<p>Multi-fidelity data fusion (MDF) is a cost-effective surrogate modeling technique, yet most methods require paired or strongly correlated data—an assumption often violated in practice. This leads to structurally unpaired and misaligned datasets, creating a major bottleneck. To address this, we propose the multi-fidelity stepwise convolutional neural network (MF-SCNN), a deep convolutional framework for unpaired data fusion. Our framework pioneers a "point-to-domain" strategy: it first generates pseudo-high-fidelity samples to enrich the data, then transforms all unpaired data into structured, image-like tensors via a novel encoding method. A multi-scale dilated convolutional network processes these tensors to extract hierarchical spatial and cross-fidelity features for final prediction. On benchmarks and engineering cases, MF-SCNN achieves a 28–31% RMSE reduction over state-of-the-art methods under challenging unpaired conditions. This demonstrates a robust solution for complex engineering problems where data alignment cannot be guaranteed, effectively overcoming a key limitation in conventional MDF.</p>