Optimization of electric motors for modern electric drive systems requires exploring an extensive design space that encompasses not only geometric parameters but also diverse topological structures to enhance performance. However, expanding the design space to cover multiple topologies, such as various rotor types and pole/slot combinations, leads to a combinatorial explosion of design candidates. This massive number of combinations makes traditional Finite Element Analysis (FEA)-based optimization impractical due to excessive computational costs. Using surrogate models is an effective alternative to address these computational burdens. However, constructing a surrogate model for such a wide design space necessitates large-scale training data. Furthermore, even if a model is established, retraining the entire model from scratch whenever the design space expands or new design candidates are added incurs significant computational redundancy, especially given the large-scale of the required dataset. To address these challenges, this study proposes a computationally efficient data-driven optimization strategy for variable topologies. We utilize large-scale datasets generated via the Equivalent Magnetic Circuit (EMC) method, which serves as a computationally efficient alternative to high-fidelity simulations. Based on this extensive data, we introduce an incremental learning strategy into the surrogate modeling process. This approach enables the model to continuously learn from new design spaces without forgetting previously acquired knowledge, thereby eliminating the need for full retraining. The proposed framework is applied to electric motor optimization problems, demonstrating its ability to effectively explore expanded design spaces with high computational efficiency.