Lithography for AI Engineers¶

A bilingual cheat-sheet that maps computational-lithography terminology onto the AI / deep-learning vocabulary you already know. Each entry has the standard definition, the closest AI analogue, a canonical reference, and a one-line Chinese summary so you can use it in either community.

Lithography terms for AI engineers — each entry has an English definition, an AI analogy, and references. A consolidated Chinese–English glossary table sits at the end of this page.

Mask¶

Definition. The patterned reticle whose transmission profile is projected through the scanner onto the wafer. Digitally, a binary or grayscale 2-D image at sub-nanometer pixel pitch.

AI analogue. Input image / output tensor of an image-to-image network. A "predicted mask" is the network's output; a "target mask" is the design intent (label).

Reference. Wong, Resolution Enhancement Techniques in Optical Lithography, SPIE Press.¶

OPC (Optical Proximity Correction)¶

Definition. Pre-distorting the mask so that, after projection through a diffraction-limited lens and resist development, the printed wafer pattern matches the design. Industrially solved by rule-based or model-based iterative correction.

AI analogue. Image-to-image inverse generation. Given the desired output (design), produce the input (mask) such that a known forward operator maps mask → design. A discriminative U-Net trained on (design, mask) pairs is the simplest learned variant.

Reference. Yang et al., GAN-OPC, DAC 2018.¶

ILT (Inverse Lithography Technology)¶

Definition. Frame OPC as a continuous optimization: parametrize the mask, define a differentiable forward model (mask → aerial image → resist), and minimize a loss against the target design.

AI analogue. Differentiable optimization with a physics-based loss — the same shape as score-matching or PINNs. Curvilinear ILT closely mirrors learned image priors regularized by a physics simulator.

Reference. Pang et al., Inverse Lithography Technology Principles in Practice, JM3 2021.¶

SRAF (Sub-Resolution Assist Feature)¶

Definition. Small auxiliary mask features placed near main features to improve process window. They print sub-resolution (i.e., do not appear on the wafer) but bias the local diffraction pattern.

AI analogue. Auxiliary input/output channels — tokens that are part of the prediction but not part of the evaluation target. Conceptually similar to side outputs / deep supervision in segmentation networks.

Reference. Liebmann et al., SPIE Advanced Lithography, 2003.¶

EPE (Edge Placement Error)¶

Definition. Distance, in nanometers, between the predicted contour and the target contour at sampled edge points. The single most important per-feature accuracy metric.

AI analogue. Edge-aligned regression loss / pixel-wise distance metric. Closer to chamfer distance than to L2 — only edge pixels contribute, and the metric is asymmetric in some formulations.

Reference. Mack, Fundamental Principles of Optical Lithography, Wiley.¶

MRC / DRC (Mask / Design Rule Check)¶

Definition. Hard-fail compliance gating: minimum width, minimum spacing, minimum area, minimum curvature radius. A mask that fails MRC cannot be manufactured regardless of optical performance.

AI analogue. Constraint satisfaction layer — a post-hoc validity filter, similar to projection back onto a feasible set in constrained optimization.

Reference. SEMI P39, EasyMRC docs.¶

Resist Model¶

Definition. Model that maps the aerial image (light intensity at the wafer) to the developed resist contour. Industrial models are calibrated empirically; learned variants substitute a CNN.

AI analogue. Forward surrogate model — the SciML pattern of replacing a PDE solver with a neural network. The "true" forward model is computationally expensive; the surrogate is cheap and differentiable.

Reference. Mack, Fundamental Principles of Optical Lithography, Ch. 11.¶

PV-Band (Process Variation Band)¶

Definition. The envelope swept by resist contours over a dose/focus process window. Width of the band measures how stable the printed pattern is under realistic manufacturing variation.

AI analogue. Aleatoric uncertainty band — analogous to a confidence interval predicted by a probabilistic model under input perturbation.

Reference. Sturtevant et al., SPIE Advanced Lithography, 2010.¶

Aerial Image¶

Definition. Light intensity distribution at the wafer plane after the projection optics, before the resist responds. Computed via Hopkins / Abbe imaging from the mask plus illumination + lens kernel.

AI analogue. Intermediate latent representation — the activation in the middle of a forward pipeline. Many learned ILT models predict the aerial image as a byproduct.

Reference. Hopkins, Proc. R. Soc. A (1953).¶

Hotspot Detection¶

Definition. Identify layout patches that are likely to fail under nominal lithography conditions (bridge, break, necking). Trained on (layout patch, hotspot label) pairs from prior tape-outs.

AI analogue. Pixel-level or patch-level anomaly classification — same task structure as defect inspection, semantic segmentation of defects, or out-of-distribution detection.

Reference. Yang et al., DAC 2017 / Lin et al., ICCAD 2016.¶

Stochastic Failure (EUV-specific)¶

Definition. EUV photons arrive Poisson-distributed; resist response is also stochastic. Below a critical dose, low-photon variance produces random bridges, breaks, and missing contacts.

AI analogue. Monte Carlo sampling under input noise — similar to evaluating a model's robustness to per-pixel Gaussian perturbation, but the noise here is physically grounded photon shot noise.

Reference. De Bisschop, J. Micro/Nanolith. MEMS MOEMS, 2017.¶

Process Window¶

Definition. The set of (dose, focus) pairs over which all CD/EPE/MRC constraints are simultaneously satisfied. A wider window means the recipe is more robust to scanner drift.

AI analogue. Feasible region in the input-perturbation parameter space — analogous to the set of latent perturbations that preserve a classifier's decision.

Reference. Sturtevant et al., SPIE Advanced Lithography, 2010.

中文对照表 (Chinese–English Glossary)¶

术语	中文	一句话说明
Mask	掩膜	在 AI 视角下就是一张二维图像张量；模型预测它，设计层提供监督
OPC	光学邻近校正	类似图像逆问题：已知期望输出，反推输入
ILT	逆向光刻	把 OPC 写成可微优化问题，物理仿真器作为可微 loss
SRAF	辅助特征	不直接成像但影响主特征的“辅助通道”
EPE	边缘放置误差	可以把它当作 chamfer distance 的工业版
MRC / DRC	掩膜 / 设计规则检查	硬约束，类似可行域投影；不满足直接拒绝
Resist Model	光刻胶模型	物理仿真的“代理网络”，让你能微分通过它优化掩膜
PV-Band	工艺变动带	对应贝叶斯模型中的不确定性区间
Aerial Image	空间像	光在晶圆平面的强度分布，在管线里相当于“中间层激活”
Hotspot Detection	热点检测	标准的分块分类 / 语义分割问题，标签来自历史 tape-out 数据
Stochastic Failure	EUV 随机失效	光子打靶的统计性涨落，可用蒙特卡洛仿真评估
Process Window	工艺窗口	各类约束同时满足的 (dose, focus) 集合，类似分类器的稳健决策区域