TL;DR
MorphOPC introduces a novel multi-scale hierarchical framework that embeds learnable mathematical morphology operations (dilation and erosion) directly into neural networks for Optical Proximity Correction (OPC). By explicitly modeling geometric transformations rather than relying purely on statistical pixel mapping, it achieves state-of-the-art printing fidelity and mask manufacturability across both edge-based and pixel-based mask optimization benchmarks.
What Problem Does This Paper Address?
In advanced semiconductor manufacturing, optical lithography is used to print circuit patterns onto silicon wafers. However, as feature sizes shrink to the nanometer scale (well below the wavelength of the exposure light), optical diffraction causes severe distortions.
To combat this, engineers use Optical Proximity Correction (OPC) and Inverse Lithography Techniques (ILT) to modify the geometries on the photomask so the final printed pattern matches the intended design. Traditional optimization-based methods are incredibly accurate but computationally prohibitive for full-chip scales due to massive iterative simulation loops.
Recently, machine learning (ML) surrogates like GANs have been introduced to accelerate this process. However, standard deep learning models treat mask generation as a statistical pixel-mapping problem. They lack explicit geometric reasoning, frequently failing to capture shape-dependent local transformations—such as adding "hammerheads" to line ends or "serifs" to corners—which leads to suboptimal mask quality and poor generalization to unseen layouts.
Methods
The authors propose MorphOPC, a framework that reframes mask generation as a sequence of learnable morphological operations applied to local layout features.
Key components include:
- MorphBasic Block: Instead of standard convolutional layers that aggregate patterns via weighted summation, this block uses learnable non-flat morphological operators (max-plus algebra for dilation/expansion, min-plus algebra for erosion/shrinkage). A per-channel gating mechanism dynamically balances whether a channel should expand or shrink features.
- MultiScaleMorph Module: To handle both fine-grained local primitives (edges, corners) and long-range optical interactions, this module splits feature channels into groups and applies morphological operations using varying structuring element (SE) sizes (e.g., $3\times3$ and $5\times5$ kernels) within a hierarchical residual framework.
- Encoder-Decoder Architecture: These modules are integrated into a U-Net-style architecture. The encoder extracts and geometrically refines features, while the decoder progressively upsamples and fuses them using skip connections to preserve high-resolution spatial details.
- Two-Stage Training: The model is first pre-trained to minimize mean squared error (MSE) against ground-truth masks, followed by a GAN-style fine-tuning stage that utilizes a forward lithography simulator to directly minimize the actual printing error on the wafer.
Datasets
The model was evaluated on three standard industry benchmarks:
- GAN-OPC: 4,875 synthetic metal-layer tiles designed under 32 nm M1 layout rules.
- LithoBench: Consisting of MetalSet (16,472 tiles for metal-layer ILT at 32 nm) and ViaSet (116,415 clips for via-layer ILT at 45 nm).
- MaskOpt: A large-scale dataset containing 104,714 metal-layer tiles for both edge-based OPC (EBOPC) and ILT, derived from real IC designs fabricated at the 45 nm node.
- ICCAD 2013 Contest Benchmark: Used strictly as an unseen test set to evaluate cross-design generalization.
Key Insights and Ideas
- Geometry over Statistics: By embedding max-plus and min-plus operations, the network is structurally incentivized to learn edge displacements (inward contraction vs. outward expansion). This directly mirrors the physical geometric reasoning used by human mask designers.
- Layer-by-Layer Sophistication: Visualizations reveal that earlier morphological layers capture broad, global edge adjustments, while deeper layers naturally learn to apply intricate, fine-grained local transformations like sharp corner serifs.
- The Hybrid Advantage: Beyond standalone use, ML-generated masks from MorphOPC serve as excellent initializations for traditional optimization-based ILT algorithms, accelerating their convergence and leading to superior final printing quality with minimal optimization slowdown.
Key Results
- Superior Printing Fidelity: MorphOPC consistently achieved the lowest $\ell_2$ printing error across nearly all datasets. On the MaskOpt (EBOPC) benchmark, it reduced $\ell_2$ error by 18.7% compared to GAN-OPC and 8.5% compared to DAMO, while slashing Edge Placement Error (EPE) by roughly 70%.
- Enhanced Manufacturability: For pixel-based ILT on MaskOpt, MorphOPC reduced the final mask "shot count" (the number of rectangular shapes needed to manufacture the mask) by over 40% compared to GAN-OPC and DAMO, significantly lowering manufacturing complexity.
- High Efficiency: MorphOPC maintains an inference runtime on the order of $10^{-2}$ to $10^{-1}$ seconds per tile. This matches the speed of standard ML baselines while performing orders of magnitude faster than traditional optimization methods (which take upwards of 4 seconds per tile).
- Robust Generalization: When trained on LithoBench and directly evaluated on the unseen ICCAD 2013 benchmark, MorphOPC outperformed all baselines in printing fidelity and process variation robustness.
Limitations and Future Work
Although MorphOPC drastically reduces mask complexity (shot counts) compared to standard generative models, specialized physics-driven ILT models (like Neural-ILT or CFNO) can still achieve lower shot counts in specific pixel-based ILT scenarios—albeit at the expense of significantly worse printing fidelity ($\ell_2$ errors over $4\times$ higher than MorphOPC).
The authors suggest that future work could explore tighter integration of lithography physics directly into the learnable structuring elements to further optimize the trade-off between mask manufacturability and wafer printing accuracy.
Summary updated