Purpose <p>With the continuous scaling of advanced semiconductor manufacturing processes, device critical dimensions (CD) are shrinking rapidly and transistor structures are evolving from planar to complex 3D architectures. These trends make critical dimension small-angle X-ray scattering (CD‑SAXS) an indispensable metrology technique for nanostructures. However, conventional CD‑SAXS workflows suffer from two major efficiency bottlenecks: time‑consuming multi‑angle data acquisition and low‑efficiency inverse modeling based on nonlinear fitting. This study aims to overcome these limitations and realize high‑efficiency CD‑SAXS measurement.</p> Methods <p>A dual‑path acceleration strategy is proposed, which integrates a ResNet34‑based deep regression network with Test‑Time Training (TTT). The neural network is pre‑trained on physically simulated scattering data and then fine‑tuned with unlabeled experimental data to reduce the simulation‑to‑reality domain shift and improve generalization. Raw scattering data is converted into a ω‑qxz coordinate format suitable for convolutional neural networks, thereby avoiding the reconstruction of qx‑qz reciprocal space patterns. The TTT mechanism ensures high prediction accuracy under sparse angular sampling.</p> Results <p>Experimental results demonstrate that the proposed method achieves measurement accuracy comparable to traditional nonlinear fitting methods, while reducing the total time of measurement and data processing by more than 30‑fold.</p> Conclusion <p>The dual‑path acceleration strategy effectively breaks through the efficiency bottlenecks of conventional CD‑SAXS. It enables accurate nanostructure measurement with greatly reduced data acquisition and processing time, showing strong potential for CD‑SAXS in real‑time, in‑line metrology of complex nanostructures in semiconductor manufacturing.</p>