Scholarlink

<p>Accurate prediction of relative density in Laser Powder Bed Fusion (LPBF) remains challenging because of complex interactions among process parameters, material characteristics, and machine-specific variability. While machine learning has shown considerable promise for LPBF modelling, most existing studies provide limited assessment of model robustness, reliability, and generalization across heterogeneous manufacturing environments. This study proposes the Reliability-Aware Physics-Guided Learning (RAPIL) framework, which integrates reduced-order thermal modelling, physics-guided feature engineering, hierarchical learning, explainable machine learning, uncertainty quantification, and comprehensive validation. A reduced-order analytical thermal model estimates intermediate thermal descriptors, including peak temperature, cooling rate, thermal gradient, and melt-pool geometry, which are incorporated into a Stage-II XGBoost model for relative density prediction. The framework was developed using a heterogeneous dataset of 1,579 LPBF experiments from multiple alloys and machine platforms. The global model achieved an R² of 0.686 with an RMSE of 3.54%. Robustness was evaluated using five-fold cross-validation, independent external validation (representative external R² = 0.773), and repeated multi-seed external validation (mean external R² = 0.671 ± 0.099). Machine-specific analyses demonstrated satisfactory predictive performance for most well-represented systems while highlighting the challenges of cross-machine generalization. SHAP analysis identified cooling rate and energy-related thermal descriptors as the dominant predictors, whereas ensemble-based uncertainty estimates showed a positive association with prediction error. Overall, RAPIL provides an interpretable, reliability-aware framework for relative density prediction across heterogeneous LPBF datasets while highlighting the remaining challenges associated with machine-dependent variability and model transferability.</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

RAPIL: a reliability-aware physics-guided learning framework for robust density prediction and generalization assessment in laser powder bed fusion

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Aswin Karkadakattil

摘要

Accurate prediction of relative density in Laser Powder Bed Fusion (LPBF) remains challenging because of complex interactions among process parameters, material characteristics, and machine-specific variability. While machine learning has shown considerable promise for LPBF modelling, most existing studies provide limited assessment of model robustness, reliability, and generalization across heterogeneous manufacturing environments. This study proposes the Reliability-Aware Physics-Guided Learning (RAPIL) framework, which integrates reduced-order thermal modelling, physics-guided feature engineering, hierarchical learning, explainable machine learning, uncertainty quantification, and comprehensive validation. A reduced-order analytical thermal model estimates intermediate thermal descriptors, including peak temperature, cooling rate, thermal gradient, and melt-pool geometry, which are incorporated into a Stage-II XGBoost model for relative density prediction. The framework was developed using a heterogeneous dataset of 1,579 LPBF experiments from multiple alloys and machine platforms. The global model achieved an R² of 0.686 with an RMSE of 3.54%. Robustness was evaluated using five-fold cross-validation, independent external validation (representative external R² = 0.773), and repeated multi-seed external validation (mean external R² = 0.671 ± 0.099). Machine-specific analyses demonstrated satisfactory predictive performance for most well-represented systems while highlighting the challenges of cross-machine generalization. SHAP analysis identified cooling rate and energy-related thermal descriptors as the dominant predictors, whereas ensemble-based uncertainty estimates showed a positive association with prediction error. Overall, RAPIL provides an interpretable, reliability-aware framework for relative density prediction across heterogeneous LPBF datasets while highlighting the remaining challenges associated with machine-dependent variability and model transferability.