EXTENDED ABSTRACT: Traditionally, inhibitors have been based on chromate or other toxic compounds that are being banned by legislation around the world. Small hetero-cyclic compounds are a promising alternative, but, as the exact molecular structure is critical to determining the inhibition efffciency, there are literally tens of thousands of possibilities. Various methods, including high throughput experimentation and computational modelling, have been developed to select or design the optimum molecular structure. A very promising approach is the use of inverse design. In this approach, high throughput experiments defining electrochemical performance and computation methods defining inhibitor characteristics and attributes are linked by a machine learning algorithm to define the molecular attributes most critical for inhibition performance. These critical molecular attributes are then used to search molecular databases and select promising candidate inhibitors that are then subject to testing for verification. Early work was able to use quantitative structural activity relationships (QSAR) methods based on a neural network approach to obtain reasonable models of the features controlling inhibition. However, while these models could represent existing data, they were not very effective in predicting performance of new molecules. It was considered that this may have been due to either the relevance of molecular characterization and attribute deffnitions or the size and coverage of both the computational and experimental database. In order to develop more relevant feature deffnition an extensive modelling program was undertaken using molecular modelling but including for the ffrst time the effect of both applied potential and solvent and this was integrated with feature engineering based on tokenizing Smiles strings, use of fragmented molecular structures, convoluted neural ffngerprints and global descriptors. Devise data preprocessing techniques such as data augmentation and data transformation to render it into gaussian form were used.

Keywords: Inhibitors, Inverse Design, Molecular Modelling, Machine Learning