Scholarlink

<p>Electrohydrodynamic (E-jet) printing allows for high-resolution additive manufacturing, but it is important to choose the right process parameters to get the right feature size. This study develops and validates a Random Forest (RF) regression model to forecast droplet diameter based on three process parameters: standoff height, applied voltage, and ink flow rate. We comprehensively compared five machine learning algorithms—Linear Regression, Support Vector Regression, K-Nearest Neighbors, Artificial Neural Networks, and Random Forest—using a 3³ full-factorial experimental design with 27 different combination and three replication each trial led to a total of 81 trials runs. The RF model performed better on independent test data (<i>n</i> = 8), with an RMSE of 23.76&#xa0;μm, a R² of 0.9747, and a MAPE of 2.99% for droplet diameter prediction. Mean Decrease in Impurity, permutation importance, and partial dependence plots all showed that flow rate was the most important predictor (70.3%), followed by standoff height (22.2%) and applied voltage (7.4%). These findings were tested&#xa0;100 times rigorously with various random seeds, and the performance was high (test R² = 0.966 ± 0.021, CV = 2.2%). An error analysis indicated that the residuals were normally distributed (Shapiro-Wilk <i>p</i> = 0.070) and that the systematic bias did not exist, so it was established that the model was reliable. This black-box approach facilitates proper optimization of E-jet processes using limited training data, which saves on experimental resources unlike the traditional trial and error techniques.</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

Improving E-jet process capabilities with black box machine learning

  • 期刊论文

  • 发表日期：

    2025年11月30日

查看全文
  • Amit Kumar Ball,
  • Raju Das,
  • Amit Kumar,
  • Shibendu Shekhar Roy,
  • Dakshina Ranjan Kisku,
  • Naresh Chandra Murmu

摘要

Electrohydrodynamic (E-jet) printing allows for high-resolution additive manufacturing, but it is important to choose the right process parameters to get the right feature size. This study develops and validates a Random Forest (RF) regression model to forecast droplet diameter based on three process parameters: standoff height, applied voltage, and ink flow rate. We comprehensively compared five machine learning algorithms—Linear Regression, Support Vector Regression, K-Nearest Neighbors, Artificial Neural Networks, and Random Forest—using a 3³ full-factorial experimental design with 27 different combination and three replication each trial led to a total of 81 trials runs. The RF model performed better on independent test data (n = 8), with an RMSE of 23.76 μm, a R² of 0.9747, and a MAPE of 2.99% for droplet diameter prediction. Mean Decrease in Impurity, permutation importance, and partial dependence plots all showed that flow rate was the most important predictor (70.3%), followed by standoff height (22.2%) and applied voltage (7.4%). These findings were tested 100 times rigorously with various random seeds, and the performance was high (test R² = 0.966 ± 0.021, CV = 2.2%). An error analysis indicated that the residuals were normally distributed (Shapiro-Wilk p = 0.070) and that the systematic bias did not exist, so it was established that the model was reliable. This black-box approach facilitates proper optimization of E-jet processes using limited training data, which saves on experimental resources unlike the traditional trial and error techniques.