PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.02214v1cs.CEcs.LG

TL;DR

Ship-structure machine learning is constrained by scarce datasets linking design geometry, performance, and rule-based constraints. MiDShip provides a multimodal cargo-hold dataset and evaluation infrastructure, with generation studies demonstrating constraint-feasible designs. The dataset is the primary contribution, while the reported constraint results apply only to the implemented research constraints.

  • Problem

    Public ship-structure benchmarks synchronizing geometry, drawings, performance information, and rule-based constraints remain scarce, limiting reproducibility, method comparison, and adoption.

    Source: §1 · p. 2

  • Method

    MiDShip is a multimodal benchmark of cargo-hold designs with synchronized parametric, CAD, drawing, materials, structural-property, and rule-evaluation data.

    Source: §5 · p. 16

  • Results

    4,952 (79.4%) repaired designs and 322 (64.9%) SGLD-inspired designs satisfied all twenty-five constraints.

    Source: §5 · pp. 16–17

  • Takeaways & Limitations

    MiDShip supports research linking parametric representations, CAD geometry, structural properties, drawings, and rule-based evaluations.

    Source: §5 · p. 17

    The twenty-five constraints are limited rule-derived evaluations, not complete ABS-vetted approval; satisfaction does not establish construction readiness or regulatory approval.

    Source: §5.1 · p. 17

Abstract

from arXiv · show

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.

1 Introduction

Cargo-hold structural design affects vessel performance and must meet classification-society rules. MiDShip addresses the shortage of structured ship-structure data with a multimodal dataset, evaluation tools, and proof-of-concept design-generation studies.

1 Introduction

  • Ship structural layout influences cargo-space arrangement, strength, manufacturability, lifecycle cost, and operational efficiency, with consequences for steel weight, capacity, fuel use, and emissions.The paper focuses on the parallel midbody, the constant-cross-section region central to longitudinal-strength evaluation.Source: §1 · p. 2
  • Public benchmarks synchronizing ship geometry, drawings, performance information, and classification-rule labels remain scarce, constraining reproducibility, method comparison, and adoption.Useful benchmarks should also support common engineering software and scripted analysis.Source: §1 · p. 2
  • MiDShip contains 12,753 cargo-hold designs across random, SGLD-inspired, and equation-informed repair subsets, with parametric, geometric, drawing, materials, and evaluation data.Each retained design includes full and mesh-ready 3D geometry, three engineering drawings with annotations, a bill of materials, twenty-five ABS-derived constraint evaluations, and structural performance metrics.Source: §1 · p. 2
  • The work releases a public multimodal dataset synchronizing parametric, geometric, drawing, material, performance, and rule-evaluation representations.The dataset license is shown as <LICENSE> in the passage.Source: §1 · p. 3
  • An accompanying parametric CAD and rule-based evaluation pipeline supports generating, modifying, and evaluating cargo-hold structures.The codebase uses Python scripting tools for Rhinoceros 3D.Source: §1 · p. 3
  • The paper preliminarily demonstrates surrogate learning of relationships among design parameters, structural properties, and twenty-five implemented constraints.Source: §1 · p. 3
  • Two demonstrations use MiDShip for constraint-aware design generation: surrogate-guided SGLD-inspired sampling and equation-informed parametric repair.Source: §1 · p. 3

2 Prior Work

Prior work spans structural evaluation, engineering datasets, AI for ship design, and drawing interpretation. Across these areas, the paper identifies a shortage of reusable ship-structure benchmarks synchronizing design modalities and rule evaluations.

2.1 Ship Structure Evaluation

  • Prior ship-structure work includes finite-element evaluation, parameterized optimization, and constraint-aware repair, while MiDShip contributes a reusable dataset linking designs to multiple engineering representations.Earlier parametric modeling systems already supported rule-informed structural design; MiDShip’s stated distinction is its synchronized, machine-learning-oriented dataset.Source: §2.1 · p. 3

2.2 Engineering Design Datasets

  • Engineering-design datasets pair structured designs with performance labels; MiDShip applies this paradigm to internal cargo-hold structures with rule evaluations and synchronized CAD, drawings, and material data.Ship-D focuses on external hull forms, while BIKED covers parametric bicycle designs and related representations.Source: §2.2 · p. 4

