PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Wenhua Huo, Fenglei Han, Wangyuan Zhao, Jialin Wu, Jiayi Han

arXiv:2610.09896v1cs.AI

TL;DR

Ship-form editing needs to connect a designer’s language to typed actions and geometric verification, beyond generating or optimizing hulls from structured inputs. NL2Hull combines Chip’s action decisions with a CFFD engine and evaluates them on SDDBench. Chip achieves 95.90% question accuracy and 99.32% FFD exact match, while end-to-end tests also expose constraint-handling limitations.

  • Problem

    Existing ship-design methods generate or optimize hull candidates from structured parameters, while an interactive system also needs to map language to localized operations and verify resulting geometry.

    Source: §1 · pp. 2–3

  • Method

    NL2Hull pairs Chip’s ordered discrete decisions for ship-form requests with a CFFD engine that executes actions, reconstructs hulls, and checks geometric constraints.

    Source: §1 · p. 3, §2.1 · p. 4

  • Results

    95.90% question accuracy and 99.32% FFD exact match were achieved by Chip on SDDBench.

    Source: §5 · p. 19

  • Takeaways & Limitations

    End-to-end tests verified execution of single edits and multi-turn edits on a KVLCC2 hull.

    Source: §5 · p. 19

    Unconstrained actions composed with constrained actions can alter final volume, and strict deck-line checks can retain small numerical residuals.

    Source: §5 · p. 19

Abstract

from arXiv · show

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.

1. Introduction

Prior hull-design and language-to-action work provide pieces of the problem, but ship-form editing still needs a link from a designer’s language to localized operations and geometric verification. NL2Hull connects typed decisions to executable FFD actions and evaluates them alongside the dataset and benchmark it introduces.

1. Introduction

  • Prior parametric hull-design work established editable geometric foundations, but representation alone does not specify intended edits or how edit sequences should be verified.The cited design line links geometric parameters to engineering requirements and practical workflows.Source: §1 · p. 2
  • Existing language-to-CAD and language-grounded action approaches motivate an interface that evaluates language understanding, typed action selection, and geometric execution as linked stages.The introduction distinguishes structured CAD sequences, executable skills, embodied action prediction, and typed decision models.Source: §1 · pp. 2–3
  • NL2Hull links natural-language requests to typed editing decisions, executable FFD actions, hull reconstruction, and geometric constraint checks.Chip predicts ordered action attributes and their probabilities; the CFFD Engine executes the actions on a separately supplied NURBS hull.Source: §1 · p. 3
  • The paper introduces the SDD Dataset and SDDBench, with a common protocol for question accuracy, probability quality, FFD exact match, and latency across Chip, JEV, and general language models.The contributions also state that Chip is validated on end-to-end editing.Source: §1 · p. 3

2. Methods

The methods connect natural-language ship-form requests to constrained geometric edits through a typed decision interface and a NURBS-based deformation engine. They also construct the SDD Dataset and SDDBench to train and evaluate these decisions.

2.1. CFFD Engine: NURBS-Based Ship Representation and FFD

  • The CFFD Engine deforms NURBS waterline control points through localized or global actions, then reconstructs the hull and checks geometric constraints.Actions use regions, operations, and six discrete magnitude levels; constraints include symmetry, deck-line preservation, displacement preservation, and valid waterline geometry.Source: §2.1 · pp. 3–5

2.2. NL2Hull Decision Interface

  • Chip maps requests and typed questions to probability distributions whose selected answers form ordered region–operation actions for geometric execution.Questions encode action count, region, operation, magnitude, and preservation requirements; the CFFD Engine receives geometric state separately.Source: §2.2 · p. 5

2.3. SDD Dataset and SDDBench

  • The SDD Dataset links ship forms, valid structured FFD actions, generated language requests, and typed decision records.Its action source contains 30,000 records spanning single edits, combined edits, successive refinements, and requests with explicit numerical magnitudes and ranges.Source: §2.3 · p. 6
  • The SDD Dataset contains 134,558 records after invalid-label augmentations and duplicate descriptions are removed, with each record expanded into ordered typed questions.The train, validation, and test partitions contain 88,604, 22,758, and 23,196 records, respectively.Source: §2.3 · p. 6
  • SDDBench selects 5,000 records and 43,496 questions from the test partition for rapid evaluation, and all compared systems receive the same requests.Evaluation separately measures question-level accuracy and probability metrics, plus exact action matches requiring agreement on action count, order, regions, and operations.Source: §2.3 · pp. 6–8

3. Results

The NURBS representation reconstructs diverse hull forms with low geometric error, while the CFFD deformation tests produce valid hulls across localized and global edits.

