Data-Quality-Navigated\nMachine Learning Strategy with\nChemical Intuition to Improve Generalization

Generalizing real-world data has been one of the most\ndifficult\nchallenges for application of machine learning (ML) in practice. Most\nML works focused on improvements in algorithms and feature representations.\nHowever, the data quality, as the foundation of ML, has been largely\noverlooked, also leading to the absence of data evaluation and processing\nmethods in ML fields. Motivated by the challenge and need, we selected\nan important but difficult reorganization energy (RE) prediction task\nas a test platform, which is an important parameter for the charge\nmobility of organic semiconductors (OSCs), to propose a data-quality-navigated\nstrategy with chemical intuition. We developed a data diversity evaluation\nbased on structure characteristics of OSC molecules, a reliability\nevaluation method based on prediction accuracy, a data filtering method\nbased on the uncertainty of <i>K</i>-fold division, and\na data split technique by clustering and stratified sampling based\non four molecular descriptor-associated REs. Consequently, a representative\nRE data set (15,989 molecules) with high reliability and diversity\ncan be obtained. For the feature representation, a complementary strategy\nis proposed by considering the chemical nature of REs and the structure\ncharacteristics of OCS molecules as well as the model algorithm. In\naddition, an ensemble framework consisting of two deep learning models\nis constructed to avoid the risk of local optimization of the single\nmodel. The robustness and generalization of our model are strongly\nvalidated against different OSC-like molecules with diverse structures\nand a wide range of REs and real OSC molecules, greatly outperforming\neight adversarial controls. Collectively, our work not only provides\na quick and reliable tool to screen efficient OSCs but also offers\nmethodological guidelines for improving the generalization of ML.