2.3 Artificial Intelligence for Ship Design

  • AI ship-design research has largely emphasized hull-form optimization, with comparatively limited attention to structural design and its data-driven methods.Prior approaches also use surrogate-guided Langevin sampling for multi-constraint inverse design, a concept applied here to ship structures.Source: §2.3 · p. 4

2.4 Drawing Evaluation and Design Verification with Artificial Intelligence

  • Engineering-drawing interpretation methods increasingly extract geometry, text, and semantic information, but ship-structure workflows lack reusable data pairing drawings with designs and performance metrics.MiDShip provides paired drawings, annotations, and CAD representations.Source: §2.4 · p. 5

2.5 Research Gap and Project Motivation

  • The authors identify MiDShip as, to their knowledge, the first public machine-learning-oriented benchmark synchronizing cargo-hold designs, CAD, drawings, bills of materials, structural properties, and rule evaluations.This claim is based on literature reviewed through August 2026.Source: §2.5 · p. 5

3 Dataset Details

MiDShip provides synthetic cargo-hold designs across three ship classes, with synchronized parametric, geometric, drawing, material, performance, and rule-evaluation representations. The section describes how these designs and selected structural evaluations are generated and what their scope does—and does not—cover.

3 Dataset Details

  • Each retained design includes parametric data, full and mesh-ready 3D geometry, drawings, a bill of materials, structural-property labels, and constraint evaluations.The 12,753 designs comprise 6,020 random, 496 SGLD-generated, and 6,237 repaired designs; the latter two subsets derive from selected random seeds.Source: §3 · p. 6

3.1 Parametric Design

  • 120 parameter fields encode principal characteristics and ship class (10), plate dimensions (27), and girder and stiffener dimensions and spacing (83).The fields represent global geometry and local structural details in a consistent tabular format for optimization and machine learning.Source: §3.1 · p. 6, Figure 2 · p. 6

3.2 Geometry Development

  • Procedural geometry assigns each element a unique identifier and records its position, orientation, dimensions, areas, and area moments of inertia.Mesh-ready files retain major members in 3D while representing smaller stiffeners as curves; no finite-element meshes or FEA results are included.Source: §3.2 · p. 8

3.3 Engineering Drawing Development

  • Each design has three engineering drawings and three bounding-box drawings, with metadata linking structural elements to their positions, dimensions, scales, and identifiers.The drawings represent longitudinal, transverse-bulkhead, and main webframe-beam-floor profiles; the procedure was adapted from an automated drawing pipeline.Source: §3.3 · p. 8

3.4 Performance Metrics

  • The maximum bending moment is a yield-based elastic proxy, not a wave-loading demand, ultimate-strength calculation, or regulatory allowable moment.It assumes 250 MPa steel yield strength and a user-selected safety factor of two, and omits buckling, nonlinear response, fatigue, and other failure modes.Source: §3.4 · pp. 9–10

3.5 Design Rule-Based Constraints

  • The 25 selected ABS-derived constraints assess scantling requirements, but exclude evaluations requiring full-vessel information and do not establish complete ship approval.The evaluator aggregates 38 rule-to-output applications from 22 clauses; a design is feasible only with respect to these implemented constraints.Source: §3.5 · pp. 10–11
  • 6,020 random designs yielded no design satisfying all 25 implemented constraints; the mean design violated 13.192.The constraints are coupled, so individual satisfaction rates cannot be multiplied to estimate the joint pass rate.Source: §3.5 · p. 12

4 Constraint-Aware Design Generation

The section evaluates two proof-of-concept approaches for guiding designs toward constraint-feasible regions; neither is presented as a novel optimization or repair algorithm. It demonstrates uses of MiDShip’s geometry-generation and constraint-evaluation pipeline.

4.1 Surrogate-Guided Inverse Design of Ship Structures

  • SGLD-inspired updates penalized predicted constraint residuals while a secondary term favored lower structural density and higher maximum bending moment.The procedure used surrogate gradients and stochastic noise; it did not claim Bayesian-posterior sampling or conventional SGLD convergence guarantees.Source: §4.1 · pp. 12–13
  • 322 of 496 SGLD designs were fully feasible (64.9%), averaging 0.409 violations, 82.7% below the seed mean and 96.9% below the random-design mean.The candidates started from interpolated near-feasible seeds, so the seed comparison is population-level rather than paired.Source: §4.1 · p. 14

