Use of Local Environments in Training Machine Learning Potentials for Low-Dimensional Carbon Nanostructures

Machine learning applications have become increasingly prominent in the field of condensed matter physics over the past decade. These methods enable the estimation of electronic properties of materials without relying on computationally expensive first-principles calculations. Similarly, molecular dynamics potentials are now being replaced by machine learning potentials trained on datasets obtained from first-principles simulations. However, the main challenge concerning these potentials is their limited transferability, as they are generally trained to fit specific physical properties of particular systems. In this presentation, we introduce our novel approach to addressing these issues from a local environment perspective. In our previous research, we demonstrated that the vibrational properties of certain low-dimensional systems can be inferred from their local atomic environments. In this work, we first present our potential generation method for studying low-dimensional carbon nanomaterials. Since one of our main goals is to contribute to the characterization of low-dimensional amorphous structures, we also describe our training dataset. Finally, we discuss our preliminary results, highlighting the performance of the trained potentials in open problems such as the estimation of energy barriers for defect formation and migration.