Two-photon polymerization (TPP), a form of nanoscale 3D printing using ultrafast laser pulses, can create structures with sub-micron detail, but getting prints to match the intended design is challenging1. The process involves complex physics: nonlinear light absorption, chemical reactions, and diffusion effects that often cause features to come out too small, too large, or distorted, especially when trying to print many features in parallel. What if we could let a computer figure out how to adjust for these effects automatically? That’s exactly what FABulous researchers from Fraunhofer IISB (Germany) and IMT Atlantique (France) accomplished in a recent study.
The researchers developed a time-dependent, differentiable forward model of the TPP process. In simple terms, it’s a physics-based simulation of how a laser builds a microstructure, but built in a machine-learning framework (PyTorch) so that it not only simulates the process but also provides gradients, information on how changes in laser power or timing would change the final printed shape. This differentiability is key: it means the model can be used for gradient-based optimization, essentially letting an algorithm tweak the laser settings via backpropagation (the same technique that trains neural networks) to achieve a desired output shape.
To showcase this, the team applied their model in three different TPP printing setups:
- Point-by-point laser writing: a conventional approach where a laser spot scans the design.
- Diffractive Optical Element (DOE) parallel exposure: a special optic splits the laser into many beams to print multiple features simultaneously.
- Spatial Light Modulator (SLM) projection: a programmable screen creates a complex light pattern, projecting an image into the resin.
In each scenario, the differentiable model enabled smarter control of the printing process. For instance, in the parallel writing test the simulation predicted that the ends of printed lines would form with less height due to oxygen-induced quenching of the polymerization reaction. By recognizing this, the algorithm automatically increased the laser dose at those line ends, compensating for the effect and producing uniform line heights. In the SLM projection case, which used a challenging low-contrast photoresist, the system combined a neural network and the physics model to find an exposure pattern that successfully produced tiny microlenses and free-form shapes true to the target design. Even the basic point-by-point writing benefited: the laser power was modulated on the fly to ensure each tiny voxel (volumetric pixel) achieved the intended height and that separate written features fused correctly into a continuous 3D form. Across all these experiments, the model-driven optimization improved the fidelity of the printed microstructures compared to using uniform or hand-tuned exposure settings.
Overall, this work represents a significant step toward automated, high-accuracy microfabrication. By merging detailed physics with modern optimization techniques, the researchers have created a tool that can perform “inverse lithography” for TPP, in that it is possible to input the thing you want to make, and the system computes how to make it. This approach can greatly reduce trial-and-error in crafting microdevices and nanostructures. The study, published in March 2026 in Journal of Micro/Nanopatterning, Materials, and Metrology, not only demonstrates impressive results in the lab but also lays a foundation for 3D printing at the microscale.
