Defect prediction is critical for advancing semiconductor manufacturing, particularly in optical lithography, extreme ultraviolet lithography (EUVL), and nanoimprint lithography (NIL). A recent study from National Taiwan University (NTU) offers new insights into predicting and mitigating defects across these three techniques, addressing key challenges for next-generation chip production.
Why Defect Prediction Matters in 2026
As the semiconductor industry pushes toward smaller nodes—down to 3 nm and beyond—defect control becomes increasingly difficult. In 2026, the adoption of high-NA EUV and advanced patterning processes heightens the risk of stochastic defects, which are random in nature and hard to predict. Likewise, NIL's growing use in emerging applications like advanced packaging and quantum devices demands robust defect management. NTU's research aligns with these trends, proposing methods that leverage machine learning and physical modeling to improve defect prediction accuracy.
Optical Lithography: Enhancing Process Window Optimization
In conventional optical lithography (using 193 nm immersion systems), defects often arise from focus variations, exposure dose errors, or mask defects. NTU researchers have developed predictive models that combine aerial image simulation with machine learning to forecast defect-prone regions on wafers. By analyzing process window parameters—such as depth of focus (DOF) and exposure latitude (EL)—their approach supports real-time process control, reducing yield loss in high-volume manufacturing. For 2026, these models are being adapted for multi-patterning techniques, where defect interactions across layers complicate prediction.
EUVL: Tackling Stochastic and Mask Defects
Extreme ultraviolet lithography (13.5 nm wavelength) introduces unique defect sources, including stochastic effects (e.g., photon shot noise), mask blank defects, and tin debris from the source. NTU's work emphasizes machine learning-based classification of defect signatures, using data from actinic inspection tools to distinguish critical from non-critical defects. A notable innovation is the integration of deep learning with rigorous electromagnetic field simulations to predict how defects on reflective masks propagate to the wafer. In 2026, this method supports the shift to high-NA EUV, where mask defects are more magnified and require sub-nanometer precision.
NIL: Predicting Defects from Template and Resist Interactions
Nanoimprint lithography, which transfers patterns via mechanical contact, faces defects like trapped air bubbles, template contamination, and resist residue. NTU's research focuses on simulating the imprint process—considering resist viscosity, template geometry, and pressure—to predict defect formation. Their models use finite-element analysis with Bayesian optimization to identify optimal process parameters, minimizing defect density. In 2026, this is particularly relevant for roll-to-roll NIL in flexible electronics and for hybrid NIL-EUV approaches being explored to extend patterning capabilities.
Comparative Insights and Future Directions
The NTU study highlights commonalities among the three techniques: all benefit from predictive analytics, but each requires tailored models due to differing physical mechanisms. Optical lithography relies on well-established imaging theory, EUVL demands stochastic-aware models, and NIL hinges on fluid dynamics. The research suggests that transfer learning could enable models to be adapted across techniques, accelerating development.
Future efforts will focus on integrating these predictive models into smart fabs, where real-time data from metrology and sensors feeds back into production systems. With the rise of AI-driven manufacturing in 2026, NTU's methods are poised to become essential tools for achieving high yield at advanced nodes.
Conclusion
National Taiwan University's work on defect prediction represents a significant step forward in semiconductor process control. By applying advanced modeling and machine learning to optical lithography, EUVL, and NIL, the research addresses current yield challenges and lays groundwork for future technologies. As the industry continues to evolve, such predictive capabilities will be vital for maintaining pace with Moore's Law.