Ultrafast laser-matter interaction involves multiple physical phenomena at different time scales. Consequently, process development for ultrafast laser processing is also a lengthy, empirical process. Different scientific models provide valuable insights on the underlying physics but are often too complex for practical use.
More recently, machine learning has proven to be very effective in predicting and optimizing micro-processing results. However, to take full advantage of these algorithms, an important number of data points are needed for training purposes. Acquiring such a dataset usually requires a significant amount of time, partially negating the benefit of machine learning.
The purpose of this work is to study the efficiency of machine learning algorithms to predict the results of a femtosecond laser engraving process, using only a small training dataset describing the engraving depth for different materials, process parameters and laser specifications. We compare the results with an engineering model based on the well-known two-temperature model, present strategies to mitigate the dataset size and compare the results with independent experimental results.
|