A hybrid machine learning approach to determine the optimal processing window in femtosecond laser-induced periodic nanostructures

Highlights

• A hybrid machine learning method is proposed for intelligent nanostructuring.

• Transfer learning is applied for feature extraction with high accuracy.

• Best operating window is determined for LIPSS fabrication.

• DT has the highest accuracy for classification.

Abstract

Surface nanostructuring could enhance surface properties such as strength, self-cleaning, anti-fog and anti-bacterial properties. Femtosecond laser-induced periodic surface structures (LIPSS) is a nanoscale structure created with laser technique. However, its quality is significantly influenced by the complicated interrelationship between the various laser processing and material parameters. Hitherto the selection of the appropriate laser parameters mainly depends on personal experience in conjunction with many time-consuming experimental trials. To have a simple, fast, and intelligent process, a hybrid machine learning method is proposed to determine the optimized processing window for femtosecond laser-induced nanostructures. Firstly, k-means clustering method was applied to automatically classify the laser-induced nanostructures into good and bad quality classes. Before clustering, dimensionality reduction methods were applied to reduce the high dimension of image data and to extract features. Different dimensionality reduction methods including principal component analysis (PCA), local linear embedding (LLE), t-random adjacent embedding (t-SNE) and transfer learning were explored. Transfer learning showed a much better result compared with other dimensionality reduction methods. Transfer learning VGG19 model achieved the highest accuracy of 90.6 %. After clustering, the image was labelled as good and bad clusters accordingly. The labeled image was trained using artificial neural network (ANN), random forest (RF), decision tree (DT), support vector machine (SVM), k-nearest neighbors (KNN) and Naive Bayesian Classifier (NBC) algorithms for the prediction of laser processing results. The results show that DT gives the best accuracy of 96.7 %. Finally, an optimal laser processing window for femtosecond laser-induced nanostructures was determined.

Introduction

Nano-manufacturing refers to processing precisions in the nanometer range. From a practical perspective, cost and reliability are major considerations. Nano-manufacturing can effectively improve the performance of materials such as strength enhancement, lightening, improving durability and adding new functions (Amanov et al., 2016). Compared with traditional nano-manufacturing methods such as chemical etching, photolithography, ion beam processing and nano-imprinting, femtosecond laser processing has its advantages, such as fast processing, low cost, high efficiency, high reliability and environmental friendliness (Wang et al., 2015a, Wang et al., 2015b). Moreover, femtosecond laser processing can be applied for almost all solid materials including metals, semiconductors, dielectrics, polymers and biological tissues. The heat-affected zone of the processed material is small and no subsequent treatment is needed. The damage to the material is low, and this makes it an ideal tool for nanofabrication (Amoako, 2019).

Since femtosecond laser machining has become an important technique in the past few decades, many studies have been carried out for its applications in industrial, aerospace, microfluidic, biological and medical fields. It has been discovered that laser-induced periodic surface structures (LIPSS) could be formed on material surfaces during laser-material interaction. When a linearly polarized laser beam interacts with a solid surface, there are two types of LIPSS formed, namely low spatial frequency LIPSS (LSFL) and high spatial frequency LIPSS (HSFL) (Jörn et al., 2017). For LSFL, it is generally created on a high absorption material surface such as a metallic surface. The direction of LIPSS is perpendicular to the laser beam polarization direction. In contrast, for HSFL, it is generated during the laser processing of transparent materials such as dielectrics. The LIPSS direction is either perpendicular or parallel to the laser polarization direction. LIPSS can be created in a single processing step for many unique applications (San-Blas et al., 2020) such as optical, mechanical and chemical property control. Surfaces with LIPSS can act as diffraction gratings for color marking. Dusser et al. (2010) developed a large area color marking technique by applying LIPSS. In addition, LIPSS structure can be combined with hologram to realize a vivid holographic display under white light irradiation (Jwad et al., 2018). Wang et al., 2015a, Wang et al., 2015b reported using of LIPSS to reduce friction of metallic surfaces. Vorobyev and Guo (2015) reported that LIPSS can be used to improve the hydrophilicity and hydrophobicity of the material and to achieve a superhydrophobic surface. Moreover, surface biocompatibility could be improved with LIPSS and the cell growth direction can be guided by the direction of LIPSS (Rebollar et al., 2008). All these investigations demonstrate that LIPSS has extraordinary charm and prospects in different applications.

