Scholarlink

<p>A physics-constrained Artificial Neural Network is developed to predict the top kerf width and volume of material removed per unit length in Abrasive Waterjet Cutting (AWJC) process. The model is trained using data based on rigorous experiments conducted on two workpiece materials viz., aluminum alloy and mild steel. The network architecture incorporates six input process parameters, such as tensile strength of the workpiece, specimen thickness, stand-off distance, abrasive mass flow rate, water jet pressure and traverse speed of the jet. A custom physics-constrained loss function enforces that the derived bottom kerf width calculated from the predicted top kerf width and specific volume must be non-negative. A Bayesian optimization algorithm is implemented <i>via</i> the Optuna hyperparameter tuning library that identifies optimal configuration. The trained feedforward model further enables an inference framework that classifies the type of cut (through or non-through) and reconstructs the bottom kerf width for through-cuts. Furthermore, the model predicts the lower and upper bounds of the top kerf width and specific volume <i>via</i> a modified backpropagation algorithm. Quantitatively, the model achieves a mean absolute percentage error (MAPE) of 2.33% for top kerf width and 2.71% for specific volume on the testing set. The cut type classification reaches 98.6% accuracy and infers the bottom kerf width for through-cuts. The predicted lower and upper bounds consistently bracket the most likely estimates with average interval widths of 0.32&#xa0;mm for top kerf width and 0.95&#xa0;mm³/mm for specific volume. The model is implemented using the PyTorch deep learning library in Python.</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 physics constrained neural network for predictive modeling of abrasive waterjet cutting

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Ngangkham Peter Singh,
  • Rajkumar Shufen

摘要

A physics-constrained Artificial Neural Network is developed to predict the top kerf width and volume of material removed per unit length in Abrasive Waterjet Cutting (AWJC) process. The model is trained using data based on rigorous experiments conducted on two workpiece materials viz., aluminum alloy and mild steel. The network architecture incorporates six input process parameters, such as tensile strength of the workpiece, specimen thickness, stand-off distance, abrasive mass flow rate, water jet pressure and traverse speed of the jet. A custom physics-constrained loss function enforces that the derived bottom kerf width calculated from the predicted top kerf width and specific volume must be non-negative. A Bayesian optimization algorithm is implemented via the Optuna hyperparameter tuning library that identifies optimal configuration. The trained feedforward model further enables an inference framework that classifies the type of cut (through or non-through) and reconstructs the bottom kerf width for through-cuts. Furthermore, the model predicts the lower and upper bounds of the top kerf width and specific volume via a modified backpropagation algorithm. Quantitatively, the model achieves a mean absolute percentage error (MAPE) of 2.33% for top kerf width and 2.71% for specific volume on the testing set. The cut type classification reaches 98.6% accuracy and infers the bottom kerf width for through-cuts. The predicted lower and upper bounds consistently bracket the most likely estimates with average interval widths of 0.32 mm for top kerf width and 0.95 mm³/mm for specific volume. The model is implemented using the PyTorch deep learning library in Python.