This paper introduces a reward‑directed diffusion framework for localized ship‑hull optimization. During dataset generation, parametric masking restricts geometric changes to selected variables—such as the bulbous bow and forebody—while preserving full‑hull continuity.\
Optimization leverages surrogate‑assisted sequential Monte Carlo sampling, targeting a reduction in predicted calm‑water resistance. Stability and displacement constraints are enforced through the training data, ensuring physically plausible designs.\
An automated viscous free‑surface CFD pipeline, validated on two benchmark container ships, evaluates the generated hull variant. At the KRISO Container Ship (KCS) displacement, the variant matches the baseline total resistance at $F_n=0.195$ but exhibits a 45%–50% increase at $F_n=0.250$ and $0.260$, a degradation missed by empirical screening.\
In contrast, GenHull E—refined by a naval architect within the same parametric representation—achieves 6%–8% lower total resistance up to $F_n=0.240$, but its resistance rises above $F_n=0.250$ as friction and pressure components reverse their trends, offsetting the gains.\
Aggregating resistance across this crossover can conceal trade‑offs that become evident when friction and pressure are examined separately. The assessment reveals the model’s inability to represent a faired bulbous bow and a reward lacking viscous‑pressure and stability terms. These findings illustrate how high‑fidelity assessment guides improvements in generative representations and objectives for physically realizable designs.\
Review