A Large Language Model-Driven 3D Small Sample Enhanced Ship Modeling Method

doi:10.16183/j.cnki.jsjtu.2024.315

Abstract

Abstract:

In the field of ship engineering, the three-dimensional (3D) small sample model is a simplified 3D digital representation of the basic geometric shape and general structure of an actual object, commonly used to validate and optimize the structure and performance of ships. Traditional 3D small sample models are represented by a series of computer-aided design (CAD) operations, which are complex and dependent on specialized modeling expertise. Large language models (LLMs) have shown excellent performance in various fields, but their application in 3D modeling suffers from the phenomenon of “hallucinations”, leading to poor accuracy and robustness. Therefore, an LLM-driven enhanced 3D small sample modeling method for ships is proposed, using a parametric modeling instruction set. First, based on retrieval-augmented generation, existing 3D small sample model code instructions are searched and called, with LLM tools being used to drive the parameters defined in the model, thereby reducing hallucinations in the generation of 3D small sample models. Then, an external ship knowledge database is introduced to supplement information to the generated 3D model with additional information such as model features, process requirements, etc., enhancing the model utility in downstream applications. The results show that the LLM-driven enhanced 3D small sample modeling method for ships can effectively handle different modeling needs, quickly generating ship 3D small sample models with up to 15 features, and significantly improving modeling efficiency.

Key words: 3D small sample ship model, large language model (LLM), retrieval-enhanced generation, reinforcement learning with dual-granularity human feedback (RLHF)

CLC Number:

U662.2

JI Kun, LÜ Chaofan, LÜ Jianhao, BAO Jinsong, MA Yanjun. A Large Language Model-Driven 3D Small Sample Enhanced Ship Modeling Method[J]. Journal of Shanghai Jiao Tong University, 2026, 60(6): 996-1007.

Figures/Tables 15

Fig.1

Fig.2

Fig.3

Fig.4

Tab.1

Fig.5

Fig.6

Fig.7

Tab.2

Tab.3

Tab.4

Fig.8

Fig.9

Fig.10

Fig.11

References 22

[1]	WANG Y L, JIANG Y B, GU Y J, et al. Intelligent 3D layout optimization design of ship engine room based on knowledge-based engineering[J]. Ocean Engineering, 2024, 295: 116966. doi: 10.1016/j.oceaneng.2024.116966 URL
[2]	KENDIBILIR A, KEFAL A. Enhanced ship cross-section design methodology using peridynamics topo-logy optimization[J]. Ocean Engineering, 2023, 286: 115531. doi: 10.1016/j.oceaneng.2023.115531 URL
[3]	WANG Z Y, YANG K, HUANG H S. Resistance spot welding nugget cross-sectional contour extraction and geometric modeling based on the improved Gabor filter[J]. Measurement, 2022, 199: 111448. doi: 10.1016/j.measurement.2022.111448 URL
[4]	ARANBURU A, CAMBA J D, JUSTEL D, et al. An improved explicit reference modeling methodology for parametric design[J]. Computer-Aided Design, 2023, 161: 103541. doi: 10.1016/j.cad.2023.103541 URL
[5]	GOPALAKRISHNAN S, HARTMAN N W, SANGID M D. A digital engineering framework to facilitate automated data exchange between geometric inspection and structural analysis[J]. Advances in Engineering Software, 2023, 183: 103498. doi: 10.1016/j.advengsoft.2023.103498 URL
[6]	ZENG A H, LIU X, DU Z X, et al. GLM-130B:An open bilingual pre-trained model[DB/OL]. (2022-10-05)[2024-08-02]. https://arxiv.org/abs/2210.02414v2.
[7]	ZHENG L M, CHIANG W L, SHENG Y, et al. Judging LLM-as-a-judge with MT-bench and chatbot arena[C]// Proceedings of the 37th International Conference on Neural Information Processing Systems. New Orleans, LA, USA: ACM, 2023: 46595-46623.
[8]	TOUVRON H, MARTIN L, STONE K, et al. Llama 2:Open foundation and fine-tuned chat models[DB/OL]. (2023-07-18)[2024-08-02]. https://arxiv.org/abs/2307.09288.
[9]	FAN W Q, WANG S J, HUANG J N, et al. Graph machine learning in the era of large language models (LLMs)[DB/OL]. (2024-04-23)[2024-08-02]. https://arxiv.org/abs/2404.14928v2.
[10]	BESTA M, BLACH N, KUBICEK A, et al. Graph of thoughts: Solving elaborate problems with large language models[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(16): 17682-17690. doi: 10.1609/aaai.v38i16.29720 URL
[11]	AGUIARI M, GAIOTTI M, RIZZO C M. Ship weight reduction by parametric design of hull scantling[J]. Ocean Engineering, 2022, 263: 112370. doi: 10.1016/j.oceaneng.2022.112370 URL
[12]	TALORDPHOP K, SUKPARUNGSEE S, AREEPONG Y. On designing new mixed modified exponentially weighted moving average-exponentially weighted moving average control chart[J]. Results in Engineering, 2023, 18: 101152. doi: 10.1016/j.rineng.2023.101152 URL
[13]	LEWIS P, PEREZ E, PIKTUS A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]// Proceedings of the 34th International Conference on Neural Information Processing Systems. Vancouver, BC, Canada: ACM, 2020: 9459-9474.
[14]	HOU X G, GOU B C, CHEN D K, et al. A semantic data-driven knowledge base construction method to assist designers in design inspiration based on traditional motifs[J]. Advanced Engineering Informatics, 2023, 56: 101987. doi: 10.1016/j.aei.2023.101987 URL
[15]	LIU H H, CHEN P Y, HU B C, et al. FaissMolLib: An efficient and easy deployable tool for ligand-based virtual screening[J]. Computational Biology and Chemistry, 2024, 110: 108057. doi: 10.1016/j.compbiolchem.2024.108057 URL
[16]	NJOMO W, OZAY G. Sequential analysis coupled with optimized substructure technique modeled on 3D-frame construction process[J]. Engineering Structures, 2014, 80: 200-210. doi: 10.1016/j.engstruct.2014.08.049 URL
[17]	CAI T C, TAN Z W, SONG X R, et al. FoRAG: Factuality-optimized retrieval augmented generation for web-enhanced long-form question answering[DB/OL]. (2024-06-19)[2024-08-02]. https://arxiv.org/abs/2406.13779v1.
[18]	WU Z Q, HU Y S, SHI W J, et al. Fine-grained human feedback gives better rewards for language model training[C]// Proceedings of the 37th International Conference on Neural Information Processing Systems. New Orleans, LA, USA: ACM, 2023: 59008-59033.
[19]	YANG S T, ZHANG S J, XIA C Y, et al. Preference-grounded token-level guidance for language model fine-tuning[DB/OL]. (2023-06-01)[2024-08-02]. https://arxiv.org/abs/2306.00398.
[20]	OUYANG L, WU J, XU J, et al. Training language models to follow instructions with human feedback[C]// Proceedings of the 36th International Conference on Neural Information Processing Systems. New Orleans, LA, USA: ACM, 2022: 27730-27744.
[21]	LIU Y, ITER D, XU Y C, et al. G-eval:NLG eva-luation using GPT-4 with better human alignment[DB/OL]. (2023-03-29)[2024-08-02]. https://arxiv.org/abs/2303.16634v3.
[22]	CHRISTIANO P F, LEIKE J, BROWN T B, et al. Deep reinforcement learning from human preferences[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, CA, USA: ACM, 2017: 4302-4310.