3.1.1. NURBS-based hull reconstruction

  • The reconstruction experiment compared twelve hulls using normalized nearest-neighbour surface distances and additional waterline and profile-fit errors.Each hull was reconstructed from 33 extracted horizontal sections fitted with cubic clamped NURBS curves.Source: §3.1.1 · p. 8
  • Across twelve hulls, fore- and aft-profile fitting RMSEs stayed below 0.0013; KCS surface RMSE was 0.0062 and waterline half-breadth RMSE was 0.0031.KCS fore- and aft-profile RMSEs were 0.0003 and 0.0008, respectively.Source: §3.1.1 · p. 8

3.1.2. FFD operation visualization and geometric validity

  • All 30 tested FFD cases passed watertightness and enclosed-volume checks; representative volume ratios were 1.037 for bow-outward, 1.033 for bulb-forward, and 1.120 for global-breadth edits.The first two edits were localized, while global breadth scaling spanned the hull.Source: §3.1.2 · p. 9

3.1. NURBS Representation and FFD Accuracy

  • The workflow connects STL input to NURBS parameters, control-point updates, reconstructed surfaces, and geometric validity checks.The experiments assess representation fidelity alongside spatially selective edits and global scaling.Source: §3.1.2 · p. 9

3.2. SDDBench Evaluation

On SDDBench, Chip leads the compared systems in question accuracy and FFD exact match, with lower probability-error measures than JEV. On JevBench, its accuracy varies with task difficulty, while typed-decision performance remains operational beyond ship design.

3.2.1. Performance on SDDBench

  • 95.90% question accuracy and 99.32% FFD exact match were Chip’s highest scores among compared systems; its NLL, ECE, and Brier score were 0.0951, 0.0032, and 0.0551.JEV reached 90.96% question accuracy and 84.52% FFD exact match; Table 3 reports performance across the comparison.Source: §3.2.1 · p. 10, Table 3 · p. 12
  • 92.87% question accuracy and 99.12% FFD exact match were reported for Claude Opus 5, while GPT-5.6 Sol reached 92.44% and 98.32%, respectively.Their action-level scores were close to Chip, but their question accuracies were lower and probability error measures higher.Source: §3.2.1 · p. 10

3.2.2. General decision capability on JevBench

  • 100.0000%, 81.9444%, and 36.9369% were Chip’s accuracies on JevBench Easy, Original, and Hard tiers, respectively.Chip solved 48 of 48 Easy tasks, 59 of 72 Original tasks, and 41 of 111 Hard tasks.Source: §3.2.2 · p. 11, Table 4 · p. 12
  • JevBench Brier scores rose from 0.0026 on Easy to 0.8782 on Hard, while ECE rose from 0.0075 to 0.3577.The Original tier had a Brier score of 0.3489 and ECE of 0.1633; the evaluation reports a performance gradient with task difficulty.Source: §3.2.2 · p. 12
  • Chip matched the best supplied JevBench accuracy on Easy, exceeded reported Laya and Intern-Decision-0.8B values on Original, and remained below the strongest public reference rows on Hard.The benchmark used the same typed-decision interface outside ship design, and all planned tasks produced valid probability distributions.Source: §3.2.2 · pp. 11–12

3.3. End-to-End Ship Form Editing Evaluation

The KVLCC2 evaluation connects natural-language instructions to sequential FFD actions, reconstructed hulls, and geometric and constraint checks. All executed edits yielded valid hulls, but multi-region success was limited mainly by constraint violations.

3.3. End-to-End Ship Form Editing Evaluation

  • The workflow converts a natural-language instruction into typed decisions and executable FFD actions, updates NURBS waterlines, reconstructs the hull, and checks constraints.Model decision-making and geometric processing were evaluated as parts of the end-to-end stage.Source: §3.3 · p. 12

3.3.1. End-to-end ship-form editing

  • The 300 KVLCC2 experiments were evenly split among single edits, multi-region edits in one turn, and two consecutive same-region edits, comprising 400 requests.Requests requiring direct continuous magnitude or spatial-extent prediction were excluded; the evaluation used qualitative magnitude requests.Source: §3.3.1 · p. 12
  • Each predicted action was applied to the current NURBS waterline state, lofted into a hull mesh, and checked for validity, constraints, action order, and qualitative magnitude.Success required matching the labelled action trajectory and passing geometric and constraint checks.Source: §3.3.1 · p. 12
  • 100% geometry validity accompanied 100% end-to-end success for single edits; continuous two-turn edits achieved 98%, with two constraint violations among 43 constrained experiments.Table 5 reports results across the three interaction settings.Source: §3.3.1 · p. 13, Table 5 · p. 14
  • Multi-region single-turn edits achieved 60% end-to-end success despite 100% execution and geometry checks; 39 of 45 constrained experiments violated a constraint.One experiment also had an action-sequence mismatch, and constraint handling was the main source of unsuccessful trajectories.Source: §3.3.1 · pp. 13–14

