Polishing Endpoint Detection Using ML Waveform Recipes

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing polishing methods struggle to accurately determine the polishing end point of substrates due to noise components in the output waveform of the driving device, requiring time-consuming manual adjustment of waveform processing parameters to create an optimal recipe.

Innovation Solution

A method utilizing a machine learning algorithm to construct a model that processes output waveforms of the driving device, calculating a simulated detection time to minimize differences from an estimated detection time, and adjusting polishing end point times based on polishing information data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If waveform processing parameters are manually adjusted to remove noise components, then measurement precision of polishing end point is improved, but loss of time and operational effort increase

Engineering Contradiction:

Improvepolishing end point detection accuracyVSAvoidtime and effort to create waveform processing recipe

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system automatically determines optimal waveform processing parameters by self-learning from measurement data, eliminating the need for manual operator adjustment. The polishing apparatus autonomously creates the waveform processing recipe by processing output waveforms and identifying parameters that minimize detection time differences, thereby resolving the contradiction between high measurement precision and time consumption.

Inventive Principle:

Principle #25Self-service

2Reliability

If waveform processing parameters are manually adjusted based on operator experience, then reliability of polishing end point determination is improved, but device complexity and operational difficulty increase

Engineering Contradiction:

Improvepolishing end point determination accuracyVSAvoidcomplexity of creating waveform processing recipe

Core Design Contradiction:

ReliabilityVSEase of operation

Solution Approach 1:

The patent replaces the manual operator-based system with an automated computational system. Instead of relying on operator experience to adjust waveform processing parameters, the system uses automatic processing of output waveforms from the driving device to determine optimal parameters. This substitution eliminates the need for operator expertise while maintaining or improving reliability.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Measurement precision

If noise components are removed from output waveform, then measurement precision is improved, but loss of information and processing complexity increase

Engineering Contradiction:

Improveoutput waveform analysis accuracyVSAvoidwaveform processing complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The system dynamically adjusts waveform processing parameters based on the specific characteristics of each output waveform. By changing parameters such as filtering levels, averaging periods, and detection thresholds according to the actual measurement conditions, the system achieves effective noise removal while adapting to different polishing scenarios, thereby managing processing complexity.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentEP4644044A1Polishing method, polishing device and recording medium

Publication Date: 2025.11.05 EBARA CORP

Figure 1
Figure 2
Figure 3~4

AI summary

The present invention relates to a polishing method, a polishing apparatus, and a recording medium. The polishing method includes inputting a measured output waveform and an estimated detection time into a model constructed by a machine learning algorithm, and outputting a waveform processing recipe, having waveform processing parameters for determining a simulated detection time that minimizes a difference from the estimated detection time, from the model.