Reservoir Computing for Substrate Process State Determination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In substrate processing apparatuses, machine learning models require a high workload for both initial setup and maintenance due to the need for extensive training data, especially when determination accuracy is affected by time-dependent changes.
Innovation Solution
A substrate processing apparatus employing a reservoir computing unit that learns weight parameters using time-series sensor data, allowing for efficient learning and prediction, thereby reducing the workload for initial setup and maintenance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a general machine learning model is used to improve determination accuracy, then determination accuracy is improved, but the workload required for learning work at start-up and maintenance increases
Solution Approach 1:
The machine learning model is segmented into two parts: a reservoir computing unit that performs complex time-series processing and a simple readout layer that requires minimal learning. This segmentation allows the system to maintain high determination accuracy while significantly reducing the learning workload, as only the readout layer parameters need to be learned rather than the entire model
Solution Approach 2:
The reservoir computing unit acts as an intermediary that transforms raw time-series sensor data into a fixed-dimensional feature space. This intermediary structure handles the complex temporal dependencies and nonlinear transformations, allowing the final readout layer to perform simple linear classification with minimal learning requirements
Data Source
AI summary
A workload required for learning work is reduced. A substrate processing apparatus includes a reservoir feature generating unit configured to receive acquired first time-series sensor data and output a reservoir feature; a learning unit configured to learn, in a learning period, a weight parameter so that prediction result data obtained by performing calculations on the reservoir feature under the weight parameter correlates with acquired second time-series sensor data; a predicting unit configured to perform calculations, in a prediction period, on the reservoir feature output from the reservoir feature generating unit in response to the acquired first time-series sensor data being input, under the learned weight parameter, to output prediction result data; and a determining unit configured to determine, in the prediction period, a state of the substrate manufacturing process by comparing the prediction result data with acquired second time-series sensor data.