The Random Forest classification model under consideration is trained on machine signal data collected from wire bond machines. The model performans binary classification at the device level and supports quality detection and automatic root cause analysis in semiconductor wire bonding.
A Random Forest classification model is a machine-learning method that makes predictions by combining the output of many decision trees. Each trained on a random subset of the data and features. By aggregating their results (through majority voting for classification), it produces accurate and robust predictions while reducing overfitting.
For this task, a random forest classification model was trained using real-world machine signal data collected from wire bonding experiments under both standard and changed settings. Each experiment comprised 44 integrated circuits (ICs), resulting in a total dataset of 176 ICs. The dataset was constructed to reflect scenarios associated with forming gas-related issues and corresponding reference conditions, enabling supervised learning of mimicked process deviations. To ensure robust and reliable performance, the data were repeatedly split into training and test sets using different random seeds, and the model was trained using default hyperparameter settings.
Key traits of Random Forest Classification:
Ensemble learning approach: Combines predictions from multiple decision trees trained on randomly sampled subsets of the data to improve robustness and generalization.
Inherent feature importance estimation: Provides global feature importance measures.
Strong performance and scalability: Particularly well-suited for high-dimensional, structured datasets (e.g. machine signals), while effectively handling large and complex feature sets without significant loss in accuracy or computational efficiency.
Limited preprocessing requirements: Requires minimal data scaling or normalization, is relatively robust to missing values, and efectively capturews nonlinear relationships and feature interactions.
Machine signal data colleced during the wire bonding process are aggregated at the device level and used as input features.
A Random Forest classification model is trained to distinguish between devices produced under standard process conditions and those associated with changed conditions linked to potential quality issues.
For the implementation of the model the scikit-learn Python module version 1.4.0 with the corresponding default hyperparametersetting was used.
Each experiment is compared against the reference experiment, which is used to define the labels.
The proposed Random Forest classification modeling approach is intended for:
Device-level quality detection: Identifying devices that are likely to have quality issues leveraging machine signal data collected during the wire bonding process.
Support for automatic root cause analysis: Linking machine-signal-related features to potential process distrubances (i.e. forming gas issues), thereby accelerating and improving root cause analysis.
Decision support in manufacturing: Assisting engineers and operators by providing data-driven insights that complement current quality control practices.
Saving data storage: Support smarter cloud-storing decisions by focusing only on relevant machine signals.
Limitations:
Dependence on labeled data: The proposed model requires labeled datasets corresponding to known process conditions or root cause, which may be costly or time-consuming to obtain in practice.
Limited scope: The model is trained on data from speciic experiments, bond recipes, machines. The performance may be degrade when broaden the scope and applying the model to different machines, recipes or extended time periods without retraining.
Sensitivity to data drift: Changes in process behavior, machine settings, or sensor characteristics over time may affect model performance.
@INPROCEEDINGS{Braakman2024,
author={Braakman, Kenneth J. and Knotter, D. Martin and Akcay, Alp and Adan, Ivo},
booktitle={2024 Winter Simulation Conference (WSC)},
title={Leveraging Machine Signals for Device-Level Quality Detection and Automatic Root Cause Analysis in Semiconductor Wire Bonding},
year={2024},
pages={1886-1897},
doi={10.1109/WSC63780.2024.10838790}}