The research subject is silicon oxide (SiO₂) trench wafers fabricated by dry etching in semiconductor copper-interconnect processes. Key structural parameters are as follows: designed trench width 25 nm, designed trench depth 120 nm, silicon-oxide dielectric-layer thickness 150 nm, and monocrystalline silicon substrate. Dry etching is performed using SF₂/O₂ mixed gas [9]. For sample preparation, wafers are diced into 10 mm × 10 mm chips. Chips are slightly cleaned in dilute hydrofluoric acid solution for 8 s to remove surface oxide residues, rinsed with deionized water and blow-dried with nitrogen. The prepared samples are directly used for SEM observation to avoid damage to internal trench structures. Meanwhile, TEM reference samples are prepared. Focused ion beam (FIB) technology is adopted to cut 100-nm-thick ultra-thin slices containing complete trench cross-sections from identical chips [10]. These slices are used for real-world TEM observation, construction of paired AI training datasets and validation of simulation outputs.

A field-emission scanning electron microscope is utilized for SEM observation with an acceleration voltage of 10 kV and working distance of 8 mm. Both secondary-electron (SE) images (for clear visualization of trench side-wall morphology) and back-scattered-electron (BSE) images (for identification of trench-bottom morphology) are collected [11][12]. Figure 2 presents signals collectable by images under various SEM viewing angles. Specifically, SE images at 0˚ (front view, for observing trench openings, surfaces and upper side-walls), SE images at 10˚ (for observing full side-wall profiles), and BSE images at 0˚ (front view, for observing full trench-bottom features) are acquired.

A transmission electron microscope operating at 200 kV acceleration voltage is used for TEM measurement. Real-world TEM images are captured targeting the corresponding trench regions characterized by SEM. Key features including side-wall verticality, bottom flatness and internal defects are recorded to construct paired datasets of “multi-view SEM fused images - ground-truth TEM images [13]”.

Figure 2. Acquisition of SEM images from different viewing angles.

Step 1: Image preprocessing is implemented in MATLAB, including Gaussian filtering for noise reduction and contrast-limited adaptive histogram equalization (CLAHE) for contrast enhancement [14].

Step 2: Sub-pixel registration and pixel-level weighted fusion of multi-view SEM images.

Taking the SE (0˚) image as the reference, offsets of SE (10˚) and BSE (0˚) images are corrected to guarantee one-to-one pixel-coordinate correspondence among three types of images. Pixel-wise weighted summation is conducted with predefined weights to generate full-field morphological fused 2D planar images, completely preserving parallax features induced by 10˚ tilt [15]:

Ifusion(x, y) = 0.5·ISE10(x, y) + 0.2·ISE0(x, y) + 0.3·IBSE0(x, y)

Although the fused output is a 2D planar image, it implicitly encodes spatial information corresponding to trench depth, longitudinal side-wall profiles and bottom contrast derived from parallax, serving as the sole input for the AI model.

Step 3: Three lightweight AI models with easy deployment are investigated, targeting SSIM ≥ 0.93 [16][17].

(1) Lightweight miniature U-Net supervised mapping model

Datasets are constructed by pairing full-field SEM fused projective 2D images with real TEM cross-sectional images obtained from FIB-prepared samples. All images are unified to identical resolution, and microscopy-oriented data augmentation is applied to expand sample volume. The dataset is split into training set, validation set and test set at a ratio of 7:2:1. On the basis of the original U-Net architecture, redundant deep convolution modules are removed for lightweight implementation, while encoder-decoder skip-connection structures are retained. The encoder extracts multi-scale features including parallax, contrast and edges from SEM fused images; the decoder reconstructs and outputs simulated TEM grayscale cross-sectional images. During training, SSIM is adopted as the characterization metric. The Adam optimizer is used for iterative training, and an early-stopping strategy based on validation-set metrics is applied to save optimal model weights. In the inference phase, simulated TEM images are generated at millisecond-level latency upon input of SEM fused images of target trenches. After grayscale normalization, outputs are quantitatively evaluated against measured TEM images via SSIM, PSNR and trench geometric parameters to realize simulation of cross-sectional morphology for trench side walls, bottoms and depth.

(2) Multi-view branch conditional GAN model

Images are not fused in advance. SE 0˚, SE 10˚ and BSE 0˚ images are fed into independent feature-extraction branches in separate channels. Multi-view parallax features are fused inside the network, and the discriminator constrains output morphology referring to real TEM data. This model delivers outstanding performance in restoring tiny side-wall damages and etch residues at trench bottoms.

