This work compares cross‑anatomy transfer with sparse interpolation for surrogate models of aortic fluid‑structure interaction (FSI) that are intended for digital‑twin applications. Four de‑identified human aortic geometries were retrieved from the Vascular Model Repository, reconstructed into separate lumen and 1.5 mm wall domains, and simulated under matched first‑cycle two‑way FSI conditions.
A geometry‑only LightGBM prior was trained using a leave‑one‑anatomy‑out (LOAO) scheme on three anatomies and then evaluated zero‑shot on the fourth. The zero‑shot transfer performed poorly across all six target fields.
A sparse field‑completion study was then conducted with 5 % anchor points (203 anchors, 3 852 evaluation nodes). After adding the prior and adapting, the oscillatory shear index (OSI) achieved $R^2 = 0.603$. Direct interpolation on the same anchors, tuned only on the three development anatomies, yielded substantially higher scores:
- Inverse‑distance weighting (IDW): $R^2_{\text{OSI}}=0.829$, $R^2_{\text{peak von Mises}}=0.617$, $R^2_{\text{mean stress}}=0.676$
- Radial basis function (RBF) interpolation: $R^2_{\text{OSI}}=0.917$, $R^2_{\text{peak von Mises}}=0.714$, $R^2_{\text{mean stress}}=0.778$
Thus, sparse labels within the same anatomy effectively support field completion, and this four‑anatomy cohort provides no evidence that a cross‑anatomy prior adds value beyond direct interpolation.
We frame the study as the first computational stage toward a measurement‑linked digital twin: the surrogate/update layer is evaluated here, while larger cohorts, converged FSI simulations, measurable patient‑side inputs, and physics‑informed learning remain future work, not a claim of a complete clinical twin.
All code, data, and computational files are released at https://github.com/ali-nourbakhsh2005/Aortic-FSI-Sparse-Field-Completion
Review: In this scenario the cross‑anatomy prior did not outperform conventional sparse interpolation, suggesting that early digital‑twin development should prioritize high‑quality intra‑anatomy data and robust interpolation techniques rather than relying on generic cross‑anatomy machine‑learning models.