Decoupling geometry-dependent dimensional errors in FDM assembly elements via a learning-based inverse design approach

3D printing is increasingly used for functional assembly parts, yet its dimensional accuracy remains a bottleneck because deviations arise from multiple superimposed sources. This study addresses whether the digital (slicer-induced) and physical components of diameter error can be decoupled and then compensated using only standard slicer settings. A full factorial experiment varying horizontal expansion, feature type (pin/hole), first printed line strategy, infill, and nominal size was combined with a custom G-code toolpath analysis, multi-factor analysis of variance, and five supervised learning models (Artificial Neural Networks, Support Vector Regression, Random Forest, Extreme Gradient Boosting, and Polynomial Regression). G-code analysis showed that slicer offsets were fully deterministic, while measurements revealed a geometry-dependent asymmetry. Pins closely followed the theoretical model, whereas holes exhibited a systematic negative offset from thermal shrinkage. The outer-first strategy reduced hole error by ~ 11% but degraded pin accuracy, motivating a feature-adaptive approach. A second-order polynomial regression achieved the lowest error (RMSE = 0.0522 mm), with nominal size found unnecessary as an input within the tested 8–12 mm range. The resulting inverse-design tool recommends feature-specific settings, enabling assembly-grade accuracy within standard slicing workflows without CAD-level geometry modification. Verified on two architecturally distinct printers (Creality Ender-6 and Snapmaker J1s), the recommended feature-specific settings achieved mean absolute diameter errors of 0.02–0.04 mm for both features, indicating low sensitivity to machine-specific hardware.