Wafer Process Control Using Reinforcement Learning Across Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semiconductor manufacturing processes lack through-stack yield optimization and fail to consider the impact of earlier processing steps on subsequent layers, leading to inefficiencies and increased costs due to dense and time-consuming measurements, as well as inadequate adjustments based on historical data from multiple layers.
Innovation Solution
The implementation of a model-free reinforcement learning (MFRL) framework that determines a sequence of states and process metrics for a semiconductor wafer, allowing for adjustments to processing parameters and operations across multiple layers to optimize yield and reduce measurement density, using algorithms like asynchronous advantage actor-critic and Q-learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If dense measurements are performed on each layer to ensure processing quality, then manufacturing precision is improved, but measurement time and costs increase
Solution Approach 1:
The system performs measurements selectively rather than densely on every layer. By using reinforcement learning to identify which layers require measurements and which can be skipped, the system achieves adequate processing quality with reduced measurement time and costs.
Solution Approach 2:
The system uses feedback from previous layer measurements and process metrics to determine measurement strategies for subsequent layers. This feedback mechanism allows the system to adjust measurement density dynamically, performing measurements only when necessary to maintain manufacturing precision.
2Ease of operation
If corrections are made based only on immediately previous layer information, then ease of operation is improved, but yield optimization deteriorates
Solution Approach 1:
The system transitions from two-dimensional layer-by-layer correction to a through-stack multi-dimensional optimization approach. By considering the entire stack of layers simultaneously and using reinforcement learning to optimize corrections across all layers, the system achieves both operational feasibility and yield optimization.
Solution Approach 2:
The system performs preliminary analysis of the entire processing stack before making corrections. By using reinforcement learning to evaluate multiple correction scenarios and their impact on final yield, the system can make informed decisions that optimize overall yield rather than just individual layer quality.
3Reliability
If through-stack yield optimization is implemented considering all previous layers, then reliability is improved, but device complexity increases
Solution Approach 1:
The system uses self-service through reinforcement learning algorithms that automatically analyze processing information from multiple layers and determine optimal corrections without requiring complex manual intervention. The reinforcement learning agent learns from historical data and autonomously optimizes yield across the entire stack.
Solution Approach 2:
The system changes the parameter of correction strategy from simple layer-by-layer adjustments to complex through-stack optimization using reinforcement learning. By implementing this parameter change in the control approach, the system achieves superior yield optimization while managing complexity through automated algorithmic decision-making.
4Measurement precision
If measurements are performed on every layer to track processing state, then measurement precision is improved, but productivity deteriorates
Solution Approach 1:
The system performs measurements partially rather than on every layer. By using reinforcement learning to identify the minimum necessary measurements required to maintain adequate processing state tracking, the system achieves sufficient measurement precision while improving productivity through reduced measurement burden.
Data Source
AI summary
A method including: determining a sequence of states of an object, the states determined based on processing information associated with the object, wherein the sequence of states includes one or more future states of the object; determining, based on at least one of the states within the sequence of states and the one or more future states, a process metric associated with the object, the process metric including an indication of whether processing requirements for the object are satisfied for individual states in the sequence of states; and initiating an adjustment to processing based on (1) at least one of the states and the one or more future states and (2) the process metric, the adjustment configured to enhance the process metric for the individual states in the sequence of states such that final processing requirements for the object are satisfied.