Scholarlink

<p>In Laser Powder Directed Energy Deposition (LP-DED), bead geometry directly influences the dimensional fidelity and mechanical integrity of multi-layer components. However, accurately and efficiently predicting some bead geometric features remain challenging due to the nonlinear coupling between multiple process parameters and melt pool dynamics. These challenges are particularly significant when only limited experimental datasets are available. This study proposes a data-driven framework to predict bead geometry using ensemble machine learning models. Experimental data from single-track Inconel 718 depositions were used to train and evaluate four supervised regression models: Random Forest, XGBoost, LightGBM, and CatBoost. Among them, CatBoost exhibited the best overall performance, achieving <i>R</i>² values above 0.95 for height, width, and angle, and over 0.90 for depth after fine-tuning. A multi-target CatBoost model further validated the framework’s robustness, achieving an R² of 0.94 and a mean squared error (MSE) of 33.3. The results indicate that CatBoost provides robust predictive capability for small experimental datasets typical of LP-DED studies. Sensitivity analysis identified powder feed rate as the most influential parameter for height and wetting angle, while depth was governed by a more complex interplay of all input variables. The proposed framework demonstrates the potential of ensemble learning for capturing complex process-geometry relationships and provides a systematic comparison of ensemble learning methods for bead geometry prediction in LP-DED.</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

Predicting bead geometry in laser powder directed energy deposition: ensemble machine learning models for complex process parameters

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Amir Shakibi,
  • Amir Yah,
  • Yue Zhou

摘要

In Laser Powder Directed Energy Deposition (LP-DED), bead geometry directly influences the dimensional fidelity and mechanical integrity of multi-layer components. However, accurately and efficiently predicting some bead geometric features remain challenging due to the nonlinear coupling between multiple process parameters and melt pool dynamics. These challenges are particularly significant when only limited experimental datasets are available. This study proposes a data-driven framework to predict bead geometry using ensemble machine learning models. Experimental data from single-track Inconel 718 depositions were used to train and evaluate four supervised regression models: Random Forest, XGBoost, LightGBM, and CatBoost. Among them, CatBoost exhibited the best overall performance, achieving R² values above 0.95 for height, width, and angle, and over 0.90 for depth after fine-tuning. A multi-target CatBoost model further validated the framework’s robustness, achieving an R² of 0.94 and a mean squared error (MSE) of 33.3. The results indicate that CatBoost provides robust predictive capability for small experimental datasets typical of LP-DED studies. Sensitivity analysis identified powder feed rate as the most influential parameter for height and wetting angle, while depth was governed by a more complex interplay of all input variables. The proposed framework demonstrates the potential of ensemble learning for capturing complex process-geometry relationships and provides a systematic comparison of ensemble learning methods for bead geometry prediction in LP-DED.