GenISys: Integrated Ship Systems Design Using Subsystem-Conditional Tabular Diffussion
Organizations: Department of Mechanical Engineering Massachusetts Institute of Technology Cambridge, MA 02139
Abstract
Ship design requires coordinating coupled subsystems to satisfy engineering constraints and balance competing performance and cost objectives. Generative artificial intelligence has supported component design, but integrated ship generation must also account for interactions among hull geometry, structure, propulsion, and operation. This paper presents GenISys, a framework for generating integrated systems, demonstrated through early-stage bulk-carrier design. Its dataset contains 217,999 proposed designs, including 16,247 that satisfy the formulated engineering constraints. The proposed simultaneous conditional denoising diffusion probabilistic model (SCDDPM) generates subsystems together, allowing each subsystem model to respond to changes elsewhere during generation. Experiments compare this approach with monolithic whole-ship generation and sequential subsystem generation across five independent training runs per architecture. In a study targeting lower cost per tonne of cargo transported, designs were generated for a required voyage range of 5,000 nautical miles. SCDDPM produced designs satisfying the formulated engineering constraints and meeting a common low-cost target in 55.70% of generation requests, compared with 51.10% for the monolithic model and 47.54% for the sequential model. These correspond to 9.0% and 17.2% more qualifying designs, respectively, while maintaining similar estimated transport costs. These results show how generative artificial intelligence can coordinate subsystems to generate feasible, high-performing integrated systems.
Figures & tables
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Family | Outputs retained | Implementation and scientific basis |
|---|---|---|
| Hull hydrostatics | Waterline length, wetted area, displacement, buoyancy centers, block/midship/prismatic coefficients, waterplane area and moments, flotation center, perpendiculars and LBP | Hull generator and hydrostatic integrator. |
| Resistance | Reynolds number, friction/form, appendage, wave, bulb, allowance and total resistance; entrance angle and bulb area intermediates | Hull design module; Holtrop–Mennen [ 33 , 34 ] . |
| Power and voyage | Effective/installed power, minimum machinery-space length, days at sea, main and auxiliary fuel demand, reserve fuel | Hull design module; Equation ( 20 ) and explicit project scenario. |
| Environmental reference | Preliminary deadweight proxy, reference and required bulk-carrier EEDI proxies, phase and reduction factor | Powering constraints; regulatory reference only. |
| Mass budget | Steel, outfit, machinery, hatch, optional deckhouse, margin, miscellaneous deadweight, fuel, ballast, cargo, lightship, total displacement, and displacement margin | Weights and system-design modules; element areas and component accounting. |
| Cargo arrangement | Hold capacities, loaded volume, fill heights, capacity margins, cargo longitudinal center of gravity (LCG) bounds/target/achievement, full-cargo displacement and CG | Cargo-hold and system-design modules; compartment and allocation records. |
| All designs (217,999) | Fully feasible (16,247) | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Index | Class | Parameter | Units | Type | Status | Minimum | Maximum | Minimum | Maximum |
| 0 | Hull | LOA | m | float | Learned | 100.459 | 349.944 | 100.459 | 314.583 |
| 1 | Hull | Lb | – | float | Learned | 0.0750004 | 0.199993 | 0.0750566 | 0.199912 |
| 2 | Hull | Ls | – | float | Learned | 0.0500068 | 0.199997 | 0.0500184 | 0.199994 |
| 3 | Hull | Bd | – | float | Learned | 0.100436 | 0.18 | 0.101449 | 0.179967 |
| 4 | Hull | Dd | – | float | Learned | 0.0750002 | 0.12796 | 0.0750047 | 0.127794 |
| Order | Subsystem | Learned controls | Count | Local checks |
|---|---|---|---|---|
| 1 | Hull Principal | LOA; Lb; Ls; Bd; Dd; Bs; WL | 7 | C078–C084 (7) |
| 2 | Hull Midbody | Rc; Rk | 2 | C070–C077 (8) |
| 3 | Hull Bow | Bow_A; Bow_B; Bk; Kappa_Bow; Delta_Bow_A; Delta_Bow_B; Drift_A; Drift_B; Drift_C | 9 | C050–C063 (14) |
| 4 | Hull Stern | bit_EP_S; Trans_A; SK; Kappa_Stern; Delta_Stern_A; Delta_Stern_B; Beta_Trans; Bc_Trans; Rc_Trans; Rk_Trans | 10 | C085–C096 (12) |
| 5 | Hull Bulbous Bow | bit_BB; Lbb; Hbb; Bbb; Lbbm; Rbb | 6 | C064–C069 (6) |
| 6 | Hull Stern Bulb | bit_SB; Kappa_SB; Lsb; HSBOA; Hsb; Bsb; Lsbm; Rsb | 8 | C097–C103 (7) |
| ID | Assigned to | Class | Check | Recorded basis or source | Units |
|---|---|---|---|---|---|
| C001 | Bottom Shell and Double Bottom | Structure | Bottom-shell absolute minimum thickness | ABS MVR 3-2-2/3.15 | mm |
| C002 | Bottom Shell and Double Bottom | Structure | Double-bottom height - absolute minimum | ABS MVR 3-2-4/1.2.ii | mm |
