Scholarlink

<p>Laser Powder Bed Fusion (LPBF) enables metal additive manufacturing, but achieving high Relative Density (RD) in new alloys requires extensive process parameter optimization. This, however, is highly constrained by sparse, imbalanced datasets and high experimental costs. State-of-the-art data-driven methods struggle to generalize across materials due to limited datasets. This study introduces Physics-Informed K-Nearest Neighbors (PIKNN), a training-free few-shot learning approach that leverages an eight-dimensional physics-constrained feature space incorporating process parameters, energy deposition metrics, and material-specific thermal diffusivities to enable cross-material RD prediction without parametric optimization. Using a publicly available dataset of 1,579 LPBF prints across six alloys, we evaluate PIKNN in a rigorous cross-material paradigm: training on four source alloys (1,244 samples) and testing on unseen Ti6Al4V and CuCrZr (335 samples). RD measurements are binned into four physics-guided quality classes. PIKNN, implemented as 1-nearest-neighbor classification in the standardized physics space, is benchmarked against a Prototypical Network and a classical SVM baseline that share identical features and episodic protocols (4-way, 1–10 shots). Results demonstrate PIKNN lifting accuracy to 58.0% on Ti6Al4V (10-shot) and 52.0% on CuCrZr (1-shot), yielding gains up to 18.3 percentage points over the trained baseline while matching Ti6Al4V 1-shot performance within 0.5 points. Ablations confirm the critical role of physics-derived features like energy density, while separability analyses (clustering metrics, linear probes, t-SNE) reveal that physics features preserve class structure better than learned embeddings. Together, these results indicate that explicit physics encoding captures cross-material transferable structure that learned embeddings fail to represent, creating a viable path for rapid LPBF material screening with a limited dataset and thus serving as a rapid screening and parameter-space reduction tool for early-stage material qualification.</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

Physics-informed few-shot learning for cross-material relative density prediction in laser powder bed fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月2日

查看全文
  • Prashant Dhakal,
  • Jae Gwang Kim,
  • Aolin Hou,
  • Xiaofei Wu,
  • Shiren Wang

摘要

Laser Powder Bed Fusion (LPBF) enables metal additive manufacturing, but achieving high Relative Density (RD) in new alloys requires extensive process parameter optimization. This, however, is highly constrained by sparse, imbalanced datasets and high experimental costs. State-of-the-art data-driven methods struggle to generalize across materials due to limited datasets. This study introduces Physics-Informed K-Nearest Neighbors (PIKNN), a training-free few-shot learning approach that leverages an eight-dimensional physics-constrained feature space incorporating process parameters, energy deposition metrics, and material-specific thermal diffusivities to enable cross-material RD prediction without parametric optimization. Using a publicly available dataset of 1,579 LPBF prints across six alloys, we evaluate PIKNN in a rigorous cross-material paradigm: training on four source alloys (1,244 samples) and testing on unseen Ti6Al4V and CuCrZr (335 samples). RD measurements are binned into four physics-guided quality classes. PIKNN, implemented as 1-nearest-neighbor classification in the standardized physics space, is benchmarked against a Prototypical Network and a classical SVM baseline that share identical features and episodic protocols (4-way, 1–10 shots). Results demonstrate PIKNN lifting accuracy to 58.0% on Ti6Al4V (10-shot) and 52.0% on CuCrZr (1-shot), yielding gains up to 18.3 percentage points over the trained baseline while matching Ti6Al4V 1-shot performance within 0.5 points. Ablations confirm the critical role of physics-derived features like energy density, while separability analyses (clustering metrics, linear probes, t-SNE) reveal that physics features preserve class structure better than learned embeddings. Together, these results indicate that explicit physics encoding captures cross-material transferable structure that learned embeddings fail to represent, creating a viable path for rapid LPBF material screening with a limited dataset and thus serving as a rapid screening and parameter-space reduction tool for early-stage material qualification.