Predicting Corrosion Inhibition Efficiency Using Large Language Models: A New Table Embedding Method for Small Datasets

This study proposes an innovative framework that uses table embedding technology of large language models (LLMs) to solve the problem of small dataset learning in corrosion inhibition efficiency prediction, opening up a new path for AI applications in materials science and industrial anti-corrosion fields. The method encodes chemical structures and experimental conditions into tables, leverages the representation capabilities of LLMs to extract deep features, achieves high-precision prediction under small samples, and provides open-source datasets and code for easy reproduction.

Corrosion causes hundreds of billions of dollars in economic losses annually. Fields such as petrochemicals rely on corrosion inhibitors, but traditional prediction methods have problems like high experimental costs, long cycles, and poor performance of traditional machine learning on small datasets. LLMs have made significant breakthroughs in natural language processing, but their application in materials science and chemistry is still in the exploratory stage. How to transfer their representation capabilities to corrosion inhibition prediction is the focus of research.

1. Table Embedding Strategy: Represent samples as structured tables containing molecular structures, experimental conditions, and concentration parameters, using the semantic understanding ability of LLMs to learn feature correlations; 2. Small Dataset Optimization: Use transfer learning to leverage pre-trained knowledge of LLMs, achieving high prediction accuracy with hundreds of samples; 3. End-to-End Process: An automated process from raw data to prediction results, facilitating engineering applications.

Experiments show that this method outperforms traditional algorithms such as random forests and support vector machines in small dataset scenarios, especially when the number of samples is less than 500. Ablation experiments verify the effectiveness of the table embedding strategy. The pre-trained knowledge of LLMs is crucial for extracting chemical representations, and simple table encoding cannot achieve the same effect.

It provides a new tool for rapid screening of corrosion inhibitors, greatly shortening the R&D cycle and reducing costs, and has important economic value for industries such as petrochemicals and offshore platforms. The generality of the method can be extended to tasks such as catalyst activity prediction and drug molecular property prediction, providing a reference solution for small-data scientific problems.

Limitations: The interpretability of the model needs to be enhanced, and the generalization ability under extreme conditions needs to be verified. Future directions: Integrate multi-modal information (molecular images, spectra), develop domain-adaptive methods, establish larger-scale corrosion databases, and promote AI-driven material design.

This study demonstrates the potential of cross-integration between AI and materials science. By solving the small dataset problem through LLM table embedding technology, it brings new technical options for industrial anti-corrosion. With the release of open-source code and datasets, we look forward to more researchers joining to jointly promote the development of AI for Science.