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

summary

The gist

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 Our…

In short

The paper explores using neural networks to improve 2-Photon Polymerization (2PP) fabrication accuracy for complex 3D structures. It demonstrates that AI can predict structural deviations and correct designs, offering a powerful alternative to traditional iterative correction methods to significantly enhance output quality.

Key concepts

Direct Correction (DC)
This approach uses the neural network to directly predict a corrected design from an initial target structure. The goal is for the network to output a design that, when printed, will be as close as possible to the desired final structure without needing multiple manual correction steps.
Iterative Correction (IC)
This method involves repeatedly using the neural network to suggest improvements. Each suggested correction is then printed in a single job under consistent conditions. This allows for quick iteration and comparison of results, although it may not always preserve the original structure's symmetry.
Customized Loss Function
The loss function used during training is tailored to prioritize specific features. It combines the Pearson correlation coefficient (CC) with the absolute difference in mean height values. This setup ensures that small surface features are given greater importance during the learning process, leading to better final print quality.
Network Architectures
The researchers tested three different neural network architectures, each optimized for a specific task—either predicting deviation or performing direct correction. The choice of architecture depends on whether the network is trying to forecast errors or generate a fully corrected design.

Terminology used across episodes

This episode discusses

The paper

First Steps towards Machine Learning for Prediction and Pre-Correction in Direct Laser Writing · Read on arXiv

Physics Department and Research Center OPTIMAS, RPTU University · Opti-Cal GmbH · Fraunhofer Institute for Industrial Mathematics ITWM

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.

DOI: 10.37188/lam.2026.128

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "First Steps towards Machine Learning for Prediction and Pre-Correction in Direct Laser Writing".

Mira: Here's the gist of the paper. 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 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 µ-structures Hence, we demonstrate, that besides conventional iterative correction methods, neural networks are a promising alternative to significantly improving the output quality in DLW.

Kai: Problem Context The major challenge in this field is the – mostly undesired – deviation between the envisaged target structure and its corresponding, experimentally 3D printed counterpart While these deviations are negligible for large structures (e.g.

Mira: those in the cm-range), they become more significant as the size of the prints decreases.

Kai: Unfortunately, this is fundamentally unavoidable with today’s photo resins.

Title and authors: Kai: So we're looking at this paper called "First Steps towards Machine Learning for Prediction and Pre-Correction in Direct Laser Writing". The title itself points directly to their main goal, which is using machine learning for both predicting what will happen and correcting it before you even start printing.

Mira: And the authors are Sven Enns, Julian Hering-Stratemeier, and Georg von Freymann. They come from the Physics Department and Research Center OPTIMAS at RPTU University Kaiserslautern-Landau.

Lev: I've seen their work on error correction methods before, so I'm interested in how they are framing this as a prediction and pre-correction problem rather than just a post-printing fix.

Kai: It seems they are moving beyond just simple adjustments and aiming for something that can generate a pre-compensated version of the design that matches the original geometry as closely as possible.

Mira: That's what they are targeting, trying to get the print to reproduce the original geometry with maximum accuracy, not just fixing errors after the fact.

The paper's summary: Kai: So they summarize their work by saying that while deviations between target structures and printed ones happen in direct laser writing because of physical processes, neural networks show good results when trained on experimental and theoretical datasets.

Mira: They are focusing on predicting the target print deviation first, which is one approach called predicting the optimal exposure parameters for two photon polymerization based on scanning electron images.

Lev: So they're trying to use machine learning to figure out what settings you need before you print so the resulting structure is closer to your original design.

Kai: They also introduced a second approach called Direct Correction, where the network directly predicts a corrected design, which they are testing with three different network architectures based on their specific tasks.

Mira: And they use a customized loss function for evaluation that includes the Pearson correlation coefficient and the absolute value of the difference between the mean height value of prediction and label.

The paper's improvements: Lev: They suggest a few things for improvement, like using iterative correction methods, which show convergent behavior towards higher correlation as you iterate through steps.

Kai: But they admit that iterative correction doesn't always preserve the symmetry of the structure, and it needs more effort than the direct correction approach because the real error doesn't always converge with more iteration numbers.

Mira: The authors also mention that their improved AI system can perform structure correction by calculating a deviation matrix to correct tilt within measurement data, which involves calculating a regression plane.

Lev: They also incorporated fabrication parameters like slicing distance, hatching distance, and laser power into the learning and prediction process so the network can account for those influences on printing results.

Kai: And they use a customized loss function that prioritizes small surface features by including a term involving the Pearson correlation coefficient, which is implemented in such a way that even small surface features are given greater importance during training.

Conclusion: Mira: So to wrap up, the paper shows promising attempts using neural networks for direct structure prediction and pre-compensation in direct laser writing.

Kai: They show that while iterative correction works well for finding good designs, the direct correction approach is a strong alternative because it directly predicts a corrected design.

Lev: From my side, I think the main challenge they tackle is moving from simple structures to more complex ones and ensuring those corrections are robust enough for real hardware.

Kai: They also flag that while they show good results, the transferability of their methods to completely arbitrary DLW geometries isn't fully proven yet.

Mira: So, overall, it's a first step towards using machine learning to make two-photon polymerization more accurate and reliable for complex three dee structures <ref:2511.23306#pg1>.

Lev: It sets up a clear path for future work on improving the neural networks further and extending their applicability to different structure types.

More episodes

← Home