Prevention of Leakage in Machine Learning Prediction for Polymer Composite Properties.

Abstract

Machine learning (ML) has facilitated property prediction for intricate materials by integrating materials and experimental features such as processing and measurement conditions. However, ML models designed for material properties have often disregarded a common issue of "leakage," resulting in an overestimation of model performance and a decrease in model transferability. This issue can arise from biases inherent in multiple data points obtained from the same experimental group. We provide a critical examination and prevention method of leakage in property prediction for polymer composites. Our proposed method utilizes data partitioning based on the experimental group to ensure that data from the same group are not mixed in both the training and test sets. Evaluation results highlight that the conventional random partitioning unintentionally inflates ML performance through the misuse of experimental features for leaking data bias within the same experimental group rather than explaining the physical causality. In contrast, the proposed method enables the leakage-free utilization of experimental features to improve prediction accuracy while ensuring model transferability. Specifically, when integrating experimental features with polymer and filler features, the conventional method overestimates the prediction performance of electrical conductivity in reducing RMSE by 26% depending on leakage, whereas the proposed method achieves a reduction in RMSE by 5% without leakage. These findings offer valuable guidance for the effective utilization of experimental features in data-driven materials science.