Integrating large language models (LLMs) into printing process optimization for binder jetting printed Ti-6Al-4V

Standard

Integrating large language models (LLMs) into printing process optimization for binder jetting printed Ti-6Al-4V. / Qian, Y.; Luo, X.; Wei, Q.; Yi, S.; Zhang, L.; Huang, D.; Qin, N.; Liu, R.; Huang, R.; Yu, K.O.; Konakov, V.G.; Wu, M.; Fu, Z.

в: Powder Technology, Том 474, 122295, 01.05.2026.

Результаты исследований: Научные публикации в периодических изданиях › статья › Рецензирование

Author

Qian, Y. ; Luo, X. ; Wei, Q. ; Yi, S. ; Zhang, L. ; Huang, D. ; Qin, N. ; Liu, R. ; Huang, R. ; Yu, K.O. ; Konakov, V.G. ; Wu, M. ; Fu, Z. / Integrating large language models (LLMs) into printing process optimization for binder jetting printed Ti-6Al-4V. в: Powder Technology. 2026 ; Том 474.

BibTeX

@article{6eacb9de73a14c2782647b4093f3ddfd,

title = "Integrating large language models (LLMs) into printing process optimization for binder jetting printed Ti-6Al-4V",

abstract = "The binder jetting additive manufacturing (BJAM) process involves multiple parameters such as layer thickness, reference voltage, powder spreading speed, drying speed, and roller speed etc. Traditional experimental parameter optimization faces severe challenges. Recent generative AI (Gen AI) advancements offer transformative potential for manufacturing. This study proposes a novel stacking NeuroSVK model integrated with a Large Language Model (LLM) to optimize printing parameters and simultaneously predict the relative density and dimensional variation of the printed Ti-6Al-4 V green bodies. A printing dataset of 43 experimental sets was evaluated using K-fold cross-validation for comprehensive model training and validation. Predictive accuracy was assessed via Mean Squared Error (MSE) and the coefficient of determination (RPorter et al. (2023)2). The LLM was fine-tuned with 100 prompts to help evaluate the reasonableness of model-optimized parameters and to develop a question-answering agent. Results show the model's recommendations are very close to the experimental optimum with R2 of 0.94, and the agent evaluates the reasonableness of the recommended parameters. Importantly, the density distribution of green bodies printed using the model's parameters is more stable than with experimentally optimized parameters. Finally, the TC4 engine piston printed using the model-optimized parameters was sintered at 1400 °C for 2 h, and a relative density ≥ 97% was reached. The microstructure consisted of α-phase with an average grain size of 28.09 ± 14.23 μm and β-phase at grain boundaries, with a microhardness of 362 HV, tensile strength of 1015 ± 23 MPa, elongation of 10.2 ± 0.6%, carbon content of 0.14 wt%, oxygen content of 0.28 wt%, and dimensional shrinkage of 15.75%, 14.55%, and 18.34% in the X, Y, and Z directions, respectively. {\textcopyright} 2026 Elsevier B.V.",

keywords = "Binder jetting, Deep learning (DL), Parameters optimization, Titanium alloy, Aluminum alloys, Binders, Deep learning, Grain boundaries, Mean square error, Optimization, Printing presses, Ternary alloys, Titanium alloys, Vanadium alloys, titanium, Green body, Language model, Optimized parameter, Parameter optimization, Printing process, Process optimisation, Relative density, Titanium (alloys), Article, artificial neural network, binder jetting, cross validation, deep learning, large language model, mean squared error, micro-computed tomography, nerve cell network, predictive model, principal component analysis, process optimization, support vector machine, surface tension, three dimensional printing, Tensile strength",

author = "Y. Qian and X. Luo and Q. Wei and S. Yi and L. Zhang and D. Huang and N. Qin and R. Liu and R. Huang and K.O. Yu and V.G. Konakov and M. Wu and Z. Fu",

note = "Export Date: 29 March 2026; Cited By: 0; Correspondence Address: X. Luo; School of New Energy and Materials, Southwest Petroleum University, Chengdu, 610500, China; email: winifreed@163.com; L. Zhang; Institute of Intelligent Manufacturing Technology, Shenzhen Polytechnic University, Shenzhen, 518055, China; email: zhangliang@szpu.edu.cn; CODEN: POTEB",

year = "2026",

month = may,

day = "1",

doi = "10.1016/j.powtec.2026.122295",

language = "English",

volume = "474",

journal = "Powder Technology",

issn = "0032-5910",

publisher = "Elsevier",

}

RIS

TY - JOUR

T1 - Integrating large language models (LLMs) into printing process optimization for binder jetting printed Ti-6Al-4V

AU - Qian, Y.

AU - Luo, X.

AU - Wei, Q.

AU - Yi, S.

AU - Zhang, L.

AU - Huang, D.

AU - Qin, N.

AU - Liu, R.

AU - Huang, R.

AU - Yu, K.O.

AU - Konakov, V.G.

AU - Wu, M.

AU - Fu, Z.

