First Steps towards Machine Learning for Prediction and Pre-Correction in Direct Laser Writing

TL;DR

This work tackles the fidelity gap in direct laser writing by developing CNN-based neural networks that predict and pre-correct prints from 2PP/DLW-enabled designs. It compares Iterative Correction and Direct Correction strategies, trained on large experimental-theoretical datasets of 2.5D structures, and demonstrates substantial improvements in print conformity for several structures while highlighting artefact-related limitations. Incorporating DLW parameters into the models preserves high correlation but can increase RMSE and surface roughness, underscoring a trade-off between correction accuracy and artefact control. The results suggest neural-network pre-correction can augment conventional methods and may generalize to other 3D printing technologies, enabling more complex, high-precision microstructures in the near future.

Abstract

Additive manufacturing using 2-Photon Polymerization (2PP, aka direct laser writing DLW) enables the fabrication of almost arbitrary complex 3D structures from the meso to the submicron scale. However, deviations between the anticipated target structure and the actual print often occur due to physico-chemical processes, limiting the accuracy and reliability of this technology. To minimize these deviations, we hereby present our latest research in developing different neural networks, targeting the above-mentioned aspect. Our networks are trained on several experimental as well as theoretical datasets and show good results in predicting fabrication deviations and (pre-) correcting 2.5D micro-structures. Hence, we demonstrate, that besides conventional iterative correction methods, neural networks are a promising alternative to significantly improving the output quality in DLW. Furthermore, there are no fundamental limitations to transferring this machine learning approach to other 3D printing technologies, as they all face the same challenge in terms of fidelity. To our point of view, the use of neural networks has the potential to enhance the capabilities of this technology, enabling the creation of complex structures with increased accuracy and precision in the near future.

Figures (11)

• Figure 1: Schematic visualization of the network architectures. (a) General architecture of the neural networks with only convolutional layers. Optionally, e.g., pooling layers were implemented as layer substitution by a convolutional, an average pooling, and a deconvolutional layer, as shown in the green inset. (b) Iterative Approach network for structure prediction including the three DLW parameters slicing, hatching, and laser power by a second branch. Therefore, also fully connected layers (FC) are used. Since the FC-outcome can be negative, a hyperbolic tangent is used as activation function, whereas all other layers still use rectified linear unit (ReLU) activation. (c) Direct Approach architecture, designed to correct for the printing deviations and predicting a corrected design with corresponding writing parameters. The field 'CNN' represents the network shown in (a).
• Figure 2: Neural network prediction. Comparison of an ideal radially symmetric $\text{CIN}_{\text{r}}$ type structure (a) with its corresponding 3D print (b) and the prediction of the neural network (c). The network was trained using mean square error as loss function. Profiles along the x-axis and the deviations are shown in (d) and (e), respectively.
• Figure 3: Exemplary neural network predictions. The target designs (a, d, g) are compared with the respective measured prints (b, e, h) and the respective network's predictions for a block (a-c), an AIR-type structure (d-f) and the logo of the former University of Kaiserslautern (g-i). For the logo, (j) and (k) show profiles along the $x$-axis and the corresponding height difference $\Delta z$ between print and target, as well as print and prediction. The estimated repetition accuracy of 500 nm (see Sec. \ref{['subsec:dataprep']}) is shaded in gray. The inset shows a table of the corresponding $rmse$ and correlation coefficients $CC$ to quantify the respective accuracies. The neural network was trained with 628 datasets.
• Figure 4: Results of the Direct Correction (DC) approach. Examples for the direct design correction of a network trained with 628 datasets. Shown are the corrections of a block (a)-(d), the logo (e)-(h), which was not included in the training data, and a $\text{CIN}_{\text{r}}$-type structure (i)-(l) including profile information (m)-(p).
• Figure 5: Artefacts illustration. Exemplary appearance of a random artefact during the correction of an AIR-type structure (a). While the corrected design (c) of the neural network shows two peaks in the middle of the structure, only one of those is visible in the measurement of the print (d).