Neural Network Feedforward Compensation for Compliant Motion Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing motion control systems in semiconductor manufacturing face challenges in accurately controlling the motion of apparatus components due to dynamic forces and time variance, leading to inaccuracies in feature reproduction on substrates.

Innovation Solution

The implementation of a motion control system that utilizes neural network feedforward control with projection-based regularization, combining a physics-guided neural network with a nominal feedforward control structure to achieve linear parameter-varying snap feedforward control. This system uses meaningful physical input signals and applies tailored wafer meandering profiles to generate training data, avoiding high frequency resonance behavior.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional motion control systems are used to control apparatus components, then the system structure is relatively simple, but motion control accuracy deteriorates due to dynamic forces and time variance

Engineering Contradiction:
Improvemotion control accuracyVSAvoidcontrol system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces traditional mechanical control systems with an artificial intelligence-based control system. Specifically, a neural network model is trained to predict and compensate for dynamic forces and time variance in motion control, substituting complex mechanical control mechanisms with an intelligent software-based solution that learns from training data to achieve higher precision without increasing physical system complexity

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the control parameters by using AI-optimized control signals instead of traditional fixed parameters. The neural network dynamically adjusts control parameters based on learned patterns from training data, enabling adaptive compensation for dynamic forces and time variance, thereby improving motion control accuracy without requiring complex hardware modifications

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If sophisticated fine-tuning steps and resolution enhancement techniques are applied to improve pattern reproduction, then manufacturing precision improves, but device complexity and process time increase

Engineering Contradiction:
Improvepattern reproduction accuracyVSAvoidcontrol system complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent substitutes complex resolution enhancement techniques and fine-tuning procedures with an AI-based motion control system. The neural network learns optimal control strategies from training data to directly achieve high-precision pattern reproduction, eliminating the need for multiple sophisticated correction steps and reducing overall system complexity

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent applies preliminary action by training the neural network model in advance with extensive training data that captures dynamic forces and time variance characteristics. This pre-trained model then automatically applies learned compensation during actual operation, achieving high manufacturing precision without requiring complex real-time adjustments or multiple fine-tuning steps

Inventive Principle:
Principle #10Preliminary action

3Manufacturing precision

If tight control loops are used to control stability of lithographic apparatus, then manufacturing precision improves, but response time and adaptability to dynamic forces deteriorate

Engineering Contradiction:
Improvestability control accuracyVSAvoidresponse speed
Core Design Contradiction:
Manufacturing precisionVSSpeed

Solution Approach 1:

The patent replaces tight mechanical control loops with an AI-based predictive control system. The neural network predicts future states and compensates for dynamic forces proactively, enabling fast response to changing conditions without the lag inherent in traditional feedback control loops, thus maintaining both precision and speed

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent applies preliminary action by using the trained neural network to predict and compensate for dynamic forces before they significantly impact system stability. The model uses learned patterns from training data to anticipate and correct deviations, achieving both tight stability control and rapid response to dynamic changes

Inventive Principle:
Principle #10Preliminary action

4Measurement precision

If conventional feedforward control is used, then the control structure is simple, but compensation accuracy for dynamic forces deteriorates

Engineering Contradiction:
Improvecompensation accuracyVSAvoidcontrol algorithm complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces conventional feedforward control algorithms with an artificial intelligence-based neural network model. The neural network learns complex compensation patterns from training data, achieving superior accuracy in compensating for dynamic forces and time variance without requiring complex mathematical models or multiple control algorithms

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the control approach by using a data-driven neural network instead of model-based feedforward control. The neural network automatically learns optimal compensation parameters from training data, adapting to specific system characteristics and achieving high compensation accuracy without manual parameter tuning or complex control logic

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4550063A1Neural network feedforward framework for compliance compensation
Publication Date: 2025.05.07 ASML NETHERLANDS BV
  • EP4550063A1 patent drawingFigure 1
  • EP4550063A1 patent drawingFigure 2~3
  • EP4550063A1 patent drawingFigure 4~5

AI summary

Neural network feedforward control with projection-based regularization is described. This enables linear parameter-varying snap feedforward control. A physics guided neural network is used in combination with a nominal feedforward control structure. The neural network has meaningful physical input signals and is small to overcome over-parameterization. The parameters of the neural network render machine specific settings. Projection-based regularization is applied such that feedforward contributions explained by a nominal model end up in this model, which are simultaneously optimized, and not in the neural network. A meandering profile is used that renders data that reflects relevant properties of a setpoint, and relevant data needed for identifying the compliant behavior that avoids exiting high frequency resonance behavior. A compliance function is obtained from the neural network rather than a table of compliance values for specific positions, which does not require interpolation or storage of tables.