-conquer principle that divides the large-scale realistic reservoir model into several types of small-scale conceptual models using human knowledge and then uses the surrogate-assisted evolutionary algorithm to obtain the infill well laws for this reservoir. The diamond inversed nine-spot well patterns are studied and summarized to provide the optimal infill well placement laws for extra-low permeability reservoirs. Additionally, the laws are implemented in W-77 actual reservoir and the oil recovery has an equivalent increase of 2.205%. The results demonstrate the proposed method’s strong engineering potential and application value, as it combines the benefits of human experience and evolutionary algorithms to determine the optimal infill well placement in a realistic extra-low permeability reservoir development scenario.