3.3.2. Evaluation time

  • Latency measurements included model requests and evaluator-side response handling, but excluded model initialization, result aggregation, and geometric processing.LLM evaluators recorded latency for failed as well as successful attempts.Source: §3.3.2 · p. 14
  • 118.01 ms mean and 235.55 ms 95th-percentile latency were recorded for Chip; the geometric stage added about 0.13 s per end-to-end experiment on average.Geometric processing averaged 0.114 s for a single edit, 0.123 s for a multi-region single-turn edit, and 0.148 s for a same-region two-turn edit.Source: §3.3.2 · p. 14

3.4. Ablation Studies

The ablations compare geometric representation, model scale, and training-data size. Within the tested ranges, expanding the training set produced larger measured gains than increasing model size, while representation changes reduced profile-fit error.

3.4.1. Sensitivity to NURBS Representation

  • The profile ablation compared five configurations on the same twelve hulls, varying control-point count and stem-knot placement separately or in combination.Contour RMSEs were measured against 80-point resampled reference contours.Source: §3.4.1 · p. 14
  • The combined profile representation had the lowest stem-contour RMSE: 0.000343 for KCS and 0.000221 averaged over twelve hulls.Increasing control count reduced both stem and stern errors; deck-knot and bulb-knot changes mainly affected the stem contour.Source: §3.4.1 · p. 15, Table 7 · p. 16

3.4.2. Effect of Chip Model Scale

  • Chip-2B achieved the highest question accuracy and lowest probability-error measures, while Chip achieved the highest FFD exact match; evaluated model sizes showed no marked performance differences.The comparison held data, probability questions, inference protocol, action projection, and SDDBench fixed across 0.8B, 2B, and 4B models.Source: §3.4.2 · pp. 15–16, Table 8 · p. 16

3.4.3. Effect of Training Data Size

  • At 1% training data, Chip reached 94.02% question accuracy and 96.86% FFD exact match, versus 80.82% and 72.24% for the 0% Kev-Base reference.Across trained checkpoints from 1% to 100%, both accuracies increased while NLL, ECE, and Brier score decreased.Source: §3.4.3 · p. 16
  • From 1% to 100% of the training data, question accuracy rose by 1.88 percentage points and FFD exact match by 2.46 points; expanding the dataset yielded larger measured gains than scaling the model.The full-data model had the best measured performance, although gains became smaller at larger data sizes.Source: §3.4.3 · p. 16, §3.4.2 · p. 16, Figure 9 · p. 16

4. Discussion

The discussion finds that end-to-end constraint failures can persist despite correct action semantics, while small local models also struggle with response formatting and action decisions. Across the tested Chip variants, training-data coverage moved measured outcomes more than parameter growth.

4.1. Constraint Semantics in End-to-End

  • Multi-region single-turn editing succeeded in 60% of experiments despite 100% execution and geometry-check rates; 39 of 45 constrained experiments violated a constraint.Across 41 violations, 34 had predicted action semantics and constraint fields matching the reference; sequential edits can leave turn-level checks unsatisfied because action compensation and final verification use different reference states.Source: §4.1 · pp. 17–18

4.2. Why Large Language Models Fail in Ship Free-Form Deformation Decision Making

  • Of 5,000 requests, Qwen3, Gemma3, and Llama3.2 had 4,225, 2,359, and 4,443 rejected records, respectively, chiefly from post-generation interface failures.Among valid responses, question accuracy was 75.61%, 66.97%, and 63.94%, while FFD exact match was 2.84%, 15.98%, and 2.16%, respectively; Chip's mean latency was 118.01 ms versus 18.41–25.28 s for those models.Source: §4.2 · p. 18

4.3. Specialized Model Scale and Data Efficiency

  • From 25% to 100% of training data, question accuracy rose from 95.66% to 95.90% and FFD exact match from 99.14% to 99.32%, while NLL, ECE, and Brier score also decreased.Increasing model size from 0.8 billion to 4 billion parameters changed question accuracy by only 0.0398 percentage points; Chip-2B led on question accuracy and calibration metrics, while Chip led on FFD exact match.Source: §4.3 · pp. 18–19

5. Conclusion and Future Work

The conclusion presents NL2Hull as a discrete constrained decision system connecting natural-language requests to executable FFD and reconstructed-hull checks. It reports strong SDDBench results while identifying sequential-constraint handling and continuous deformation control as areas for further development.