(3) Depth-prior-aided lightweight auto encoder (AE)

Simple 2D depth-distribution priors are generated merely via parallax interpolation without constructing complete three-dimensional meshes. Depth priors together with SEM fused images are fed into the auto encoder. The SSIM loss function is adopted to optimize outputs, and this model requires fewer labeled samples.

Step 4: Quantitative validation methods.

Ground-truth TEM and AI-generated simulated TEM images are converted into standardized 2D planar images via orthographic projection to eliminate perspective discrepancies. Structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) are employed to quantitatively evaluate simulation fidelity.

Table 1 summarizes SEM imaging parameters and corresponding image-quality metrics for different viewing perspectives. All parameters satisfy requirements for subsequent image fusion and AI-driven simulation. SE signals are collected at 0˚ and 10˚. The 10˚ tilted perspective generates obvious pixel parallax carrying information of trench depth and longitudinal side-wall variation. Shadow artifacts on trench side walls are insignificant. BSE signals acquired at 0˚ are free of shadow occlusion and completely reveal trench-bottom morphology and internal residues. Complementary information from SE and BSE signals captures full-range spatial features of trenches accurately and guarantees reliability for image fusion and AI simulation.

Table 1. Imaging parameters and quality indicators of SEM images at different viewing angles.

Full-field morphological fused images are generated after sub-pixel registration and pixel-level weighted fusion. The outputs are merely 2D grayscale images without explicit 3D coordinates, yet they fully retain trench-depth information and side-wall gradient derived from parallax of SE 10˚ images, together with opening boundaries from SE 0˚ and bottom morphological features from BSE 0˚. No local spatial-information loss occurs. As inputs for AI models, fused images can equivalently substitute spatial-structure data obtained from 3D reconstruction and provide complete information sources for AI to learn imaging rules of trench three-dimensional cross-sections.

As shown in Table 2, the optimal input combination is SE (0˚ + 10˚) dual-view coupled with BSE (0˚). All three lightweight AI models achieve stable SSIM above 0.92 and reproduce three-dimensional features including trench side walls, bottoms and depth.

1) The lightweight U-Net model achieves optimal comprehensive performance with SSIM of 0.93. It balances simulation accuracy and inference speed and is suitable for on-line batch inspection.

2) The multi-view branch GAN obtains SSIM = 0.92 and exhibits prominent capability in restoring micro-defects, applicable for high-precision defect analysis scenarios.

3) The depth-prior auto encoder yields SSIM ranging from 0.92 to 0.93 and demands fewer training samples, suitable for off-line laboratory simulation.

Under single-view conditions, simulation performance of SE mode surpasses that of BSE mode. Without parallax input from the SE 10˚ tilted perspective, restoration accuracy for trench depth and longitudinal side-wall morphology drops significantly. It demonstrates that multi-view parallax information serves as the core foundation for AI to realize equivalent substitution for three-dimensional reconstruction.

Table 2. Comparison between simulated TEM images and real TEM images under different SEM-input combinations.

Preprocessed fused multi-view SEM images are fed into the well-trained lightweight U-Net model to generate simulated TEM images that reconstruct complete side-wall and bottom morphology. Quantitative comparison against real TEM data is summarized in Table 3. SE (0˚ + 10˚) dual-view inputs supply full-segment side-wall details, and BSE (0˚) provides bottom-contrast features. The fused images deliver complete trench-feature inputs for networks and realize excellent fine-detail restoration for simulated outputs. For the optimal SE (0˚ + 10˚) + BSE (0˚) combination, average SSIM reaches 0.93 (threshold = 0.9), PSNR is 34.8 dB. Simulated TEM images accurately reproduce real micro-morphology of trenches and supply reliable micro-data for copper-filling process optimization.

Table 3. Comparison between simulated and actual TEM images.

To unify comparison benchmarks and eliminate 3D perspective deviations, both real and simulated TEM outputs are converted into standardized 2D planar images via orthographic projection following conventional semiconductor-characterization standards for quantitative evaluation of structural similarity and detail-restoration performance. Under the optimal imaging combination, simulated and real planar TEM images show high contour overlap and complete matching of micro-features without obvious deviation among evaluation metrics, further verifying stability and accuracy of the proposed simulation method. Benefiting from this non-destructive high-efficiency inspection approach, cumbersome TEM sample-preparation procedures can be skipped. Micro-quality of post-etching trenches can be rapidly judged, and inspection data can be fed back to production lines in real time to realize dynamic adjustment of process parameters. It provides powerful technical support for closed-loop quality control in continuous semiconductor copper-interconnect manufacturing, improves process efficiency substantially and reduces wafer scrapping costs.