N1 - Export Date: 29 March 2026; Cited By: 0; Correspondence Address: X. Luo; School of New Energy and Materials, Southwest Petroleum University, Chengdu, 610500, China; email: winifreed@163.com; L. Zhang; Institute of Intelligent Manufacturing Technology, Shenzhen Polytechnic University, Shenzhen, 518055, China; email: zhangliang@szpu.edu.cn; CODEN: POTEB

PY - 2026/5/1

Y1 - 2026/5/1

N2 - The binder jetting additive manufacturing (BJAM) process involves multiple parameters such as layer thickness, reference voltage, powder spreading speed, drying speed, and roller speed etc. Traditional experimental parameter optimization faces severe challenges. Recent generative AI (Gen AI) advancements offer transformative potential for manufacturing. This study proposes a novel stacking NeuroSVK model integrated with a Large Language Model (LLM) to optimize printing parameters and simultaneously predict the relative density and dimensional variation of the printed Ti-6Al-4 V green bodies. A printing dataset of 43 experimental sets was evaluated using K-fold cross-validation for comprehensive model training and validation. Predictive accuracy was assessed via Mean Squared Error (MSE) and the coefficient of determination (RPorter et al. (2023)2). The LLM was fine-tuned with 100 prompts to help evaluate the reasonableness of model-optimized parameters and to develop a question-answering agent. Results show the model's recommendations are very close to the experimental optimum with R2 of 0.94, and the agent evaluates the reasonableness of the recommended parameters. Importantly, the density distribution of green bodies printed using the model's parameters is more stable than with experimentally optimized parameters. Finally, the TC4 engine piston printed using the model-optimized parameters was sintered at 1400 °C for 2 h, and a relative density ≥ 97% was reached. The microstructure consisted of α-phase with an average grain size of 28.09 ± 14.23 μm and β-phase at grain boundaries, with a microhardness of 362 HV, tensile strength of 1015 ± 23 MPa, elongation of 10.2 ± 0.6%, carbon content of 0.14 wt%, oxygen content of 0.28 wt%, and dimensional shrinkage of 15.75%, 14.55%, and 18.34% in the X, Y, and Z directions, respectively. © 2026 Elsevier B.V.

AB - The binder jetting additive manufacturing (BJAM) process involves multiple parameters such as layer thickness, reference voltage, powder spreading speed, drying speed, and roller speed etc. Traditional experimental parameter optimization faces severe challenges. Recent generative AI (Gen AI) advancements offer transformative potential for manufacturing. This study proposes a novel stacking NeuroSVK model integrated with a Large Language Model (LLM) to optimize printing parameters and simultaneously predict the relative density and dimensional variation of the printed Ti-6Al-4 V green bodies. A printing dataset of 43 experimental sets was evaluated using K-fold cross-validation for comprehensive model training and validation. Predictive accuracy was assessed via Mean Squared Error (MSE) and the coefficient of determination (RPorter et al. (2023)2). The LLM was fine-tuned with 100 prompts to help evaluate the reasonableness of model-optimized parameters and to develop a question-answering agent. Results show the model's recommendations are very close to the experimental optimum with R2 of 0.94, and the agent evaluates the reasonableness of the recommended parameters. Importantly, the density distribution of green bodies printed using the model's parameters is more stable than with experimentally optimized parameters. Finally, the TC4 engine piston printed using the model-optimized parameters was sintered at 1400 °C for 2 h, and a relative density ≥ 97% was reached. The microstructure consisted of α-phase with an average grain size of 28.09 ± 14.23 μm and β-phase at grain boundaries, with a microhardness of 362 HV, tensile strength of 1015 ± 23 MPa, elongation of 10.2 ± 0.6%, carbon content of 0.14 wt%, oxygen content of 0.28 wt%, and dimensional shrinkage of 15.75%, 14.55%, and 18.34% in the X, Y, and Z directions, respectively. © 2026 Elsevier B.V.

KW - Binder jetting

KW - Deep learning (DL)

KW - Parameters optimization

KW - Titanium alloy

KW - Aluminum alloys

KW - Binders

KW - Deep learning

KW - Grain boundaries

KW - Mean square error

KW - Optimization

KW - Printing presses

KW - Ternary alloys

KW - Titanium alloys

KW - Vanadium alloys

KW - titanium

KW - Green body

KW - Language model

KW - Optimized parameter

KW - Parameter optimization

KW - Printing process

KW - Process optimisation

KW - Relative density

KW - Titanium (alloys)

KW - Article

KW - artificial neural network

KW - binder jetting

KW - cross validation

KW - deep learning

KW - large language model

KW - mean squared error

KW - micro-computed tomography

KW - nerve cell network

KW - predictive model

KW - principal component analysis

KW - process optimization

KW - support vector machine

KW - surface tension

KW - three dimensional printing

KW - Tensile strength

UR - https://www.mendeley.com/catalogue/1958ea30-9f27-357b-8917-9f3453d6d3d7/

U2 - 10.1016/j.powtec.2026.122295

DO - 10.1016/j.powtec.2026.122295

M3 - Article

VL - 474

JO - Powder Technology

JF - Powder Technology

SN - 0032-5910

M1 - 122295

ER -

ID: 151902489