Reliability of Data-Driven Tensile-Property Prediction for Additively Manufactured Photopolymer Gyroid Lattices Under Material Domain Shift
Accurate average predictions for a few unseen resins do not establish whether a model can support individual material choices. This study evaluates the reliability of a published gyroid tensile-property predictor by aligning validation, residual calibration, and interval assessment with whole-material hold-outs. Using the existing experimental dataset, matched training counts reveal 2.59 times greater normalized error under material-grouped validation than under random specimen validation. Material-grouped calibration increases mean coverage of nominal 90% intervals from 62.2% to 87.2%, while exposing substantial failures for individual resins. A further analysis uses designed porosity and training-derived performance thresholds to assess screening before fabrication. Wider grouped intervals resolve fewer candidate decisions but reduce their observed error rates; many candidates remain undecided, and property-space distance misses some unreliable materials. Prediction errors in physical units increase when designed porosity replaces measured porosity, although some base-scaled errors decrease. The contribution is an empirical assessment of transfer reliability and screening resolution across the observed materials. It establishes which performance claims survive material-level evaluation and which require additional characterization and prospective tensile tests.
Authors
- Zhanhao Liang (ORCID: https://orcid.org/0009-0003-6382-3688)
- Jie Yang
- Yu Qin
- Almazbek Arzybaev
- Shenggui Chen
- Yuze Ma
Institutions
- Kyrgyz State Technical University (KG)
Publication Details
- Journal
- Micromachines
- Published
- 2026-11-16
- DOI
- https://doi.org/10.3390/mi17111330
- Primary Topic
- Cellular and Composite Structures
- Type
- article
- Field-Weighted Citation Impact
- 0.00