TLDR: Researchers have developed a groundbreaking AI model, the Waveguide Neural Operator (WGNO), to revolutionize the simulation of extreme ultraviolet (EUV) light diffraction in lithography masks. This hybrid approach combines the accuracy of traditional physics-based methods with the speed of neural networks, drastically reducing computation time and training duration. The WGNO achieves state-of-the-art accuracy and inference speed, offering a highly efficient solution to accelerate the design workflows of advanced semiconductor devices.
The relentless march of semiconductor technology, often described by Moore’s Law, has pushed the boundaries of integrated circuit feature sizes down to the nanometer scale. At this minuscule level, Extreme Ultraviolet (EUV) lithography, using light with wavelengths around 13.5 nm, has become the cornerstone for manufacturing advanced semiconductor devices. However, at such tiny scales, the diffraction of electromagnetic waves from the photomask becomes a dominant physical effect, causing the pattern printed on the silicon wafer to deviate significantly from the original mask design.
To counteract these anticipated diffraction effects and ensure the desired circuit pattern is produced, processor manufacturing relies heavily on Optical Proximity Correction (OPC) techniques. A crucial step in the OPC process is the forward simulation: accurately and rapidly predicting how the electromagnetic field will distribute after diffracting from a complex, multi-layered EUV mask.
Traditionally, this simulation task has presented a trade-off. Simple models are fast but lack the necessary accuracy for modern manufacturing. Rigorous numerical solvers, such as the Finite Element Method (FEM) or the Waveguide (WG) method, offer high accuracy but are computationally very expensive, often becoming a bottleneck in the design process. More recently, deep neural networks like Convolutional Neural Networks (CNNs) have been proposed for their speed, but they typically require large datasets, extensive training times, and sometimes struggle to achieve the required accuracy and generalization.
A New Paradigm: Physics-Informed AI
A recent research paper introduces an alternative deep learning paradigm to address these challenges: Physics-Informed Neural Networks (PINNs) and a novel neural operator. Unlike traditional neural networks that rely on vast pre-computed datasets, PINNs are trained in an unsupervised manner, directly leveraging the governing physical equations during their training process. This eliminates the need for large, labeled datasets and allows the networks to learn the underlying physics.
Introducing the Waveguide Neural Operator (WGNO)
To overcome the limitations of both the computationally intensive WG method and the sometimes less accurate PINNs, the researchers propose a hybrid solution: the Waveguide Neural Operator (WGNO). The core innovation of WGNO is to replace only the most computationally expensive part of the traditional WG method—solving a large linear system—with a neural network. This multi-layer perceptron (MLP) is trained to learn the complex mapping from input parameters to the solution of this linear system.
The rest of the WG pipeline, including calculating waveguide modes and reconstructing the field, remains unchanged. This hybrid approach operates in a mesh-independent Fourier space, inheriting the robust physical structure of the WG method while directly targeting its primary computational bottleneck. The network is trained to minimize the discrepancy between its predicted solution and the physically consistent outcome.
Also Read:
- China Unveils ‘Meteor-1’ Optical Chip, Advancing AI and Challenging Traditional Computing
- Bridging Physics and AI: A New Approach to Predicting Complex Systems with Fourier Spectral Transformers
Remarkable Performance and Efficiency
Numerical experiments on realistic 2D and 3D lithography masks demonstrate the significant advantages of the WGNO. While PINNs showed reasonable accuracy, they struggled with more complex problems and required longer training times. In contrast, the WGNO achieved state-of-the-art accuracy, with errors several orders of magnitude smaller than PINNs, and remarkably short training times. For instance, in 2D mask simulations, WGNO achieved excellent accuracy with training times as low as 0.012 seconds, compared to thousands of seconds for PINNs. For the challenging 3D mask problem, the WGNO maintained high accuracy while achieving a speedup of over 200 times compared to the rigorous WG solver, with the entire training process taking only about 18 seconds.
This level of performance—high accuracy, extremely fast inference, and rapid training—makes the Waveguide Neural Operator a highly promising tool for accelerating the Optical Proximity Correction design cycle in semiconductor manufacturing. It offers a powerful solution for complex diffraction problems in EUV lithography simulation, potentially streamlining the development of next-generation microchips. You can read the full research paper here.
