image: Figure 1| Basic architecture of the Convolutional Neutral Networks. a Basic network architecture without fabrication parameters showing the different network layers including their corresponding amount and sizes of applied filters. b Including fabrication parameters to predict the deviation of the print to its design. This prediction can be used for design correction. c Direct correction of the CNN with a corrected design with corresponding fabrication parameters, when given the target design as input.
Credit: Sven Enns et al.
Direct laser writing (DLW) employing two-photon absorption allows to fabricate almost arbitrary 3D structures down to the sub-micron scale. DLW is one of the fastest techniques for laser-based additive manufacturing reliably offering feature sizes down to 100 nm. This performance is realized by focusing a femto-second pulsed laser beam into a photo-sensitive material (photo resin). Two photon absorption triggers a small volume in the very focus of the laser beam to polymerize. The fabrication of three-dimensional structures is achieved through a relative movement between laser focus and resin. While feature sizes in the order of ~100 nm can be achieved, the deviations between designed and printed structure can reach up to 30%, amounting to tens of microns in worst-case scenarios. Although many of the chemico-physical processes leading to deviations are known, there is a lack of literature on predicting or calculating the effect of these processes on a specific structure. In order to achieve more precise structures, one common approach involves iterative manual corrections of the design according to the observed deviations. However, this approach is time consuming, leading to considerable costs. Identifying more effective methodologies for accounting for deviations is thus desirable.
In the recent paper “First steps towards machine learning for prediction and pre-correction in direct laser writing“ published in Light: Advanced Manufacturing, a group led by Professor Georg von Freymann from RPTU University Kaiserslautern-Landau and Leibniz University Hannover in Germany presents initial findings from employing machine learning algorithms for the prediction of manufacturing deviations. Compared to the manual approach, which demands printing and measuring of each iteration, the neural network prediction can directly correct the design prior to printing, thereby considerably reducing time and resources required.
To enable the computer to learn the various deviation mechanisms, convolutional neural networks (CNNs) have been trained. CNNs are particularly suitable for grid-like data, such as images, and are frequently employed for tasks such as image classification and object and feature detection. These capabilities are essential to the presented approach. The training datasets, which consist of pairs of design and measurement data, are presented to the network to enable a supervised learning process. This process specifically allows for mapping between the design and the corresponding measurement. It also enables the network to comprehend the impact of deviation mechanisms on the printed output. This knowledge about the deviation can then also be used to correct the initial design, so that the printed structure is as close to the target as possible.
The generation of training data is arguably the most crucial aspect of this approach. Achieving optimal results for all potential printed structures requires extensive training data. The current project utilizes 2.5D, which refers to structures where only the surface is modulated, and the most impactful DLW parameters. However, the results are very promising, and the final goal is to extend the models to three-dimensional structures with all fabrication parameters.
In the current state of the project, the neural network has the potential to reduce deviations by up to 90% for certain benchmark structures.
These results demonstrate the potential of machine learning as a tool to enhance the fabrication outcomes in DLW, making further development of the models a worthwhile effort.
Journal
Light: Advanced Manufacturing
Article Title
First Steps towards Machine Learning for Prediction and Pre-Correction in Direct Laser Writing