SABER-BIM: A Component-Level Adaptive Lightweighting Framework for Digital Twin BIM Models

Lightweighting Building Information Modeling (BIM) models for digital-twin applications requires balancing aggressive geometric reduction with component-level engineering tolerances and mesh usability. Most geometric simplification pipelines apply uniform ratios or hand-tuned heuristics, which struggle to accommodate the strong heterogeneity of BIM components in functional role, geometric complexity, and detail distribution. End-to-end learning-based simplification can be adaptive, but it often entangles decision-making with geometric editing, making engineering constraints difficult to enforce and audit. We present Semantic-Geometric Co-driven Adaptive Budget Estimation and Reduction for BIM (SABER-BIM), which formulates lightweighting as a component-level face-budget allocation problem. Conditioned on Industry Foundation Classes (IFC) types and structure-sensitive geometric descriptors, SABER-BIM predicts target face counts for individual components and then meets a user-specified global budget through global scaling. The predicted budgets are executed by a robust geometric backend (e.g., quadric error metrics, QEM), yielding an auditable and easily deployable pipeline. To address the absence of direct supervision, we introduce an offline pseudo-ground-truth procedure that searches for the minimum feasible target face count for each component under semantic-aware tolerance and mesh-validity constraints. Experiments on the IFCNet dataset show that SABER-BIM allocates budgets more effectively under identical global constraints, improving stability in both geometric error control and engineering usability.