Ion Implantation Sensor and Process Target Data for Predicting Ion Beam Tuning in Semiconductor Manufacturing

Dataset Description:

This dataset is designed to predict ion beam tuning setup processes in semiconductor manufacturing, in terms of tuning success or failure, and tuning duration. It is split into X and y to allow for supervised learning approaches.

X represents the current equipment condition and the process targets of the currently processed and the upcoming lot, as defined within recipes.

y represents the ion beam tuning setup report, which informs about the tuning success ratio and tuning duration. These setups are necessary, when switching between recipes to prepare the equipment for processing the next lot. y contains three labels, enabling classification of (1) tuning success or fail, and (2) prolonged tuning, as well as (3) estimation of tuning duration as a regression task.

About X:

Each lot is processed with a specific recipe to achieve the process target. The tuning takes place before the first wafer of the to-be-tuned recipe is processed. Each row in X includes logistical information such as the equipment used for processing and parsed recipe / process target information for the current and upcoming lot. The majority of data consists out of aggregated metrics of equipment-internally tracked sensor traces, recording physical parameters such as gas flows, temperatures, voltages and currents. When analyzed in conjunction with the processed recipe, these sensors provide insights into the current equipment condition. 

About y:

The setup_result column indicates the success or failure of tuning - with setup_result=0 indicating tuning success, while setup_result=1 signals tuning failure. If the first tuning attempt fails, there may be follow-up attempts, but these are not included in this dataset. The duration column represents the tuning duration in seconds, as used for regression analysis. The duration_interval column is a binary label for prolonged tunings, i.e. duration_interval=1 for instances, which take more than 6 minutes to tune.

For reproducibility of the corresponding paper's results:

The dataset contains the same carefully curated subset of features.

The train_test_split() has already been performed, thus we provide x_train and x_valid separately.

To reduce the effect of outliers in the data, the sensor data has already been scaled, as derived from x_train.

In summary, these datasets (X, y) provide comprehensive information for predicting ion beam tuning in semiconductor manufacturing, making it a valuable resource for researchers and practitioners in the field.

Python Code for Reproducibility:

Furthermore, we share a jupyter notebook ionbeamtuning.ipynb with Python code to train the best performing model on the provided data, as described in the paper. To execute the code, you may need to install any missing packages specified in the requirements.txt, as indicated within the notebook.