5. Conclusion and Future Work

  • On SDDBench, Chip achieved 95.90% question accuracy and 99.32% FFD exact match with well-calibrated probabilities.The representation study reported mean stem-contour error of 0.000221 over 12 hulls, and end-to-end experiments verified single and multi-turn edits on a KVLCC2 hull.Source: §5 · p. 19
  • The CFFD Engine checks displacement at turn level after action-level compensation, so unconstrained edits composed with constrained ones can alter final volume.The conclusion proposes composition-aware constraint handling, final-state projection, and tolerances tied to reconstruction and meshing resolution.Source: §5 · p. 19
  • The current interface uses ordinal magnitude levels and default regional extents rather than predicting continuous deformation magnitude or spatial extent.Future work is to add continuous numerical outputs trained against geometric outcomes and evaluate their calibration with discrete action accuracy.Source: §5 · p. 19

A. Prompt templates

The appendix presents English translations of runtime prompts that were written in Chinese, documenting the prompts without changing them.

A. Prompt templates

  • The prompts used in the code are written in Chinese and translated into English for readability.The translations document the runtime prompts without changing them.Source: A. Prompt templates · p. 20
  • The prompt blocks use fenced code blocks to display the translated prompts verbatim.The passage describes these as providing the LaTeX equivalent of fenced Markdown code blocks.Source: A. Prompt templates · p. 20
  • Braced fields are replaced with the current request, typed questions, or structured actions.Source: A. Prompt templates · p. 20

A.1. SDDBench probability-output prompt

The probability-output prompt presents each typed question with its candidate keys and asks the model to return a probability distribution for every question as strict JSON.

A.1. SDDBench probability-output prompt

  • Each question block provides an identifier, type, instruction, and candidate keys with descriptions.Source: §A.1 · p. 20
  • The prompt asks the model to output a probability distribution for every question based on the state and questions.Source: §A.1 · p. 20
  • The required response is strict JSON with matching question and option keys, probabilities from 0 to 1 that sum to 1, and no extra text.Choice, noul, and score questions use their prescribed key forms.Source: §A.1 · p. 20

A.2. Natural-language description-generation prompt

The prompt templates generate Chinese ship-edit requests that preserve the structured actions for each editing turn. They vary wording while enforcing content and format constraints, with a retry instruction for invalid outputs.

A.2. Natural-language description-generation prompt

  • Each language variant expresses the same structured action with different sentence structure while preserving its relative magnitude.The prompt permits variation in sentence structure, verbs, degree adverbs, and connectors.Source: §A.2 · p. 20
  • Generated requests must preserve regions, operations, magnitudes, extents, and constraints while following fixed regional and directional vocabulary.The prompt also prohibits internal intensity terminology, added engineering objectives, and physical units.Source: §A.2 · pp. 20–21
  • Prompts distinguish single-action, multi-action, and follow-up turns, carrying prior actions as context when needed.Multi-action requests express all current actions in one sentence; follow-up descriptions cover only the current turn.Source: §A.2 · pp. 20–21
  • When validation fails, the original prompt is supplemented with an instruction to regenerate one Chinese sentence that meets format and semantic requirements.The retry instruction bars internal magnitude-level terminology and asks for no explanation of validation.Source: §A.2 · p. 21

B. Chip training configuration

Chip was adapted from the Kev-0.8B decision checkpoint using the SDD Dataset training partition. The configuration also defines larger model variants and nested training-data-size experiments.

B. Chip training configuration

  • Chip used all 88,604 training-partition records for two epochs, rank-16 low-rank adaptation, and a learning rate of 2 × 10^-5.Training used an effective batch size of eight and bfloat16 computation on an NVIDIA RTX A6000.Source: B. Chip training configuration · p. 22
  • The model-scale experiment adapted Chip-2B and Chip-4B on the same SDD Dataset partition using the main ship-design adaptation settings.Chip-2B began from a decision checkpoint trained on the original Kev dataset, while Chip-4B used the open-sourced Kev-4B decision checkpoint.Source: B. Chip training configuration · p. 22
  • The training-data-size experiment used nested subsets ranging from 1% to 100% of the training partition, or 886 to 88,604 records.Each checkpoint was independently initialized from the same Kev-0.8B checkpoint, with random seed 0.Source: B. Chip training configuration · p. 22

Related papers

Loading 2610.09896v1…