Federated Learning for Semiconductor Process Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Machine learning models used for predicting maintenance and calibration needs in semiconductor manufacturing processes face challenges due to the confidential nature of data from suppliers and customers, leading to reluctance in data sharing and inaccurate predictions when using non-confidential data.

Innovation Solution

The implementation of vertically federated learning allows for the training of machine learning models using aligned time series data sets from different participants while preserving privacy, by receiving and combining model parameters from multiple participants without exposing their confidential data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning models are trained using time series data from multiple participants, then prediction accuracy for maintenance and calibration needs is improved, but data confidentiality and privacy are compromised

Engineering Contradiction:
Improveprediction accuracyVSAvoiddata confidentiality
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent segments the training data horizontally across multiple participants, where each participant contributes time series data from their own equipment. The federated learning system trains a global model by aggregating gradients or updates from each participant's local model, ensuring that raw confidential data remains distributed and never centralized, thus maintaining data confidentiality while achieving improved prediction accuracy through multi-participant data diversity

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a federated learning server as an intermediary that coordinates the training process without accessing any participant's raw data. This intermediary facilitates the exchange of model updates and gradients between participants, enabling collaborative model training while preserving the confidentiality of each participant's proprietary equipment data through encrypted communication and centralized secure aggregation

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If confidential data is shared between suppliers and customers for model training, then model performance is improved, but data security risks increase

Engineering Contradiction:
Improvemodel performanceVSAvoiddata security risks
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent implements preliminary actions by having each participant train their own local model on their confidential data before participating in federated learning. This preliminary local training allows the system to leverage the performance benefits of participant-specific data without requiring actual data sharing, as only model parameters and gradients (not raw data) are exchanged with the federated server, thus improving model performance while eliminating data security risks

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by creating and exchanging copies of model parameters, gradients, and updates instead of sharing the actual confidential raw data. Each participant's equipment data remains locally stored and never leaves their premises, while the federated learning system aggregates copies of model information to improve the global model's performance, achieving reliable predictions without exposing sensitive information or creating data security vulnerabilities

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4538796A1Vertically federated training of a machine learning model used by different participants for configuring a semiconductor manufacturing process
Publication Date: 2025.04.16 ASML NETHERLANDS BV
  • EP4538796A1 patent drawingFigure 1
  • EP4538796A1 patent drawingFigure 2
  • EP4538796A1 patent drawingFigure 3

AI summary

Training a machine learning model used by different participants is described. Vertically federated learning is used to train the model with time series data sets. Time series data sets received from different participants are aligned, with each time series data set comprising different features corresponding to one or more samples common to each participant, but preserving the privacy of each time series data set for participants. First model parameters are received from a first participant. These are determined based on the aligning and first features provided by the first participant in a first time series data set. Second model parameters are received from a second participant. These are determined based on the aligning, second features provided by the second participant in a second time series data set, and semiconductor manufacturing process outputs associated with the second features. The model is trained based on the first and second model parameters.