GenISys lets 20 ship subsystems denoise at once: 9% more low-cost, constraint-satisfying bulk-carrier concepts
Synopsis
GenISys splits early-stage bulk-carrier concept design into 20 subsystems and uses a simultaneous conditional denoising diffusion model so each subsystem reads the others' intermediate states at every reverse step, instead of generating them in a fixed order. In a five-fit confirmation at a fixed 5,000-nautical-mile voyage range, the model returned designs that both satisfied the 187 formulated engineering constraints and met a common low-cost target in 55.70% of generation requests, versus 51.10% for the monolithic model and 47.54% for the sequential model, corresponding to 9.0% and 17.2% more qualifying designs with similar estimated transport costs.
Interpretation
A generative model can now propose a whole coupled bulk-carrier concept rather than only a hull shape, placing hull, structure, subdivision, powering and operating parameters in one evaluable design space. Earlier generative ship work concentrated on hull form, such as ShipHullGAN, ShipGen and C-ShipGen; integrated vessels still required repeated designer evaluation and revision to reconcile disciplines. The dataset holds 217,999 proposed designs, 16,247 of which satisfy the formulated engineering constraints; none of the 100,000 randomly combined hull, structural and operating designs satisfied them, and feasible examples came from an equation-informed repair procedure.
The simultaneous conditional denoising diffusion model (SCDDPM) assigns 133 controlled parameters to 20 subsystem denoisers, and each denoiser reads the noisy states of the other subsystems at every step, so later subsystems can influence earlier ones. The sequential HCDDPM completes each block's full reverse chain in a fixed order, so early errors become downstream conditions that cannot be revised; the monolithic DDPM uses one network for all parameters and has no subsystem interfaces. In the confirmation at a fixed 5,000-nautical-mile range with five fresh fits per architecture, SCDDPM averaged a 55.70% qualifying yield, DDPM 51.10% and HCDDPM 47.54%; 29,970 of 30,000 requests completed evaluation and 17,110 designs were fully feasible.
The subsystem partition also yields two operating interfaces: feasibility guidance can be applied to one subsystem's parameters, or the remaining parameters can be locked while a chosen subsystem is resampled. A monolithic model can only guide the whole ship and cannot confine a change to one subsystem. On 100 starting designs that failed only the hopper-structure checks, guidance raised fully feasible child yield from 10.6% to 17.7% for DDPM, 17.4% to 25.3% for HCDDPM and 15.4% to 23.9% for SCDDPM.
Perspective
The demonstrated scope is early-stage bulk-carrier concept design with 133 controlled parameters, 187 rule-based engineering constraints and five performance objectives, evaluated with Holtrop–Mennen resistance estimation, a representative three-hold structural model and economic coefficients inherited from an earlier benchmark. Satisfying these constraints establishes feasibility within the modeled design problem only; full-vessel stability, structural strength and cost still require detailed engineering assessment.
The authors note that the three architectures were compared as complete generation pipelines whose model sizes, training effort and guidance configurations differ, so the observed differences cannot be attributed entirely to simultaneous conditioning. Training data come from a synthetic population whose feasible examples were produced by the repair procedure and may not cover solutions found in other design settings; generalization to other representations and engineering systems remains to be tested. At the more permissive training-median cost target, HCDDPM instead has the highest qualifying yield (60.10%).