Cobalt (Co) has emerged as a promising interconnect material to replace copper. Due to its susceptibility to corrosion, controlling the balance between complexing agents and inhibitors during chemical mechanical polishing (CMP) is crucial. Molecular dynamics (MD) simulation can reveal the atom removal mechanisms in CMP. However, existing force field parameters fail to accurately describe the complex interactions between Co and C/N groups in liquids. This study aims to design a closed-loop paradigm based on an active learning framework to generate the data set and optimize the ReaxFF parameters for Co CMP, and to reveal the atom removal mechanism by using the MD method. First, high-precision density functional theory (DFT) calculations for some basic configurations were performed to generate energy and force data as the training set. Systems involving key chemical groups during the Co-CMP process, and the adhesion of corrosion inhibitors triazole (TAZ) and benzotriazole (BTA), and complexing agents such as citric acid (CA) and glycine (GLY), were examined. Then, the ReaxFF parameters were optimized through machine learning within the Jax-ReaxFF framework. Inspection of the MD results was used to check the final structure, and unacceptable structures were sent back to DFT calculations for new training set generation. MD simulation results were compared with DFT analysis and experimental verification to demonstrate the reliability of the ReaxFF-MD. The adsorption geometries, charge distribution, and RDF results reveal that GLY has a stronger complexing ability, and BTA is a better corrosion inhibitor than TAZ. Both physisorption and chemisorption affect the adsorption. The “steric blocking plus stable bonding” is the key mechanism to control corrosion. This research not only provides a more accurate force field description for Co CMP but also offers a new simplified sampling scheme for the development of force field parameters.
Huang et al. (Sat,) studied this question.