Scholarlink

<p>Conventional machine learning approaches for <i>in-situ</i> tribological monitoring in Directed Energy Deposition (DED-LB) often treat Wear Rate (WR) and Coefficient of Friction (COF) as isolated tasks, ignoring their shared physical representation while operating as opaque systems. To overcome this, this study introduces Tribo-MtGB, an interpretable Multi-Task Gradient Boosting framework. Utilizing real-time thermal-geometric molten pool descriptors, the architecture leverages a shared-representation layer and inverse-variance task-weighting to simultaneously optimize WR and COF predictions across a wide gradient of Stellite 6/WC-12Co metal matrix composites. Validation demonstrates that Tribo-MtGB significantly outperforms state-of-the-art single-task models (XGBoost, MLP), achieving global <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> values exceeding 0.97 for all carbide-reinforced MMCs while operating at an ultra-low inference latency of 0.19&#xa0;ms, enabling true Near Real-Time (NRT) deployment. Furthermore, by integrating Shapley Additive exPlanations (SHAP), the framework autonomously maps how localized thermal history physically governs tribological performance. This work establishes a scalable, non-destructive, and transparent digital methodology for the continuous characterization of multi-material coatings.</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

A Multi-Task Learning Framework for Simultaneous In-Situ Prediction of Wear and Friction in DED-LB MMCs Coatings

  • 期刊论文

  • 发表日期：

    2026年8月1日

查看全文
  • Viridiana Humarán-Sarmiento,
  • Enrique Martinez-Franco,
  • Angel-Iván García-Moreno

摘要

Conventional machine learning approaches for in-situ tribological monitoring in Directed Energy Deposition (DED-LB) often treat Wear Rate (WR) and Coefficient of Friction (COF) as isolated tasks, ignoring their shared physical representation while operating as opaque systems. To overcome this, this study introduces Tribo-MtGB, an interpretable Multi-Task Gradient Boosting framework. Utilizing real-time thermal-geometric molten pool descriptors, the architecture leverages a shared-representation layer and inverse-variance task-weighting to simultaneously optimize WR and COF predictions across a wide gradient of Stellite 6/WC-12Co metal matrix composites. Validation demonstrates that Tribo-MtGB significantly outperforms state-of-the-art single-task models (XGBoost, MLP), achieving global \(R^2\) values exceeding 0.97 for all carbide-reinforced MMCs while operating at an ultra-low inference latency of 0.19 ms, enabling true Near Real-Time (NRT) deployment. Furthermore, by integrating Shapley Additive exPlanations (SHAP), the framework autonomously maps how localized thermal history physically governs tribological performance. This work establishes a scalable, non-destructive, and transparent digital methodology for the continuous characterization of multi-material coatings.