Scholarlink

<p>Physics-informed machine learning (PIML) represents a promising area within Laser-Assisted Machining (LAM) process. However, previous approaches have heavily relied on extensive datasets for their success, posing challenges given the limited availability of data in LAM process. Addressing this issue, the current study introduces new Architecture-guided PIML (APIML) framework established on deep unfolding, tailored by scenarios with sparse datapoints. Specifically, APIML is designed to predict thermal histories in the LAM process. The architectural design of APIML draws inspiration from iterative temperature model equations, where each iteration corresponds to a neural network layer. APIML framework hyperparameters are meticulously optimized, and their performance rigorously evaluated. Notably, when tested with thousand data points split at an 80:20 ratio, APIML achieves a mean absolute percentage error (MAPE) of 3.6% and an R<sup>2</sup> value of 0.941 Comparative analysis pits APIML against traditional machine learning models including artificial neural network, decision tree regressor, random forest regressor, and support vector regressor. Results demonstrate APIML’s superior performance: it achieves a 56.98% lower MAPE and a 17.1% higher R<sup>2</sup> compared to the best-performing decision tree regressor model among the traditional approaches. In essence, APIML showcases promising advancements in leveraging physics-informed machine learning to overcome data scarcity challenges in LAM process, offering more accurate predictions of thermal histories crucial for optimizing machining processes.</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

Architecture-Guided Physics-Learned Machine Learning for Temperature Prediction in Laser-Assisted Turning Process

  • 期刊论文

  • 发表日期：

    2025年2月5日

查看全文
  • Mondi Rama Karthik,
  • Thella Babu Rao

摘要

Physics-informed machine learning (PIML) represents a promising area within Laser-Assisted Machining (LAM) process. However, previous approaches have heavily relied on extensive datasets for their success, posing challenges given the limited availability of data in LAM process. Addressing this issue, the current study introduces new Architecture-guided PIML (APIML) framework established on deep unfolding, tailored by scenarios with sparse datapoints. Specifically, APIML is designed to predict thermal histories in the LAM process. The architectural design of APIML draws inspiration from iterative temperature model equations, where each iteration corresponds to a neural network layer. APIML framework hyperparameters are meticulously optimized, and their performance rigorously evaluated. Notably, when tested with thousand data points split at an 80:20 ratio, APIML achieves a mean absolute percentage error (MAPE) of 3.6% and an R2 value of 0.941 Comparative analysis pits APIML against traditional machine learning models including artificial neural network, decision tree regressor, random forest regressor, and support vector regressor. Results demonstrate APIML’s superior performance: it achieves a 56.98% lower MAPE and a 17.1% higher R2 compared to the best-performing decision tree regressor model among the traditional approaches. In essence, APIML showcases promising advancements in leveraging physics-informed machine learning to overcome data scarcity challenges in LAM process, offering more accurate predictions of thermal histories crucial for optimizing machining processes.