Diffusion Model Parameter Inference for Wafer Metrology

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current metrology data processing techniques are computationally intensive, prone to getting stuck in local minima, and fail to accurately model patterned substrates due to insufficient parameterization, especially when extra layers like oxidation are present.

Innovation Solution

Employ a generative neural network, specifically a diffusion model, to iteratively refine substrate model data, making it more consistent with metrology data, using sensor configuration data and class information to enhance robustness and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If iterative optimization algorithms are used to infer parameters from metrology data, then parameter estimation can be obtained, but the computational burden is intensive and the process is time-consuming

Engineering Contradiction:
Improveparameter estimation accuracyVSAvoidcomputational time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent pre-trains a diffusion model on a large dataset of substrate structures and their corresponding metrology data before actual parameter inference. This preliminary training enables the model to learn complex relationships offline, so that during actual use, parameter inference can be performed rapidly without iterative optimization, thus reducing computational time while maintaining accuracy

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses a diffusion model that generates parameter estimates by learning from trained data distributions rather than performing iterative optimization from scratch. The model creates a probabilistic copy of the parameter inference process based on patterns learned during training, enabling fast approximation without repeated forward model evaluations

Inventive Principle:
Principle #26Copying

2Manufacturing precision

If sophisticated fine-tuning steps and resolution enhancement techniques are applied to achieve low k1 lithography, then pattern reproduction accuracy is improved, but the process complexity increases

Engineering Contradiction:
Improvepattern reproduction accuracyVSAvoidprocess complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent infers multiple substrate parameters simultaneously (including thickness, refractive index, absorption coefficient, and roughness) using a unified diffusion model. By changing from traditional single-parameter optimization to multi-parameter joint inference, the system achieves accurate pattern reproduction while simplifying the overall process through integrated parameter estimation

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If traditional parameter inference methods are used, then parameter estimates can be obtained, but the methods fail to accurately model substrates with extra layers such as oxidation

Engineering Contradiction:
Improvesubstrate model accuracyVSAvoidmodel adaptability to varying substrate structures
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent employs a diffusion model that is trained on diverse substrate structures including various layer configurations, oxidation layers, and different material compositions. This universal training enables the single model to accurately infer parameters for substrates with extra layers or unexpected structures without requiring model reconfiguration, thus achieving both high accuracy and adaptability

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4700481A1Diffusion models for parameter inference with application to wafer metrology
Publication Date: 2026.02.25 ASML NETHERLANDS BV
  • EP4700481A1 patent drawingFigure 1
  • EP4700481A1 patent drawingFigure 2
  • EP4700481A1 patent drawingFigure 3~4

AI summary

A generative network, conditioned on metrology data obtained from at least one patterned substrate, is used to process an input, to generate an output. The input is substrate model data defining a substrate model of the at least one substrate, and the output of the generative network is used to produce updated substrate model data defining an updated substrate model. In one or more further iterations, the generative network processes an input based on the updated substrate model generated in the preceding operation. The output of the generative network in the last iteration is used to generate substrate description data describing the at least one substrate.