A physics-informed AI model for on-demand design of complex optical coatings

The inverse design of multilayer optical coatings is a computationally intensive, non-convex optimization problem. We address this using a neural network framework trained to directly generate physical layer structures from target spectra. Our approach couples a ‘design’ network with a parameter-free, differentiable Transfer Matrix Method (TMM) model. During training, the network proposes a layer stack, the TMM model calculates its optical response, and the network’s parameters are adjusted to minimize the difference between this response and the target. This end to-end process allows the network to learn the physics of optical design directly without relying on any prior design data. Using this approach, we have generated a broadband dispersive mirror with performance characteristics comparable to those obtained by established commercial optimization tools. Building on this validated approach, we now show that this framework can be used to create a general, pre-trained model that rapidly generates initial designs for new problems, functioning as a powerful tool for optical coating design. Furthermore, the model architecture is inherently adaptable; it can be quickly fine-tuned on specific, user-defined design problems to further enhance its performance. This work demonstrates a practical pathway towards fast, generalizable, and continually improving AI tools for optical system design.

Content

Subjects

Optical Coatings

Automated Design

Neural Networks

Differentiable Optics

Ultrafast Lasers

Data-Free Design

DDC Class

600: Technology

005: Computer Programming, Programs, Data and Security

Funding References

Funding(s)

Helmholtz Graduate School for the Structure of Matter - Phase 2