Deep Learning X-Ray Scatterometry for Noisy HAR Measurements

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing X-ray analysis methods for measuring critical dimensions of semiconductor structures require expert intervention, are resource-intensive, and suffer from noise-related distortions and process variations, leading to suboptimal measurement quality and efficiency.

Innovation Solution

Employing deep learning techniques, specifically convolutional neural networks (CNNs), to analyze X-ray scatterometry data, enabling automated alignment and measurement of high aspect ratio (HAR) structures by training neural networks with synthetic and real data to improve measurement accuracy and reduce resource requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If traditional X-ray analysis methods are used, then measurement capability is provided, but expert intervention and operational resources are required, reducing automation

Engineering Contradiction:

Improveautomation of X-ray analysisVSAvoidneed for expert intervention

Core Design Contradiction:

Extent of automationVSDevice complexity

Solution Approach 1:

The system employs deep learning models that automatically perform alignment and measurement tasks without requiring expert intervention. The neural networks self-adjust parameters and interpret diffraction patterns autonomously, transforming a process that previously required skilled operators into an automated self-service system.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent replaces manual expert analysis with computational algorithms. Deep learning models substitute the mechanical process of expert visual inspection and manual measurement with automated image processing and pattern recognition systems that operate without human intervention.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Productivity

If traditional X-ray analysis methods are used, then measurements can be performed, but resource consumption is high, reducing productivity

Engineering Contradiction:

Improvethroughput of X-ray analysisVSAvoidoperational resources

Core Design Contradiction:

ProductivityVSLoss of energy

Solution Approach 1:

The system performs preliminary alignment and parameter optimization using deep learning models before actual measurement. The neural networks pre-process diffraction patterns and predict optimal measurement parameters, reducing the need for repeated measurements and minimizing operational resource consumption during production.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent utilizes deep learning models that dynamically adjust measurement parameters based on learned patterns from training data. This adaptive parameter optimization reduces measurement time and resource consumption by selecting optimal measurement conditions without requiring extensive trial-and-error experimentation.

Inventive Principle:

Principle #35Parameter changes

3Measurement precision

If traditional X-ray analysis methods are used, then measurements are performed, but noise and process variations cause suboptimal measurement quality

Engineering Contradiction:

Improvemeasurement qualityVSAvoidnoise and process variations

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The system employs deep learning models trained on extensive datasets that include various noise conditions and process variations. The models learn to distinguish signal from noise through feedback during training, enabling them to maintain high measurement precision even when subjected to real-world disturbances and variations.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent pre-trains deep learning models with diverse synthetic and experimental data that encompasses expected noise patterns and process variations. This beforehand cushioning prepares the models to robustly handle real measurement conditions, reducing the impact of noise and variations before actual measurements are performed.

Inventive Principle:

Principle #11Beforehand cushioning (Prior cushioning)

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

Enhances the automation, throughput, and quality of X-ray analysis by reducing the need for expert engineers and operational resources, while improving measurement repeatability, reproducibility, and sensitivity.

Implementation Method 1

a diffraction image indicative of X-ray photons diffracted from one or more structures formed in a sample responsively to directing an incident X-ray beam at an angle relative to the sample

Methodology Applied

Scientific EffectDiffraction: Diffraction

Data Source

PatentUS12535438B2Analysis of x-ray scatterometry data using deep learning

Publication Date: 2026.01.27 BRUKER TECH LTD

AI summary

A method for training a neural network (NN), the method includes: receiving a training dataset including: (a) multiple pairs of: (i) a diffraction image indicative of X-ray photons diffracted from structures formed in a sample responsively to directing an incident X-ray beam at an angle relative to the sample, and (ii) a label, including: a first parameter indicative of at least a first property of the structures, and a second parameter indicative of at least a second property of the incident X-ray beam, and (b) multiple predefined outputs for the multiple pairs, respectively. The NN is trained to obtain the predefined outputs by: (i) applying the NN to at least a given pair of the pairs, and (ii) responsively to receiving from the NN an estimated output of the given pair, providing the NN with a given predefined output of the predefined outputs that corresponds to the given pair.