Text Embedding-Assisted Prediction of Corrosion Inhibition Efficiency Using Machine Learning under Small-Sample Conditions
Year: 2026
Type: preprint
Abstract: Machine learning has been increasingly applied to corrosion prediction, but limited experimental data remain a challenge for small-sample prediction. This study investigates how text embedding can be used as a molecular representation for predicting the corrosion inhibition efficiency of organic inhibitors for ZE41 magnesium alloy. An experimental dataset containing 75 inhibitors was used, with 60 samples for training and 15 samples for testing. Conventional molecular descriptors were combined with text embeddings generated using ChemBERTa and Qwen-text-embedding-v4, and five feature construction strategies were compared to examine the effects of embedding dimensionality and feature integration. Direct incorporation of high-dimensional ChemBERTa embeddings provided limited improvement, whereas dimensionality reduction improved the prediction performance. A selective hybrid strategy was further investigated, in which selected descriptors were incorporated into the text embedding while other descriptors were retained in their original numerical form. This strategy achieved a coefficient of determination of 0.61 and a Pearson correlation coefficient of 0.79, compared to those of 0.12 and 0.43, respectively, for the conventional descriptor-based representation. The results suggest that text embedding may be more useful as a complementary representation to conventional molecular descriptors than as a direct replacement, when experimental data are limited.
Source: ChemRxiv
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204275000