4.2 LLM-Assisted, Constraint-Informed Parametric Repair

  • LLM-assisted code analysis mapped constraints to relevant parameters, informing rule-based repairs; the LLM was not used during design generation or evaluation.Only parameters associated with violated constraints were eligible for modification, while several hull, ship-type, and hatch-opening parameters remained fixed.Source: §4.2 · pp. 14–15

4.3 Sampling and Evaluation of Repaired Designs

  • Repair targets exceeded thresholds by 10%-35% for depth and thickness, 35%-60% for section modulus, and 20%-30% for spacing constraints.The study sought a feasible derivative subset using explicit equations, not out-of-sample generalization of a learned repair model.Source: §4.3 · p. 15
  • 4,952 of 6,237 repaired designs were fully feasible (79.4%), averaging 0.296 violations—a 97.1% paired reduction from their source designs.Every successfully evaluated repaired design had fewer violations than its source; 17 candidates were excluded after repeated CAD-generation errors.Source: §4.3 · p. 16
  • Mean nearest-neighbor spacing was 3.495 for repaired sets, 1.144 for SGLD batches, and 3.729 for random sets under the equal-size-set protocol.The metric does not by itself establish global design-space coverage or engineering realism.Source: §4.3 · p. 16
  • Residual repair failures reflected dependencies that were not fully represented, including CAD-adjusted stiffener spacing and ship-class-dependent requirements for hopper stiffeners.Hopper-stiffener section modulus was the most common remaining failure, occurring in 745 designs.Source: §4.3 · p. 16

5 Conclusion

MiDShip provides a synchronized benchmark and evaluation pipeline for cargo-hold structures; proof-of-concept generation studies produced many designs satisfying the implemented constraints, within stated dataset and methodological limits.

5 Conclusion

  • MiDShip synchronizes parametric designs, CAD and mesh geometries, drawings, bills of materials, structural properties, and rule evaluations across 12,753 cargo-hold designs.The authors identify the dataset and its reproducible geometry-generation and evaluation pipeline—not either generation method—as the primary contribution.Source: §5 · pp. 16–17
  • 322 SGLD designs (64.9%) were fully feasible and averaged 0.409 violations; SGLD-batch mean MNN was 1.144 versus 3.729 for random-design sets.The SGLD mean violation count was 82.7% below the seed mean and 96.9% below the random-subset mean.Source: §5 · p. 16, Table 2 · p. 17
  • 4,952 repaired designs (79.4%) were fully feasible, averaging 0.296 violations, 97.1% below the paired-source mean; repaired-design mean MNN was 3.495.Under the equal-size-set protocol, repaired-design mean MNN was 3.495, compared with 1.144 for SGLD batches and 3.729 for random-design sets.Source: §5 · p. 17, Table 2 · p. 17

5.1 Dataset Limitations

  • The synthetic cargo-hold regions omit global arrangements and vessel-level loading, and their prescribed ranges should not be used to infer operational-vessel configuration prevalence.The designs also use simplified geometry and omit localized reinforcement and other complex forms.Source: §5.1 · p. 17
  • The 25 implemented constraints are limited scantling evaluations, not a complete or ABS-vetted approval process; satisfying them does not establish construction readiness or regulatory approval.Loading, global-strength, fatigue, buckling, and manufacturability evaluations remain outside the pipeline.Source: §5.1 · p. 17
  • Generation studies use selected starting designs and lack matched baselines, so they do not establish performance from arbitrary starts, superiority, sample efficiency, or state-of-the-art performance.The repair procedure primarily increases dimensions or changes spacing and does not optimize material weight or other structural performance objectives.Source: §5.1 · p. 18
  • The repair implementation depends on supplied code, model version, development context, and human-guided iteration, and its results do not establish a general LLM capability for ship-structure design.Missed CAD post-processing dependencies and ship-class-specific requirements contributed to remaining violations.Source: §5.1 · p. 18

5.2 Future Work

  • Future work targets automated meshing and finite-element analysis, more loading conditions and rules, joint feasibility-performance optimization, less favorable starts, and matched method baselines.Proposed extensions also include richer structural features, additional ship classes, and multimodal tasks linking parameters, 3D geometry, and drawings.Source: §5.2 · p. 18

Datasheet for the MiDShip dataset

The datasheet describes MiDShip’s purpose, synthetic cargo-hold instances, multimodal contents, generation and preprocessing, uses, and distribution and maintenance arrangements.

