Simulation-based deep learning can help photomask and semiconductor teams address a basic challenge: training neural networks requires diverse data, but real manufacturing records may not contain enough examples of rare defects or unusual conditions. For applications involving mask and wafer shapes, the training process also needs to represent curvilinear geometries and relevant manufacturing effects. GPU-accelerated simulation offers a way to generate this data at scale for research and development.
A kit introduced by D2S, TrueMask DLK, was designed to bring together computational design tools, mask and lithography simulators, and pre-trained neural networks for semiconductor manufacturing R&D. Its approach illustrates how simulation, rather than production data alone, can support deep learning development for complex photomask and wafer applications.
Why simulation-based deep learning matters in semiconductor manufacturing
Deep learning is a form of machine learning that identifies patterns in large datasets and learns transformations from examples. In semiconductor design and manufacturing, potential applications depend on networks learning how geometric features and process conditions relate to observed outcomes.
Training a network to perform reliably requires examples spanning a broad range of configurations. Real mask and wafer data is valuable, but it may not provide enough instances of uncommon defects or anomalous conditions. Capturing thousands of scanning electron microscope (SEM) images can also be impractical, and naturally occurring defect events may be too limited to represent the full range of cases a model should encounter.
Simulation can complement measured data by generating training examples for conditions that are rare in manufacturing records. That makes the quality and relevance of the simulation models important: synthetic examples need to reflect the geometries and physical effects that matter to the intended application. In mask and wafer workflows, this includes support for curvilinear data rather than relying solely on simplified shapes.
Curvilinear data and full-chip processing
Curvilinear mask shapes can represent complex patterns used in advanced photomask designs. The source describes real mask and wafer data as curvilinear and identifies curvilinear processing as a requirement for simulation-based training data. A training workflow that cannot represent these shapes may miss relevant geometry in the examples used to teach a neural network.
Processing this data across an entire chip can be computationally demanding. GPU acceleration is intended to make full-chip curvilinear simulation faster, enabling the generation and processing of larger or more varied datasets for model development. GPUs and other single instruction, multiple data (SIMD) devices perform many similar operations in parallel, a capability used by deep learning workloads and by the computational tools that prepare training data.
For engineering teams, the practical consideration is not simply whether a workflow uses a GPU. It is whether the simulation represents the geometries and process effects relevant to the target task, and whether the data can be generated at a useful scale.
Combining mask and lithography models
TrueMask DLK was presented as a combination of D2S computational design software and simulation capabilities from partner companies. The package included the D2S TrueMask model-based mask data preparation solution, with native curvilinear mask and lithography simulators, geometry libraries, and verification checks. It also offered additional simulation options: rigorous e-beam simulation from aBeam Technologies and ArF and EUV lithography simulation, including mask 3D effects, from Fastlitho.
These components address different parts of the data-generation challenge. E-beam simulation relates to mask fabrication, while ArF and EUV lithography simulation can represent how mask patterns interact with lithographic processes. Including mask 3D effects in the lithography simulation is relevant when those effects need to be reflected in training examples. The appropriate model depends on the application and the manufacturing behavior the neural network is expected to learn.
What a deep learning kit can provide R&D teams
A packaged environment can bring together software and models that teams would otherwise need to assemble for a training workflow. The capabilities described for TrueMask DLK included:
- GPU-accelerated, model-based mask data preparation for complex masks.
- Curvilinear mask and lithography simulation, along with geometry libraries and verification checks.
- Optional rigorous e-beam and ArF/EUV simulation capabilities from partner companies.
- A pre-trained neural network model for mask and wafer fabrication.
- A D2S Computational Design Platform based on the sixth-generation platform and NVIDIA Tesla V100 GPU accelerator.
The Center for Deep Learning in Electronics Manufacturing (CDLe), a partnership of NuFlare, Mycronic, and D2S, was identified as an early adopter. Its director emphasized the role of simulation in creating examples of unusual situations that may be scarce in real manufacturing data.
A pre-trained model can provide a starting point, but manufacturing organizations still need to consider their own data and objectives. The kit was described as a quick-start approach intended to help companies ramp up deep learning using their data. That framing highlights an important distinction: a general training resource does not remove the need to assess whether the examples and model behavior fit a specific process or use case.
Practical considerations for semiconductor teams
For system integrators, process engineers, and R&D leaders evaluating deep learning, several questions follow from this approach:
- Are rare conditions represented? Identify where production data is sparse and whether simulation can generate useful additional examples.
- Does the simulation reflect relevant geometry? Check that curvilinear patterns and applicable mask or lithography effects are represented.
- Can data be processed at the required scope? Assess whether the computing platform can support the intended full-chip simulation workflow.
- How will the model relate to company data? Define how simulated examples, measured data, and any pre-trained network will be evaluated for the target application.
Conclusion
Deep learning for photomask and semiconductor manufacturing depends on more than neural network architecture. It also requires training data that captures relevant geometries, process effects, and uncommon conditions. Simulation can extend limited real-world datasets, while GPU acceleration can support full-chip curvilinear processing. Integrated kits such as TrueMask DLK bring these elements together as a starting point for R&D teams, but the value of the approach rests on choosing simulations and data that match the manufacturing problem at hand.