Modeling pulsed laser micromachining of micro geometries using machine-learning techniques

Abstract

A wide range of opportunities are emerging in the micro-system technology sector for laser micro-machining systems, because they are capable of processing a various types of materials with micro-scale precision. However, few process datasets and machine learning techniques are optimized for this industrial task. This article aims to show how the process parameters of micro-laser milling influence the final features of the microshapes that are produced and aims to identify, at the same time, the most accurate machine learning technique for the modelization of this multivariable process. We studied the capabilities of laser micromachining by performing experiments on hardened steel with a pulsed Nd:YAG laser. Arrays of micro-channels were manufactured using various scanning speeds, pulse intensities and pulse frequencies. The results are presented in terms of the main industrial requirements for any manufactured good: dimensional accuracy (in our case, depth and width of the channels), surface roughness and material removal rate (which is a measure of the productivity of the process). Different machine learning techniques were then tested on the datasets to try to build high accuracy models for each output variable. The selected techniques were: k-Nearest Neighbours, neural networks, decision trees and linear regression models. Our analysis of the correlation coefficients and the mean absolute error of all the generated models show that neural networks are better at modelling channel width and that decision trees are better at modelling surface roughness; both techniques are similar for depth and material removal rate. In all cases these two techniques are more accurate than the other two. It can be concluded that decision trees can be used for modelling laser micro-machining of micro geometries, if the dimensional accuracy of the workpiece is the main industrial requirement, while neural networks are better in the other cases.

Artificial intelligence for automatic prediction of required surface roughness by monitoring wear on face mill teeth

“…RF has an even greater accuracy of between 33% for the complete dataset and 42% for the uncompleted dataset, with a statistically significant difference in both cases. Considering the standard deviation of the roughness in the dataset, 0.88 μm, the RMSE can be considered a little high (30% for the best model), although still acceptable from the industrial point of view compared with similar works (Bustillo et al 2011a, b;Maudes et al 2017;Teixidor et al 2015). Figure 4 shows the dataset prediction error for each instance (cross size) for the RF model to analyze this fact.…”

“…The last two regressors are often used as baseline methods for the purpose of comparison with the naïve approach and the regression model, as stated by previous works on surface roughness (Maudes et al 2017;Teixidor et al 2015). The naïve approach uses the mean value of the output as a prediction that is independent of the input values; in this study, the naïve approach will always predict a surface roughness of 1.20 μm (the mean roughness value considering the whole dataset).…”

“…Some complex manufacturing problems like laser polishing (Bustillo et al 2011b), maintenance planning of five-axis milling (Freiburg et al 2014), laser micromachining of cavities (Teixidor et al 2015), and deep drilling (Bustillo et al 2016) have been successfully modeled using regression trees. In all these cases, the decision trees create models that are as accurate as other standard machine-learning techniques, such as ANNs, and are suitable for both continuous and discrete outputs.…”

Experimental Optimization of Nimonic 263 Laser Cutting Using a Particle Swarm Approach

“…Pandey and Dubey [19] used pulsed Nd:YAG laser to cut the titanium alloy and showed that lower pulse width and pulse frequency and higher cutting speed result in a better cut. Texidor et al [20] investigated the laser microcutting, and Savriama et al [21] explained the novel pattering effects during the high-frequency laser microcutting of hard ceramics. Jarosz et al [22] presented effect of the cutting speed on the surface roughness and HAZ of stainless steel, while the other control factors were kept constant.…”

Semi-supervised roughness prediction with partly unlabeled vibration data streams

“…Importantly, the method allows robust integration of various supervised methods, already used in industrial settings, such as (k)Nearest Neighbors ((k)NN) and Random Forest. In particular, the work shows how techniques such as kNN, previously used for roughness modeling elsewhere (Teixidor et al 2015), can be used when partly unlabeled data becomes available.…”