Motivation

  • MiDShip was created as a publicly released, machine-learning-oriented benchmark for cargo-hold structural design and evaluation.Source: Motivation · p. 19

Composition

  • Each design includes a parametric vector, 3D and mesh-ready geometries, structural-element properties, performance and constraint evaluations, and three drawings with annotated versions.The three engineering drawings and their labeled versions total six drawing files per design.Source: Composition · p. 19

Data Collection Process

  • The 6,020 random designs were generated from prescribed input ranges with rounding, coupling, and ship-class activation rules; SGLD and repaired subsets derive from selected random designs.The three subsets are related, not statistically independent samples.Source: Data Collection Process · p. 19

Preprocessing, Cleaning, and Labeling

Uses

  • The dataset is suitable for training ship-structure evaluation models; the paper also reports a trained constraint surrogate and prediction models used in its SGLD demonstration.Source: Uses · p. 20

Distribution

  • The dataset and software are released under GNU GPL version 3.0 for noncommercial research and educational purposes.Source: Distribution · p. 20

Maintenance

  • The authors maintain the dataset through Hugging Face; users can submit corrections and updates through GitHub issues and pull requests.Source: Maintenance · p. 20

Appendix: Design Parameters

The appendix describes the MiDShip parameter fields and clarifies that nominal random-generation limits are not universal feasibility bounds.

Appendix: Design Parameters

  • MiDShip design vectors contain 120 parameter fields with nominal random-generation specifications.Source: Appendix: Design Parameters · p. 21
  • The fields encode dimensions, member properties, and ship-category indicators; the stored T field is unused by geometry generation and rule evaluation, which instead use Trule = 0.667D.The appendix defines h, t, and w as member height, thickness, and flange width, respectively.Source: Appendix: Design Parameters · p. 21
  • Nominal parameter limits are not universal feasibility bounds: coupled or class-specific sampling, post-sampling adjustments, and repair can produce values outside those ranges.Source: Appendix: Design Parameters · p. 21

Principal-Dimension and Ship-Class Coupling Rules

The generation rules couple principal dimensions and apply ship-class-specific parameter activation and geometry cleanup.

Principal-Dimension and Ship-Class Coupling Rules

  • The modeled three-hold length is sampled uniformly from 40 to 80 m, then temporary full-ship length is set to LOA = sL L3h, with sL sampled from 3 to 4.5.The temporary LOA is used during sampling and is not included in the released vector.Source: Principal-Dimension and Ship-Class… · p. 21
  • Overhang is sampled as 0.03–0.05 of L3h, cleanup enforces Rb ≤ Db/1000, and rule calculations requiring draft use Trule = 0.667D instead of stored T.The stored T field is retained but unused by the geometry generator and rule evaluator.Source: Principal-Dimension and Ship-Class… · p. 21
  • Ship class is sampled uniformly among tanker, containership, and bulk carrier; class-specific parameters are zeroed or activated, with bracket-leg ranges varying by class.Tankers sample relevant bracket legs from 500–1,000 mm; bulk carriers sample them from 3,000–5,000 mm.Source: Principal-Dimension and Ship-Class… · p. 21
  • Additional cleanup activates or zeros longitudinal-bulkhead and hatch-opening fields, removes inner-side-shell values below 10 mm, and rounds fields to intended types.Ship-class fields are projected to a single active one-hot category.Source: Principal-Dimension and Ship-Class… · pp. 21–22

Appendix: Source and Generated-Design Violation Distributions

Table 5 describes violation-count distributions for method-specific source populations and retained generated designs, with distinct comparison bases for SGLD and repair.

Appendix: Source and Generated-Design Violation Distributions

  • Table 5 compares violation-count distributions for method-specific source populations and retained generated designs using the implemented evaluator.The caption does not provide distribution values.Source: Table 5 · p. 24
  • The SGLD comparison is population-level because candidates were initialized from interpolations of seed vectors.Source: Table 5 · p. 24
  • The repair-source column covers 3,127 unique source designs, while the reported paired-source mean weights 6,237 retained pairs and differs slightly.Source: Table 5 · p. 24

Appendix: ABS Constraint Information

The appendix defines twenty-five evaluated ABS constraint outputs from selected rule applications and specifies how overlapping minimum requirements are consolidated.

Appendix: ABS Constraint Information

Loading 2610.02214v1…