Detection and correction of complex system properties in wave front shaping using neural networks

• Applicant: Professor Dr.-Ing. Michael Schmidt
• Subject Area: Joining and Separation Technology
• Term: since 2018
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 397970984

Wave front shaping with liquid crystal displays (SLM from Spatial Light Modulator) enables tailoring arbitrary intensity profiles, which increase the ablation quality and efficiency in ultrashort pulse laser materials processing. However, pure wave front shaping also causes the formation of so-called speckle, a light/dark contrast within the target distribution, because only a part of the light field - in this case the wave front, but not the amplitude - is shaped. In the first funding phase, methods for high-precision and uniform beam shaping were developed. In addition to reducing speckle, these methods maintain the high light efficiency of the liquid crystal display and the small amount of non-diffracted light vanishes which would otherwise overlay the target structure. In theory, all those methods give perfect results. However, this is practically not realisable as several process parameters cannot be controlled exactly: A non-perfect calibration of the Look-up-Table, Cross-Talk of adjacent pixels, and system-specific distortions like aberrations severely affect the beam-shaped result. In order to provide an accurate representation of the desired target distribution, these complex system properties must be precisely characterized, which in turn allows them to be included in the calculation of the phase mask. As the required experimental characterization and modeling is highly time-consuming and requires enormous computational effort, machine learning will be used instead. The overall objective is to develop a fully interpretable neural network, which generates phase masks for beam shaping on basis of experimental training data. Those phase masks should directly incorporate system-specific peculiarities of the experimental setup such as aberrations, as well as characteristics of the spatial light modulator, as for example Cross-Talk or the calibration of the resulting phase shift, which makes the cumbersome calibration of the system no longer necessary. Accordingly, the neural network should reflect physical properties and thus also allow for conclusions to be drawn on the experimental setup. To improve transparency and interpretability of the neural network, physical relationships that can be modelled well should be directly included (e.g. beam propagation). Based on a general pre-trained network, independent experimental setups should be able to be re-trained with only a few measurement data to obtain high quality results with only little effort and without detailed process knowledge.

• DFG Programme: Research Grants