Partha Sarathi Dutta1,2,Aditya Koneru1,2,Adil Muhammed1,2,Sukriti Manna1,2,Karthik Balasubramanian1,2,Henry Chan2,Troy David Loeffler2,1,Subramanian Sankaranarayanan1,2
University of Illinois at Chicago1,Argonne National Laboratory2
Partha Sarathi Dutta1,2,Aditya Koneru1,2,Adil Muhammed1,2,Sukriti Manna1,2,Karthik Balasubramanian1,2,Henry Chan2,Troy David Loeffler2,1,Subramanian Sankaranarayanan1,2
University of Illinois at Chicago1,Argonne National Laboratory2
Silicene is a promising two-dimensional material consisting of a single layer of silicon, has a honeycomb lattice configuration, and has many similarities with graphene. Silicene has potential applications in transistors, flexible optoelectronic applications, energy storage, catalysis, sensors, quantum computing, spintronics, etc. There exist many potential models to perform dynamical simulations of Silicene. However, most of these models have been trained to a limited number of ab-initio data and therefore do not capture all the properties of all the Silicene polymorphs. Using a computationally cheap Tersoff formalism, we demonstrate that one can significantly improve upon the existing parameterizations using the advances in ML approaches. The training data for five different polymorphs of Silicene were computed using ab-initio calculations and our objective was to fit the crystal lattice structure, cohesive energy, equation of state, and elastic constants of all the polymorphs as well as the phonon dispersion of the most stable polymorph. We choose an in-house reinforcement learning (RL) based continuous-Monte Carlo Tree Search (c-MCTS) to solve this complex, non-linear, high-dimensional, multi-objective problem and get the optimum tersoff potential parameters. The optimization problem involves a trade-off between exploitation (navigating local search space of parameters for better fitting) and exploration (searching unexplored search space with the hope of getting better parameters). We use a reward mechanism wherein the reward was decided based on the agreement between the target property and the predicted property for a given set of potential model parameters. Our RL-trained potential parameters improve upon the performance of several existing models and adequately describe many of the properties of Silicene. We will discuss and compare the performance of our model with state-of-the-art physics-based and ML models for Silicene. We demonstrate that the RL-trained model performs very well in describing the thermal and mechanical properties of Silicene polymorphs. This opens a plethora of opportunities for studying the mechanical flexibility (for flexible electronics), thermal management, thermoelectric applications, strain engineering (for strain-tailored devices), and nanoelectromechanical systems (NEMS) of Silicene.