Appearance-preserving simplification of non-watertight CAD models using NeuS-guided outer shell extraction

Abstract Large-scale industrial computer-aided design (CAD) assemblies in shipbuilding and process plants routinely contain dense interior geometry that hinders real-time visualization and digital twin deployment, and tessellated CAD data exchanged through neutral formats often suffer from small inter-part gaps that render conventional outer shell extraction unreliable. Under these circumstances, visibility-based techniques such as ray casting can preserve the coordinate fidelity of original mesh faces without reconstruction-induced distortion, but are prone to leakage through micro-gaps, causing interior structures to be misclassified as exterior surfaces. Neural implicit surface reconstruction methods, such as neural implicit surfaces (NeuS), can recover watertight topology by closing such gaps, but often degrade geometric fidelity by over-smoothing thin components and sharp mechanical details. To address this trade-off, we propose an automated NeuS-guided outer shell extraction framework that combines geometry-faithful visibility analysis with watertight topological guidance. The proposed method introduces anisotropic cube normalization to improve NeuS reconstruction of high-aspect-ratio industrial structures, uses the NeuS reconstruction as a watertight guide surface rather than as the final output, and refines ray-casting candidates through 6-point asymmetric tolerance filtering to preserve exterior details while rejecting interior faces. Experiments on three non-watertight tessellated benchmark models, comprising one industrial ship assembly and two publicly available polygonal assets, demonstrate that the proposed method improves leakage-robust interior elimination while retaining the coordinate fidelity of the original mesh faces that remain in the extracted shell. On the primary ship benchmark, the method achieves an inner removal rate of 95.26%, an outer preservation rate of 98.21%, and an area-weighted harmonic score of 96.71%. On the car and aircraft benchmark models, area-weighted harmonic scores of 87.18% and 96.87% are obtained, respectively; the car case reveals a near-contact limitation in which improved interior removal is accompanied by reduced exterior preservation, which a model-specific tolerance adjustment partly mitigates. Additional qualitative tests on Helicopter, Submarine, and Train models without manual ground-truth labels further illustrate the applicability of the pipeline to other interior-containing tessellated assemblies.