MiDShip: Multimodal Dataset of Ship Cargo Hold Structures for Engineering Design

Noah J. Bagazinski; Md. Ferdous Alam; Jaya Manideep Rebbagondla; Faez Ahmed

cs.CESep 4, 2026

MiDShip: Multimodal Dataset of Ship Cargo Hold Structures for Engineering Design

Authors: Noah J. Bagazinski, Md. Ferdous Alam, Jaya Manideep Rebbagondla, Faez Ahmed

Organizations: Department of Mechanical Engineering, Massachusetts Institute of Technology Cambridge, MA 02139 · George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology Atlanta, GA 30318

Abstract

Ship structures govern vessel strength, safety, and manufacturability, but their design must satisfy hundreds of classification society requirements, making the process complex and iterative. Data-driven approaches are limited by the lack of structured datasets linking design geometry, structural performance, and rule-based constraints. This paper presents MiDShip, a multimodal dataset of 12,753 synthetic cargo-hold structural designs: 6,020 random, 496 generated by an SGLD-inspired procedure, and 6,237 generated by an equation-informed repair procedure. Each design includes parametric data, full and mesh-ready 3D geometry, engineering drawings and annotations, a bill of materials, and preliminary structural evaluations. Twenty-five constraints derived from a subset of ABS MVR are also evaluated. None of the random designs satisfies all constraints. Among the SGLD-inspired designs, 322 (64.9%) were fully compliant, with an average of 0.409 violations, 82.7% below the seed mean and 96.9% below the random-design mean. The repair procedure, developed through LLM-assisted code analysis, produced 4,952 fully compliant designs (79.4%), averaging 0.296 violations, 97.1% below the paired-source mean. In equal-size comparisons, mean nearest-neighbor distances in the scaled 120-parameter space were 3.495 for repaired designs, 1.144 for SGLD batches, and 3.729 for random designs. The primary contribution is the synchronized dataset and its generation and evaluation infrastructure; the generation studies demonstrate its utility rather than proposing new optimization algorithms. MiDShip supports machine learning, generative design, and automated rule-based evaluation for ship structures.

Figures & tables

Figure 1: Overview of the MiDShip dataset, a multimodal dataset for ship cargo-hold structural design and evaluation. Each design is represented by a parametric design vector used to generate full 3D and mesh-ready geometries. Corresponding 2D engineering drawings are produced alongside labels for structural density, longitudinal cross-sectional area, center of gravity, maximum bending moment, and twenty-five local scantling evaluations derived from ABS design rules.

Figure 2: Each ship structure is defined by a set of 120 parameters grouped into three categories: principal characteristics and ship class (10 parameters), shell and plate dimensions (27 parameters), and stiffener/girder spacing and sizing (83 parameters). There are twenty unique stiffeners characterized by the parametric designs.

Figure 3: Two-dimensional t-SNE visualizations of the design distributions, colored by ship class and the presence of longitudinal bulkheads. Panel (a) shows the 6,020 random designs; panel (b) overlays all 500 attempted SGLD candidate vectors; panel (c) overlays a 100-design sample of repaired candidate vectors; and panel (d) overlays all 6,254 attempted repair candidate vectors. The projection provides a qualitative view of the jointly embedded parameter vectors and is not used to quantify global design-space coverage or engineering diversity.

Figure 4: Examples of dataset designs. The leftmost column shows two containership designs; the two middle columns show tanker designs with varying longitudinal bulkheads and hatch openings; and the rightmost column shows two bulk-carrier designs.

Figure 5: Example mesh-ready geometry. This representation is identical to the full 3D geometry except that smaller stiffeners are replaced with curves. The structural properties of each stiffener are stored in a list to support future finite-element meshing.

Figure 6: Example 2D engineering drawing of the longitudinal structure. The drawing contains a slice of the ship structure, a title block, two information blocks, and formatted margins.

Figure 7: The 2D drawing in Figure 6 with labeled bounding boxes around structural elements. The positions, dimensions, and unique identifiers of the structural elements and bounding boxes are stored in a separate metadata file.

Figure 8: Distributions of structural properties for the 6,020 retained random designs. The plots show the volume-normalized structural steel mass, longitudinal cross-sectional area, and calculated maximum bending moment released with the dataset.

Figure 9: Examples of random designs with high calculated maximum bending moments (top) and low structural densities (bottom).

