基于机器学习的飞秒激光加工微坑阵列特征预测  

Machine Learning-Assisted Feature Prediction of Micro Pit Arrays in Femtosecond Laser Processing

作　　者：王泽林[1] 王冰[1] 宋海英[1] 刘世炳[1] WANG Zelin;WANG Bing;SONG Haiying;LIU Shibing(Strong-Field and Ultrafast Photonics Lab,Faculty of Materials and Manufacturing,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学材料与制造学部强场与超快光子学实验室,北京100124

出　　处：《材料保护》2023年第10期67-73,129,共8页Materials Protection

摘　　要：微坑阵列结构具有耐磨损、抗腐蚀、提高生物相容性与抗菌性等特点,应用十分广泛。飞秒激光独特的超快加工效应使其在高质量微坑加工中独具优势。应用随机森林回归(RFR)算法和人工神经网络(ANN)算法对飞秒激光加工的微坑阵列几何形状和质量进行了预测,分析了激光加工参数对微坑的直径、深度和表面粗糙度(Ra)的影响。通过均方根误差、确定系数以及平均绝对误差对RFR与ANN 2种模型的预测能力进行了评估。结果显示:ANN模型的整体预测准确率相比RFR略高一些,R^(2)值为0.81,直径、深度、粗糙度预测的R^(2)分别为0.67、0.79、0.85。利用数据增强方法对数据集进行了扩增,ANN模型的准确率进一步提高,整体R^(2)为0.91,直径、深度、粗糙度预测的R^(2)分别为0.81、0.91、0.95。研究结果表明,ANN模型在飞秒激光加工微坑阵列的预测中相比RFR具有更优异的预测性能,且随着数据量的增加,这种优势更加明显,也进一步验证了ANN模型的潜力。With the characteristics of wear resistance,corrosion resistance,improved biocompatibility and antibacterial properties,the micropits array structure has been widely used.femtosecond laser has unique advantages in high-quality micro pit processing,because of its unique ultra-fast processing effect.In this work,Random Forest Regression(RFR)algorithm and Artificial Neural Network(ANN)algorithm were applied to predict the geometry and quality of micro-pit arrays processed by femtosecond laser.Additionally,the effects of laser processing parameters on the diameter,depth and surface roughness(Ra)of micro-pits were analyzed.The predictive capabilities of the RFR and ANN models were evaluated through the root mean square error,coefficient of determination and mean absolute error.Results showed that the overall prediction accuracy of ANN model was slightly higher than RFR model.The R^(2) for ANN model was 0.81.For diameter,depth and surface roughness,the R^(2) was 0.67,0.79 and 0.85,respectively.Data augmentation method was applied to augment the dataset,and the ANN model prediction accuracy was further improved after data augmentation.The overall R^(2) increased to 0.91.The R^(2) for diameter,depth and surface roughness was 0.81,0.91 and 0.95,respectively.In general,the ANN model had better prediction performance than Random Forest in predicting micro pit arrays processed by femtosecond laser processing.As the amount of data increased,this advantage became more obvious,which further verified the potential of the ANN model.

关 键 词：机器学习 人工神经网络 激光加工 参数预测 

分 类 号：TQ181[化学工程]