Laser processing parameters, such as laser power, laser scanning speed, and the laser scanning times, have a great influence on the formation and quality of LIPSS. The effects of these processing parameters on the formation of LIPSS are nonlinear. Hitherto the selection of the appropriate laser parameters mainly depends on personal experience in conjunction with many time-consuming experimental trials. Thus, it is desirable to develop a controllable and optimized fabrication system for LIPSS.

With increasing computing power and development of machine learning algorithms, the advantages of machine learning for predictive modeling in material processing are gradually being recognized. Machine learning algorithms have been applied for the prediction of laser drilling (Ghoreishi and Nakhjavani, 2008), laser cutting (Chaki et al., 2020), laser microgrooving (Dixit et al., 2019) and laser milling (Leone et al., 2019). Bakhtiyari et al. (2021) discussed the application of artificial intelligence in laser machining. The applications of the common machine learning methods for geometry characteristics, metallurgical characteristics, surface quality and material removal rate were summarized. Teixidor et al. (2015) used different machine learning methods to predict the geometries of laser machined micro-channels. The advantages and limits of different machine learning methods were discussed including KNN, neural networks, DT and linear regression model for laser milling considering different process outputs. It was found that both neural networks and DT had better accuracy than the other two models. DT has better accuracy for material removal rate modeling, while neural networks perform better on depth prediction. Maudes et al. (2017) adopted different machine learning methods for optimization of the laser microfabrication process of the stent. Linear SVM method was shown suitable for identifying the good and bad cutting conditions. Random Forest ensembles of the regression trees got best performance for geometries prediction among other machine learning methods. Jain et al. (2019) used an ANN method to predict the laser drilled hole circularity and heat affected zone. A multi-objective genetic algorithm was applied to optimize the process and to determine the suitable laser processing parameters. Alizadeh and Omrani (2018) applied a Taguchi method to design the laser cutting experiments. The kerf width and kerf taper were measured and ANN method was applied to predict partial experimental results. An optimal robust data envelopment analysis method was applied for the best quality parameters, with experimental verification. For the laser micro-grooving process, Dixit et al. (2019) reported a combined method of design of experiments (DOE), response surface method (RSM), and ANN together with a genetic algorithm (GA) to predict and optimize the upper width, lower width and depth of micro-grooves. Parandoush et al. (2015) used a finite element method (FEM) to model laser grooving process. Subsequently, fuzzy logic method was applied to predict the modeled groove depth. There are also increasing researches on the application of deep learning methods in laser machining, which are typically for image analysis and predictive image generation. Oh and Ki (2019) applied CNN and cGAN method to predict the hardness of laser treated sample. The hardness distribution image was generated through the model based on a 3D simulated temperature distribution image. The average prediction accuracy was up to 94.4 %. Sun et al. (2020) adopted a LeNet-5 network to predict the cleanliness of substrate after laser cleaning. Chen et al. (2019) proposed a deep generative network to predict the laser-browned dough. Photorealistic images of browned dough could be predictively generated with the model. Mills et al. (2018) applied a cGAN method to predict laser machined surface topography (SEM images). With laser spatial intensity profile as input images, the SEM image of laser machined sample can be predictively generated. The generated images are shown well consistent with experimental SEM images. They further reviewed the application of machine learning methods including deep learning in laser machining (Mills and Grant-Jacob, 2021). There are increasing research on the applications of ANN, CNN and GAN method. Zhang et al. (2020) reported a deep learning method for feature extraction of laser machining data. The AlexNet with multi-task learning shows a better performance than ResNet and single-task model.

For LIPSS fabrication, most of the current research focuses mainly on exploring the formation mechanisms. A significant amount of experimental processing information has not been utilized for process prediction and optimization. The application of machine learning for intelligent, simple and fast fabrication of LIPSS is not well explored. Thus, we proposed here a hybrid machine learning method in our attempt to determine the optimized processing window of LIPSS fabrication process. The proposed method includes experimental sampling, dimensionality reduction, data clustering, classification and decision boundary determination. The three-dimensional (3D) design space of laser processing parameters was explored through mean sampling experiment design method. LIPSS SEM image features were extracted by applying a transfer learning method. Subsequently k-means clustering method was used to automatically cluster and label samples as good and bad quality classes. Different machine learning methods were applied to predict the fabrication results. Finally, an optimized operating window was determined.