Speaker
Description
Hydrogenation of graphene is a promising route toward tailoring its electronic properties, including opening the band gap required for transistor applications and advancing hydrogen-storage platforms. Industrial processing techniques—such as exposure to atomic beams, plasmas, or chemical functionalization—depend on the consecutive deposition of individual atomic species onto the surface. As local adatom concentrations increase, cooperative electronic interactions, localized lattice strain, and hydrogen-bonding networks dynamically alter the energy landscape for subsequent adsorption events. However, the precise atomistic mechanisms governing this dynamic coverage evolution, local cluster formation, and structural reorganization remain poorly understood.
Addressing this knowledge gap presents severe computational and experimental challenges. Direct real-time experimental observation of individual atomic adsorption events under processing conditions remains unfeasible. On the theoretical side, density functional theory (DFT) is limited to small unit cells or unrealistically ordered periodic configurations that fail to reflect the disordered, low-coverage topologies present at the beginning of a deposition process. Conversely, classical molecular dynamics (MD) lacks the accuracy required to model bond breaking and shifting atomic interactions as local concentrations change.
Integrating molecular dynamics with on-the-fly machine learning from DFT data provides the necessary framework to overcome these limitations. A machine-learned interatomic potential (MLIP) for hydrogenated graphene was developed using on-the-fly learning with Bayesian uncertainty quantification implemented in the Vienna Ab initio Simulation Package (VASP). The training dataset incorporated pristine graphene as well as structures containing Stone–Wales and monovacancy defects. The potential was iteratively refined through active learning and rigorous validation on adsorption, desorption, and cluster-formation processes.
The final MLIP reproduces DFT energies and forces with high accuracy while reducing the computational cost by two to three orders of magnitude. Reliable simulations were achieved for hydrogen coverages up to approximately 10%. By combining quantum-level precision with the efficiency needed for large-scale systems, this approach captures realistic, non-periodic surface arrangements, avoids unphysical static geometries, and maps the true dynamic evolution of low-ordered, low-coverage adsorbate systems under experimental conditions. These results demonstrate that machine-learning interatomic potentials provide an efficient, highly accurate framework for studying adatom dynamics and enable large-scale investigations into adsorption-driven modifications of graphene's electronic properties.