Population	Designs	Feasibility rate ( $\uparrow$ )	Mean CV count ( $\downarrow$ )	MNN distance
Random designs	6,020	0.000	13.192	$3.729\pm 0.052$
SGLD seed pool	25	0.000	2.360	–
SGLD-generated designs	496	0.649	0.409	$1.144\pm 0.104$
Repair source designs	3,127	0.000	10.288	–
Repaired designs	6,237	$\mathbf{0.794}$	$\mathbf{0.296}$	$3.495\pm 0.133$

Table 1: Comparison of feasibility and constraint-violation statistics for the complete dataset subsets and the method-specific starting populations. The SGLD seed pool and repair source designs are drawn from the random subset and are not additional dataset instances. Feasibility is the fraction of designs satisfying all twenty-five implemented constraints, and the mean constraint-violation (CV) count is the total number of violations divided by the number of designs. Candidate parameter sets that failed CAD generation are excluded. MNN values are means

\pm 2s

across equal-size 100-design sets, where

s

is the sample standard deviation. The random and repaired results use thirty sets sampled without replacement within each set, with possible overlap across repeated draws; the SGLD result uses the five actual 100-candidate generation batches. MNN was not calculated for the method-specific starting populations. The MNN ranges describe between-set dispersion and are not confidence intervals. Full violation-count distributions are reported in Appendix Table 8 . Arrows indicate the preferred direction for feasibility and CV count; MNN is reported descriptively, and boldface denotes the best feasibility and CV values.

Constraint	Random satisfaction rate (%)	Validation $R^{2}$	Validation $NRMSE$ (%)
Double-bottom height	83.3	0.931	6.3
Bottom-floor thickness	71.9	0.977	5.0
Bottom-girder thickness	13.6	0.967	5.1
Bottom-floor spacing	73.3	0.972	5.4
Inner-bottom deck thickness	8.9	0.891	8.3
Hopper-plate thickness	44.2	0.912	7.5

Table 2: Per-constraint satisfaction rates in the retained random subset and constraint-surrogate validation performance. Random satisfaction is the percentage of the 6,020 retained random designs that satisfy each constraint. The

R^{2}

and normalized root mean squared error (

NRMSE

) values use all 1,204 designs in the class-stratified validation set and are calculated in logistic-transformed residual space. SM denotes section modulus.

Constraint	Number of designs	Frequency among evaluated repaired designs (%)
Hopper stiffener section modulus	745	11.94%
Transverse-bulkhead thickness	505	8.10%
Horizontal transverse-bulkhead stiffener section modulus	340	5.45%
Vertical transverse-bulkhead stiffener section modulus	182	2.92%
Deck-plate thickness	37	0.59%
Hopper-plate thickness	25	0.40%

Table 3: Constraints remaining violated after repair and evaluation using the implemented constraint calculations. Percentages use the 6,237 successfully evaluated repaired designs as the denominator.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Index	Name	Units	LL	UL
0	L_3h	m	40	80
1	B	m	15	80
2	T	frac. D	0.25	0.667
3	D	m	10	40
4	R_b	frac. Db	0.75	1
5	l_overhang	frac. L_3h	0.03	0.05

Appendix

Table 4: The 120 MiDShip parameter fields and their nominal random-sampling specifications, including units, lower limits (LL), and upper limits (UL).

Index	Name	Units	LL	UL
40	blkhd trans stiff h	mm	120	320
41	blkhd trans stiff t	mm	7	12
42	blkhd trans stiff w	mm	14	200
43	blkhd trans stiff torc	bool	0	1
44	blkhd vert stiff h	mm	120	320
45	blkhd vert stiff t	mm	7	12

Appendix

Table 5: The 120 MiDShip parameter fields and their nominal random-sampling specifications, including units, lower limits (LL), and upper limits (UL) (continued).

Index	Name	Units	LL	UL
80	bulk car bottom hopper t	mm	15	35
81	bulk car top hopper t	mm	15	35
82	num bulk car bottom hopper stiff	int	3	6
83	bulk car bottom hopper stiff h	mm	120	320
84	bulk car bottom hopper stiff t	mm	7	12
85	bulk car bottom hopper stiff w	mm	14	200

Appendix

Table 6: The 120 MiDShip parameter fields and their nominal random-sampling specifications, including units, lower limits (LL), and upper limits (UL) (continued).

Requirement Number	Structural Constraint
ABS Part 3, Ch. 2, Sec. 4/3.1	Minimum Double Bottom Height
ABS Part 3, Ch. 2, Sec. 4/1.2	Minimum Double Bottom Height
ABS Part 3, Ch. 2, Sec. 4/5.1	Minimum Bottom Floor Thickness
ABS Part 3, Ch. 2, Sec. 4/5.1	Minimum Bottom Girder Thickness
ABS Part 3, Ch. 2, Sec. 4/5.1	Maximum Bottom Floor Spacing
ABS Part 3, Ch. 2, Sec. 4/9.2	Minimum Inner Bottom Deck Thickness

Appendix

Table 7: ABS rule-to-output applications used to formulate twenty-five constraint outputs for selected spacing, sizing, and stiffness requirements. For outputs governed by multiple applicable minimum requirements, the greatest required value is used.

