Scholarlink

<p>3D printing parameter selection has traditionally relied on expert knowledge and trial-and-error simulations to meet target consumption criteria such as printing time, material length, and weight. This manual process is time-consuming, non-scalable, and prone to inconsistency, especially under varying design constraints. To address this challenge, we introduce an intelligent framework that predicts optimal 3D printing parameters from geometric features and desired consumption goals using a novel, explainable autoencoder variant. Our proposed framework combines a forward-pretrained model based on XGBoost with a reverse model structured as a deep multi-layer perceptron (MLP), forming a bidirectional autoencoder. The forward model learns to estimate consumption outcomes from geometric and printing parameters, while the reverse model infers the required printing parameters to meet desired consumption goals. A dynamic, phase-wise training strategy is employed to sequentially optimize both reverse reconstruction loss and forward-consistency loss. SHAP-based interpretability is integrated to explain the influence of each feature on the predictions. Experimental evaluation on a dataset of 1,366 samples shows that the forward model achieves high accuracy (R² = 0.9751), while the reverse model, when trained with dynamic weight scheduling, produces printing parameters that yield consumption predictions with R² = 0.9577. The model outperforms traditional regression and static-loss approaches in both accuracy and stability. SHAP analysis highlights the key contributors among geometric and consumption inputs, enhancing transparency and deployment feasibility. This study provides a scalable, explainable, and efficient alternative to expert-based 3D printing design workflows. It demonstrates that an intelligently trained bidirectional autoencoder can autonomously determine printing parameters aligned with multi-objective consumption constraints. The framework has strong practical relevance for low-cost production, design automation, and green manufacturing. Future work will focus on domain adaptation across printer types and online correction via feedback loops.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable bidirectional autoencoder framework for intelligent 3D printing parameter design under multi-objective constraints

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Bich-Ngoc Mach,
  • Duong Thi Kim Chi,
  • Hoa-Cuc Nguyen,
  • Trinh Thi Nhu Quynh,
  • Thanh Q. Nguyen

摘要

3D printing parameter selection has traditionally relied on expert knowledge and trial-and-error simulations to meet target consumption criteria such as printing time, material length, and weight. This manual process is time-consuming, non-scalable, and prone to inconsistency, especially under varying design constraints. To address this challenge, we introduce an intelligent framework that predicts optimal 3D printing parameters from geometric features and desired consumption goals using a novel, explainable autoencoder variant. Our proposed framework combines a forward-pretrained model based on XGBoost with a reverse model structured as a deep multi-layer perceptron (MLP), forming a bidirectional autoencoder. The forward model learns to estimate consumption outcomes from geometric and printing parameters, while the reverse model infers the required printing parameters to meet desired consumption goals. A dynamic, phase-wise training strategy is employed to sequentially optimize both reverse reconstruction loss and forward-consistency loss. SHAP-based interpretability is integrated to explain the influence of each feature on the predictions. Experimental evaluation on a dataset of 1,366 samples shows that the forward model achieves high accuracy (R² = 0.9751), while the reverse model, when trained with dynamic weight scheduling, produces printing parameters that yield consumption predictions with R² = 0.9577. The model outperforms traditional regression and static-loss approaches in both accuracy and stability. SHAP analysis highlights the key contributors among geometric and consumption inputs, enhancing transparency and deployment feasibility. This study provides a scalable, explainable, and efficient alternative to expert-based 3D printing design workflows. It demonstrates that an intelligently trained bidirectional autoencoder can autonomously determine printing parameters aligned with multi-objective consumption constraints. The framework has strong practical relevance for low-cost production, design automation, and green manufacturing. Future work will focus on domain adaptation across printer types and online correction via feedback loops.

Graphical abstract