Machine Learning Techniques for Lithography Optimization

Machine learning is transforming lithography optimisation by offering data-driven alternatives to labour-intensive physical simulations and manual parameter tuning. These techniques address key design-for-manufacturability challenges, including hotspot detection, optical proximity correction, mask synthesis and process-parameter calibration. Convolutional neural networks learn complex non-linear mappings between layout geometries and printability metrics, enabling rapid identification of defect-prone regions. Dimensionality-reduction algorithms and autoencoder networks furnish compact representations of lithographic patterns, streamlining model calibration and inverse lithography for mask generation. Transfer learning and synthetic pattern generation mitigate data scarcity by enriching training sets with artificial or simulated examples, bolstering performance on previously unseen designs. Regression-based and tree-based models approximate outputs of computationally expensive lithography simulations, facilitating real-time process control. Multi-task learning frameworks jointly optimise classification and localisation objectives, reducing false alarms while enhancing detection accuracy. These approaches have driven improvements in critical-dimension fidelity, yield enhancement and throughput acceleration, supporting sub-10 nm technology nodes and emerging direct-write lithography. Hybrid workflows that integrate physical models with data-driven methods are gaining traction, balancing accuracy, interpretability and computational efficiency within advanced semiconductor manufacturing toolchains.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Multi-task deep-learning frameworks have recently demonstrated exceptional performance in lithographic hotspot detection by combining classification and localisation within a single model. By employing synthetic pattern generation to augment limited datasets, these approaches have improved detection of never-seen-before hotspots and reduced false-alarm rates, achieving accuracy levels above 98 % with marked reductions in spurious alerts. In scanning probe lithography, an intelligent machine learning framework has been introduced for in-situ characterisation and process-parameter optimisation, enabling real-time extraction of global feature statistics and automated adjustment of fabrication parameters. This has yielded finer critical-dimension control and scalable direct-write nano-patterning over large areas. Complementing defect detection and process tuning, adaptive pattern-representation methods have been applied to optical proximity correction (OPC) model calibration. Comparative studies of autoencoder-based and Fourier-based feature vectors reveal that domain-augmented autoencoder representations more effectively capture lithographic variabilities, enhancing pattern selection for OPC model training and accelerating calibration convergence. Together, these contributions illustrate a unified trend towards integrated, data-driven lithography workflows that harness synthetic data, representation learning and multi-objective optimisation to meet the exacting demands of next-generation semiconductor fabrication.

Machine Learning Techniques for Lithography Optimization publication trend

The graph below shows the total number of articles in machine learning techniques for lithography optimization across all publications each year (not limited to Nature Index journals).

Technical terms

Lithographic hotspot: A layout region prone to printing defects caused by optical or process proximity effects, leading to critical-dimension deviations.

Optical proximity correction (OPC): A technique that adjusts mask patterns to counteract distortions introduced by the imaging system, ensuring accurate feature reproduction.

Convolutional neural network (CNN): A deep learning architecture that applies convolutional filters to extract hierarchical spatial features from image-based inputs such as layout patterns.

Autoencoder: A neural network that learns compact feature representations by encoding input data into a lower-dimensional space and reconstructing it with minimal error.

Synthetic pattern generation: The creation of artificial layout samples using algorithmic or simulation-driven methods to augment training datasets and improve model generalisation.

References

1. Enhanced Lithographic Hotspot Detection via Multi-Task Deep Learning With Synthetic Pattern Generation. IEEE Open Journal of the Computer Society (2024).
2. Towards smart scanning probe lithography: a framework accelerating nano-fabrication process with in-situ characterization via machine learning. Microsystems & Nanoengineering (2023).
3. Intelligent Photolithography Corrections Using Dimensionality Reductions. IEEE Photonics Journal (2019).
4. Feature Vector Effectiveness Evaluation for Pattern Selection in Computational Lithography. Photonics (2024).
5. Lithography Hotspot Detection Method Based on Transfer Learning Using Pre-Trained Deep Convolutional Neural Network. Applied Sciences (2022).