Violation count	SGLD seeds ( $n=25$ )	Retained SGLD designs ( $n=496$ )	Repair sources ( $n=3{,}127$ )	Retained repaired designs ( $n=6{,}237$ )
0	0	322	0	4,952
1	3	145	3	914
2	10	29	10	206
3	12	0	19	142
4	0	0	44	23
5	0	0	103	0

Appendix

Table 8: Violation-count distributions for the method-specific source populations and retained generated designs. All counts use the implemented evaluator underlying the reported results. The SGLD comparison is population-level because candidates were initialized from interpolations of seed vectors. The repair-source column contains the 3,127 unique source designs; the paired source mean reported in the text is weighted over the 6,237 retained source–repair pairs and therefore differs slightly from the unique-source mean shown here.

Explore similar work

Oct 7, 2026cs.AI

NL2Hull: A Natural Language-Driven Constrained Ship Design Decision Framework

Ship-form design combines smooth geometric representation, local shape editing, and constraints on the resulting hull. We present the Natural-Language-to-Hull Framework (NL2Hull Framework), which formulates ship-form editing as a typed discrete decision problem and connects language decisions to numerical geometry. Its Constrained Free-Form Deformation Engine (CFFD Engine) represents hull waterlines with non-uniform rational B-splines (NURBS), applies free-form deformation (FFD) to their control points, reconstructs the hull, and checks geometric constraints. We construct the Ship Design Decision Dataset (SDD Dataset) with 134,558 cleaned records and evaluate compared models on its subset Ship Design Decision Benchmark (SDDBench), containing 5,000 records and 43,496 typed questions. We propose Chip, a constrained ship-design decision model for processing natural-language requests. Chip reaches 95.90% question accuracy and 99.32% FFD exact match, with a negative log-likelihood of 0.0951, an expected calibration error of 0.0032, and a Brier score of 0.0551. The NL2Hull Framework provides a reproducible interface for evaluating language-based ship-form decisions while identifying the geometry and continuous-control components that require further development. Our code and dataset is available at https://github.com/wenhuahuo/NL2Hull.

Wenhua Huo, Fenglei Han, Wangyuan Zhao +2

Harbin Engineering University, 145 Nantong Street, Nangang District, Harbin, P. R. China

Shapes Constraint Language

Ask PDF

Apr 30, 2026cs.CE

Design Structure Matrix Modularization with Large Language Models

Design Structure Matrix (DSM) modularization, the task of partitioning system elements into cohesive modules, is a fundamental combinatorial challenge in engineering design. Traditional methods treat modularization as a pure graph optimization, without access to the engineering context embedded in the system. Building on prior work on LLM-based combinatorial optimization for DSM sequencing, this paper extends the method to modularization across five cases and three backbone LLMs. Our method achieves near-reference quality within 30 iterations without requiring specialized optimization code. Counterintuitively, domain knowledge, beneficial in sequencing, consistently impairs performance on more complex DSMs. We attribute this to semantic misalignment between the LLM's functional priors and the purely structural optimization objective, and propose the semantic-alignment hypothesis as a testable condition governing knowledge effectiveness with LLMs. Ablation studies identify the most effective input representation, objective formulation, and solution pool design for practical deployment. These findings offer practical guidance for deploying LLMs in engineering design optimization.

Shuo Jiang, Jianxi Luo

Department of Systems Engineering City University of Hong Kong, Hong Kong

Design Structure Matrices Modular Framework Combinatorial Optimization

Ask PDF

Jan 30, 2026cs.AI

Retrieval Augmented (Knowledge Graph), and Large Language Model-Driven Design Structure Matrix (DSM) Generation of Cyber-Physical Systems

We explore the potential of Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), and Graph-based RAG (GraphRAG) for generating Design Structure Matrices (DSMs). We test these methods on two distinct use cases--a power screwdriver and a CubeSat with known architectural references--evaluating their performance on two key tasks: determining relationships between predefined components, and the more complex challenge of identifying components and their subsequent relationships. We measure the performance by assessing each element of the DSM and overall architecture. Despite design and computational challenges, we identify opportunities for automated DSM generation, with all code publicly available for reproducibility and further feedback from the domain experts.

H. Sinan Bank, Daniel R. Herber

Department of Systems Engineering, Colorado State University, Fort Collins, CO 80523, USA

Design Structure Matrices Engineering Graph-Based Retrieval Augmented Generation Knowledge Graphs Cyber-Physical System

Ask PDF