| C003 | Bottom Shell and Double Bottom | Structure | Double-bottom height - breadth twentieth | ABS MVR 3-2-4/1.2.ii | mm |
| C004 | Bottom Shell and Double Bottom | Structure | Double-bottom height - draft and breadth equation | ABS MVR 3-2-4/3.1.1(c) | mm |
| C005 | Bottom Shell and Double Bottom | Structure | Inner bottom plating - cargo pressure | ABS MVR 3-2-4/9.2 | mm |
| C006 | Bottom Shell and Double Bottom | Structure | Inner bottom plating - excavator wheel loading | ABS MVR 3-2-4/9.7; 3-2-3/5.17 | mm |
| Study | Training and fits | Requests and locks | Guidance and reverse steps | Role and statistical unit |
|---|---|---|---|---|
| Feasibility-strength sweep | Feasible or completed training; one saved fit per architecture/population. | 46,000 requests; 46 conditions; all controls editable. | ; DDPM system + hull, HCDDPM local, SCDDPM local/system/combined. Guidance ; endpoint . | Feasibility-guidance comparison (Appendix E ); ten disjoint 100-request subsets per condition, conditional on saved fits. |
| Noise/guidance ablation | Feasible training; one saved fit per architecture. | 12,000 requests; four arms per architecture; all controls editable. | Ordinary/reduced noise crossed with absent/present guidance ( ). DDPM system + hull, HCDDPM local, SCDDPM system. ; endpoint . | Primary constraint analysis (Appendix E ); ten paired 100-request batches per arm; nine Holm-adjusted contrasts. |
| Locally feasible training | 22,192–119,143 passing rows per block; one new HCDDPM and one new SCDDPM, compared with saved globally feasible fits. | 18,000 new requests; 18 conditions; all controls editable. | Frozen feasible-only transforms/classifiers; and supported local/system/combined modes. ; endpoint . | Supporting training study (Appendix G ); ten disjoint 100-request subsets; 18 matched comparisons per outcome. |
| Free-range performance | Feasible training; one saved fit per architecture. Performance regression models use completed training rows. | 900 validation requests; nine arms; range, speed, and dimensions editable. Pilot and refinement cohorts separate. | ; DDPM system + hull, HCDDPM/SCDDPM local. Feasibility window ; performance (DDPM/SCDDPM) or (HCDDPM); endpoint . Full-state/prefix performance regression models (Appendix H ). | Independent sampling-cohort validation after pilots (Appendix H ); paired requests/descriptive response distributions, not training replication. |
| Fixed-range strength selection | Feasible training; one saved fit per architecture; performance regression models use completed training rows. | 36,000 requests; 36 cells; range locked at 1,000, 5,000, or 8,000 nmi. | , ; performance-study contexts. Guidance ; endpoint . | Select frozen settings for fresh-fit confirmation (Appendix J ); ten paired 100-request batches; 27 Holm-adjusted contrasts. |
| Cost Reduction-Guided Sampling Design Study | Feasible training; five fresh fits per architecture; guidance assets and transforms fixed. | 30,000 requests; two arms per fit; 5,000-nmi range locked. | ; versus DDPM/SCDDPM 0.10 or HCDDPM 0.03. Performance-study contexts. Guidance ; endpoint . | Primary repeatability study (Appendix I ); five complete fits per architecture; paired guidance and Welch architecture tests, Holm correction across five planned tests. |
| Architecture | Fit | Arm | Completed | Feasible | Qualifying count | Mean RFR (£/t) |
|---|---|---|---|---|---|---|
| DDPM | 1 | F | 1000 | 566 | 375 | 23.67 |
| DDPM | 1 | FR | 1000 | 529 | 527 | 17.63 |
| DDPM | 2 | F | 1000 | 547 | 322 | 24.75 |
| DDPM | 2 | FR | 1000 | 497 | 492 | 18.01 |
| DDPM | 3 | F | 1000 | 556 | 325 | 24.92 |
| DDPM | 3 | FR | 1000 | 525 | 520 | 18.12 |
| Architecture | Range (nmi) | Completed | Feasible (%) | RFR | Primary reference yield (%) | Median reference yield (%) | Upper-quartile reference yield (%) | |
|---|---|---|---|---|---|---|---|---|
| DDPM | 1,000 | 0.00 | 1000 | 56.9 | 9.18 | 40.2 | 55.6 | 56.8 |
| DDPM | 1,000 | 0.01 | 1000 | 57.3 | 8.88 | 44.2 | 56.7 | 57.3 |
| DDPM | 1,000 | 0.03 | 999 | 55.8 | 8.43 | 47.9 | 55.3 | 55.8 |
| DDPM | 1,000 | 0.10 | 1000 | 51.8 | 7.29 | 51.1 | 51.8 | 51.8 |
| DDPM | 5,000 | 0.00 | 999 | 52.2 | 24.07 | 33.7 | 49.1 | 52.0 |
| DDPM | 5,000 | 0.01 | 1000 | 54.1 | 23.20 | 37.9 | 52.3 | 54.0 |
| Architecture | Range (nmi) | Gain (pp) | 95% interval (pp) | ||
|---|---|---|---|---|---|
| DDPM | 1,000 | 0.01 | 4.0 | [2.18, 5.82] | 0.00151 |
| DDPM | 1,000 | 0.03 | 7.7 | [5.82, 9.58] | |
| DDPM | 1,000 | 0.10 | 10.9 | [8.60, 13.20] | |
| DDPM | 5,000 | 0.01 | 4.2 | [2.63, 5.77] | 0.000777 |
| DDPM | 5,000 | 0.03 | 11.4 | [9.06, 13.74] | |
| DDPM | 5,000 | 0.10 | 16.7 | [8.64, 24.76] | 0.00151 |
Constrained Generative Modeling
Constrained generative modeling develops methods that produce samples satisfying structural, physical, semantic, or user-defined requirements.
See all 211 papers