Iterative Error‐Driven Ensemble Labeling (IEDEL) Algorithm for Enhanced Data Quality Control for the Atmospheric Radiation Measurement (ARM) Program User Facility

For over three decades, the Atmospheric Radiation Measurement (ARM) Program user facility has provided researchers with invaluable benchmark atmospheric data. Ensuring the accuracy and integrity of ARM data is vital, and to achieve this, the ARM Data Quality Office (DQO) has implemented customized quality control tests tailored to each variable, with guidance from instrument mentors. These tests are designed to pinpoint common issues, such as data exceeding valid ranges or persisting with little change over extended periods, and ARM offers tools for users to review and exclude contaminated data efficiently. However, certain quality issues, such as spikes in time series or data drift over time, sometimes evade detection by existing tests and require manual identification by data analysts and instrument mentors through visualization tools. To tackle these challenges more efficiently, the DQO has developed and implemented the Iterative Error‐Driven Ensemble Labeling (IEDEL) algorithm with unanimous voting and transfer learning techniques to efficiently generate labeled data at scale. This initiative has empowered the creation of high‐performing machine learning models, enabling real‐time monitoring of data quality issues within the ARM data and thereby enhancing data integrity and accessibility. Plain Language Summary For more than 30 years, the Atmospheric Radiation Measurement (ARM) Program user facility has been providing scientists with important atmospheric data. Ensuring these data are accurate and trustworthy is crucial. To achieve this, the ARM Data Quality Office (DQO) establishes tailored quality control (QC) checks for each data variable, based on thresholds designed by the ARM instrument mentors, who are experts in meteorology. These checks help identify common data issues, such as data falling outside the normal range or not changing as expected over time. However, some problems, like sporadic data spikes or shifts in the average of data over time, might not be detected by these QC checks. These issues require visual identification by data analysts and ARM instrument mentors using ARM's visualization tools. To become more efficient at detecting these problems, the DQO has developed a new method called the Iterative Error‐Driven Ensemble Labeling algorithm to label data issues and used a machine learning algorithm to categorize them. This innovative approach enables the DQO to build intelligent applications that monitor data in real time, around t