TL;DR
This paper introduces a novel method for designing masks for double-exposure lithography (DEL) using inverse lithography (ILT). The approach automatically synthesizes two masks that, when used in sequence, can print complex patterns with improved contrast and resolution, overcoming limitations of single-exposure techniques.
What Problem Does This Paper Address?
Modern semiconductor manufacturing requires printing increasingly intricate patterns at ever-smaller scales. Traditional lithography techniques struggle with these demands, particularly when dealing with "phase conflicts" – situations where a single mask cannot simultaneously satisfy the optical requirements for different parts of a pattern. Double-exposure lithography (DEL) offers a solution by using two masks and two exposures to build up the final pattern, effectively splitting a complex design into two simpler, more manageable components. However, designing these two masks optimally is a significant challenge. Inverse lithography (ILT) is a powerful technique for designing masks by working backward from the desired wafer pattern. This paper tackles the problem of combining ILT with DEL to automatically generate the necessary masks for DEL systems, aiming to achieve high-fidelity pattern printing with improved contrast and resolution.
Methods
The authors propose a pixel-based inverse lithography framework for double-exposure lithography. Their approach involves:
- Forward Model: They define a mathematical model that describes how two input masks, when exposed sequentially, combine to form the final aerial image on the wafer. For simplicity, they initially focus on a coherent imaging system.
- Optimization Problem: The core of the method is an optimization problem that seeks to find two masks (represented as pixel arrays) that minimize the difference between the desired target pattern and the pattern produced by the DEL system.
- Gradient Calculation: They analytically calculate the gradient of the cost function with respect to the mask parameters. This is crucial for efficient optimization.
- Optimization Algorithms: Two optimization algorithms are explored:
- Simultaneous Descent: Updates both masks simultaneously in each iteration.
- Cyclic Coordinate Descent: Updates each mask independently in a sequential manner. This method is found to be more efficient.
- Mask Synthesis: The optimization process yields continuous-tone masks, which are then thresholded to produce physically realizable chromeless phase-shift masks (PSM). These masks have transmission values of either -1 (180-degree phase shift) or +1 (zero-degree phase shift).
- Regularization: A regularization framework is mentioned as a way to handle the ill-posed nature of the problem and incorporate prior information, such as promoting desirable mask properties like tolerance to alignment errors or ease of manufacturing.
Datasets
Not discussed in the paper. The paper focuses on synthesizing masks for given target patterns, rather than training on a dataset of existing patterns.
Key Insights and Ideas
- Automatic Pattern Decomposition: The DEL approach, guided by ILT, naturally decomposes complex target patterns into two sparser sub-patterns. This decomposition is not arbitrary but seems to aim at increasing feature pitch, reducing frequency content, and decreasing density, making the individual exposures easier to print.
- Resolving Phase Conflicts: The method effectively resolves phase conflicts that plague single-exposure lithography. For instance, in the "T-joint" example, the second exposure fills in the gap at the intersection that would otherwise be a "hot-spot" in a single-exposure scenario.
- Superior Contrast and Reduced Hot-spots: The final aerial images produced by the DEL masks exhibit superior contrast and a notable absence of hot-spots compared to patterns generated by single-exposure ILT.
- Multiple Solutions Exist: The ill-posed nature of the inverse DEL problem means that multiple pairs of masks can achieve a similar desired output. The specific decomposition can depend on the optimization algorithm and regularization used.
- Cyclic Coordinate Descent Efficiency: The cyclic coordinate descent algorithm demonstrates faster convergence than simultaneous descent for this problem, making it a more superior choice for searching the solution space.
Key Results
- Successful Synthesis of Chromeless PSMs: The framework successfully synthesized two chromeless phase-shift masks for various target patterns, including a T-joint, a U-joint, and a more complex pattern.
- Improved Aerial Images: The resulting aerial images showed significantly better contrast and uniformity compared to single-exposure ILT for the same target patterns. For the T-joint, the DEL approach eliminated the hot-spot at the junction.
- Faster Convergence: The cyclic coordinate descent method converged faster than simultaneous descent, as illustrated by the cost function behavior plots.
Limitations and Future Work
- Coherent Illumination Assumption: The presented results are for an idealized coherent imaging system (σ=0). Extending the framework to partially coherent imaging systems with different source configurations is a key area for future work.
- Alignment Tolerance: The current framework does not explicitly account for critical alignment steps between the two exposures. Future work should focus on using the regularization framework to synthesize masks that are robust to alignment errors.
- Optimization Algorithms: While cyclic coordinate descent is effective, exploring more advanced optimization algorithms like Conjugate Gradient or Quasi-Newton methods could further improve search efficiency.
- Physical Realizability: While the masks are thresholded to be discrete, further investigation into other desirable properties like low complexity, ease of manufacturing, and low MEEF (Mask Error Enhancement Factor) is ongoing.
- Generalization: The paper presents preliminary results, and further validation with more complex patterns and realistic imaging conditions would be beneficial.
Summary updated