建模规则	功能描述	适用场景
γ	将小样模型分解为独立的功能模块	每个模块可以单独设计和优化
Θ	通过参数化设计快速调整和优化小样模型	小样库中已经存在的三维小样模型
Λ	确保所有设计遵循行业标准和规范	所有船舶设计阶段
Γ	用于创建具有自相似性质的结构	设计船舶的复杂结构和装饰元素
Ω	根据预设条件控制模型的生成	根据特定环境条件进行设计调整
Ψ	引入随机性来模拟不规则性和多样性	为船舶模型添加自然磨损效果或其他随机纹理效果

奖励模型粒度	评价粒度
奖励模型粒度	整体建模序列	子模型的建模序列	子特征的建模序列
序列级别	log loss(${\hat{\mathit{R}}}_{\mathit{\phi }}$, r)^[21]	$\stackrel{\mathrm{L}}{\sum _{\mathrm{p}=1}}$log loss(${\hat{\mathit{R}}}_{\mathit{\phi }}$[p], r_p)^[22]	$\stackrel{\mathit{L}}{\sum _{\mathit{p}=1}}$α_MSE(${\hat{\mathit{R}}}_{\mathit{\phi }}$[p], Agg_q(R_pq))
标记级别	log loss(Agg_τ(${\hat{\mathit{r}}}_{\mathit{\tau }}$), r)^[19]	$\stackrel{\mathit{L}}{\sum _{\mathit{p}=1}}$log loss(Agg_τ(${\hat{\mathit{r}}}_{\mathit{\tau }}$)[p], r_p)	$\stackrel{\mathit{L}}{\sum _{\mathit{p}=1}}$α_MSE(Agg_τ(${\hat{\mathit{r}}}_{\mathit{\tau }}$)[p], Agg_q(R_pq))

方法	多面组合体	方管	弯外连接管	角铁圆法兰
目标模型
几何建模	2.8 min	2.4 min	3 min	3 min
参数化建模	1.2 min	0.8 min	3 min	2.8 min
混合建模	1.2 min	0.8 min	2.9 min	2.8 min
本文	0.8 min	0.3 min	2.4 min	1.7 min

方法	3个特征		6个特征		9个特征		12个特征
方法	平均时间/min	准确度	平均时间/min	准确度	平均时间/min	准确度	平均时间/min	准确度
几何建模	6.4	1.0	9.8	1.0	13.2	1.0	21.8	1.0
参数化建模	2.3	1.0	7.4	0.9	10.3	1.0	13.4	1.0
程序化建模	7.0	1.0	11.1	1.0	18.4	0.9	33.6	0.9
混合建模	5.5	1.0	9.3	0.8	13.1	1.0	18.4	1.0
本文	1.64	1.0	3.68	1.0	3.42	0.8	4.32	1.0