AI-driven discovery of high-performance corrosion inhibitors using a BERT-GPT framework for molecular generation

Abstract

The discovery of high-efficiency corrosion inhibitors is constrained by the structural diversity of existing molecular entities. This study introduces a framework that integrates BERT and GPT models to predict, screen, and generate molecules with potential corrosion inhibiting properties. Predictive models for inhibition efficiency (IE) and toxicity on carbon steel surfaces in 1 M HCl environments were built by utilizing chemical language embeddings and molecular descriptors. The screening of the PubChem library identified 1,336,971 candidate molecules with a predicted IE greater than 90 % at a 1 mM concentration and a predicted LD50 (lethal dose for 50 % of test animals) exceeding 2000 mg/kg (indicating low acute toxicity). Scaffold analysis highlighted an enrichment of critical scaffold, particularly 6 imidazole scaffolds in conjunction with property constraints (IE = 90 %, LD50 = 1000 mg/kg), informed the conditional generation of molecules using MolGPT. A total of 310,962 valid molecules were generated, with 25.17 % (78,258 molecules) satisfying all specified conditions, which is much higher than the hit ratio (4.08 %) of screening from PubChem database. This study demonstrates an AI-driven molecular generation for corrosion inhibitor discovery.