Transfer learning-based modeling of synthesis-particle size relationships in metal-organic frameworks

Abstract

Metal–organic frameworks (MOFs) are promising materials for applications such as gas storage and drug delivery, where their performance is strongly affected by morphology such as particle size. However, the relationships between synthesis conditions and the resulting particle size are complex and remain poorly understood. This work develops a transfer learning-based modeling framework to quantitatively predict how synthesis parameters affect the particle size of zeolitic imidazolate framework-8 (ZIF-8). Using an extreme gradient boosting algorithm (XGBoost), we first build a baseline predictive model trained on a combined dataset of literature and in-house experimental data. To improve performance, we explore transfer learning that pretrains a model on literature data to fix structural hyperparameters such as the number of trees and then fine-tune decision splits with a weighting scheme to emphasize in-house data. To further address data scarcity, we adopt synthetic data augmentation through local interpolation in synthesis parameter space. New data are generated by interpolating between nearby in-house synthesis conditions, preserving the physical meaning of augmented samples. Our results show that transfer learning and data augmentation significantly improve model accuracy and interpretability. This work demonstrates the potential of transfer learning to bridge heterogeneous data sources and accelerate data-driven materials synthesis optimization.