SLGAN Learning Rate Control for Semiconductor Defect Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor inspection methods struggle to accurately detect and characterize micron-scale to sub-micron-scale features on semiconductor workpieces, such as threading edge dislocations and stacking faults, due to the complexity and variability of these features, which are not adequately addressed by conventional metrology solutions.
Innovation Solution
A stabilized learning generative adversarial network (SLGAN) is employed, with regulated learning rates for the discriminator and generator networks, to enhance the detection and characterization of semiconductor workpiece features by stabilizing the learning process and improving the accuracy of feature detection and classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional metrology solutions are used for inspection, then the inspection process is simple, but the detection precision of micron-scale to sub-micron-scale features is insufficient
Solution Approach 1:
The patent introduces an intermediary stabilization mechanism that mediates between the generator and discriminator networks. This stabilization mechanism regulates the learning rates of both networks to maintain a balanced adversarial relationship, enabling the complex SLGAN model to achieve superior detection precision for micron-scale to sub-micron-scale features while managing the complexity through controlled interaction dynamics
2Productivity
If the learning rate is increased to speed up training, then the training speed improves, but the training stability deteriorates
Solution Approach 1:
The patent implements dynamic learning rate regulation where the learning rates of the generator and discriminator networks are adjusted dynamically during training based on their respective losses. This dynamic adjustment mechanism allows the system to maintain high training speed while ensuring stability by adapting the learning rates to the current training state, preventing the discriminator from overwhelming the generator
Solution Approach 2:
The patent employs feedback mechanisms where the losses of both the generator and discriminator networks are continuously monitored and fed back to regulate their learning rates. This feedback loop ensures that when the discriminator becomes too powerful, the generator's learning rate is increased relative to it, maintaining training stability while progressing efficiently
3Measurement precision
If the discriminator network is made more powerful to improve feature detection, then the detection accuracy improves, but the generator network cannot keep up, causing training instability
Solution Approach 1:
The patent changes the learning rate parameters of the generator and discriminator networks dynamically during training. By regulating these parameters based on the relative performance and losses of the two networks, the system maintains an balanced adversarial relationship where the generator can keep up with the powerful discriminator, ensuring both high feature detection accuracy and training stability
Data Source
AI summary
Systems and methods for inspecting semiconductor workpieces are provided. In one example, a method includes obtaining workpiece data for a semiconductor workpiece. The method includes providing the workpiece data as input to an inspection model, the inspection model being a stabilized learning generative adversarial network (SLGAN) trained model, wherein the SLGAN trained model is associated with a regulated learning rate for one or more of a discriminator network or a generator network. The method also includes obtaining an output from the inspection model, the output associated with one or more characteristics of the semiconductor workpiece.