Masked Autoencoder for Data Recovery in Polymer Research: Mitigating Data Integrity Threats

Polymer research critically depends on high-quality datasets to understand material properties, develop new materials, and optimize existing ones. However, these datasets are increasingly vulnerable to data integrity threats such as accidental deletions, system failures, and cyber attacks, leading to significant data loss and corruption. Traditional data recovery methods, including mean and median imputation, often fail to accurately restore missing data due to their inability to capture complex data relationships. While machine learning techniques like autoencoders offer improved recovery by learning intricate data patterns, their effectiveness in polymer research data remains largely unvalidated. In this paper, we propose an autoencoder-based data recovery approach that utilizes random masks to guide the recovery process. This method is specifically tailored to the complexities of polymer datasets, aiming to achieve higher accuracy and robustness in data recovery. We evaluated the proposed approach using cloud point temperature datasets for binary polymer solutions. The evaluation results demonstrate it outperforms conventional imputation methods, effectively mitigating the consequence of data loss and compromised data integrity.