Technology
Tech Breakthrough
Texas A&M University Achieves Cross-Alloy Density Prediction for LPBF via Physics-Informed K-Nearest Neighbors
2026-06-02
A team at Texas A&M University proposes PIKNN (Physics-Informed K-Nearest Neighbors), a training-free few-shot learning method. By constructing an eight-dimensional physically constrained feature space incorporating laser power, scan speed, layer thickness, hatch spacing, volumetric energy density, linear energy density, and material thermal diffusivity, they achieve cross-material relative density prediction across 1,579 samples from six alloys. Trained on four source alloys (1,244 samples), PIKNN achieves 58.0% accuracy on unseen Ti6Al4V (10-shot) and 52.0% accuracy on CuCrZr (1-shot), outperforming supervised baselines (Prototypical Networks, SVM) by up to 18.3%. Ablation studies confirm that physically derived features such as energy density play a critical role, with explicit physical encoding capturing cross-material transferable structures that data-driven learned embeddings cannot express.
📰 News Source
This article was translated from a Chinese source.
Source URL not available
Show Chinese original
德州农工大学通过物理信息K近邻实现LPBF跨合金致密度预测
德州农工大学团队提出PIKNN（物理信息K近邻），一种无需训练的少样本学习方法。通过构建包含激光功率、扫描速度、层厚、扫描间距、体能量密度、线能量密度及材料热扩散率的八维物理约束特征空间，在6种合金1579个样本上实现跨材料相对密度预测。在4种源合金（1244样本）训练下，PIKNN对未见过的Ti6Al4V（10-shot）准确率达58.0%，对CuCrZr（1-shot）准确率达52.0%，比有监督基线（原型网络、SVM）最高提升18.3%。消融实验证实能量密度等物理派生特征起关键作用，显式物理编码可捕捉数据驱动学习嵌入无法表达的跨材料可迁移结构。