We develop a neuroevolution-potential (NEP) framework for generating neural network-based machinelearning potentials. They are trained using an evolutionary strategy for performing large-scale molecular dynamics (MD) simulations. A descriptor of the atomic environment is constructed based on Chebyshev and Legendre polynomials. The method is implemented in graphic processing units within the open-source GPUMD package, which can attain a computational speed over 10 7 atom-step per second using one Nvidia Tesla V100. Furthermore, per-atom heat current is available in NEP, which paves the way for efficient and accurate MD simulations of heat transport in materials with strong phonon anharmonicity or spatial disorder, which usually cannot be accurately treated either with traditional empirical potentials or with perturbative methods.

We study the structural and mechanical properties of nanoporous (NP) carbon materials by extensive atomistic machine-learning (ML) driven molecular dynamics (MD) simulations. To this end, we retrain a ML Gaussian approximation potential (GAP) for carbon by recalculating the a-C structural database of Deringer and Csányi adding van der Waals interactions. Our GAP enables a notable speedup and improves the accuracy of energy and force predictions. We use the GAP to thoroughly study the atomistic structure and pore-size distribution in computational NP carbon samples. These samples are generated by a melt-graphitization-quench MD procedure over a wide range of densities (from 0.5 to 1.7 g/cm3) with structures containing 131 072 atoms. Our results are in good agreement with experimental data for the available observables and provide a comprehensive account of structural (radial and angular distribution functions, motif and ring counts, X-ray diffraction patterns, pore characterization) and mechanical (elastic moduli and their evolution with density) properties. Our results show relatively narrow pore-size distributions, where the peak position and width of the distributions are dictated by the mass density of the materials. Our data allow further work on computational characterization of NP carbon materials, in particular for energy-storage applications, as well as suggest future experimental characterization of NP carbon-based materials.

Amorphous silicon (a-Si) is an important thermal-management material and also serves as an ideal playground for studying heat transport in strongly disordered materials. Theoretical prediction of the thermal conductivity of a-Si in a wide range of temperatures and sample sizes is still a challenge. Herein we present a systematic investigation of the thermal transport properties of a-Si by employing large-scale molecular dynamics (MD) simulations with an accurate and efficient machine learned neuroevolution potential (NEP) trained against abundant reference data calculated at the quantum-mechanical density-functional-theory level. The high efficiency of NEP allows us to study the effects of finite size and quenching rate in the formation of a-Si in great detail. We find that a simulation cell up to 64 000 atoms (a cubic cell with a linear size of 11 nm) and a quenching rate down to 10 11 K s −1 are required for almost convergent thermal conductivity. Structural properties, including short-and medium-range order as characterized by the pair-correlation function, angular-distribution function, coordination number, ring statistics, and structure factor are studied to demonstrate the accuracy of NEP and to further evaluate the role of quenching rate. Using both the heterogeneous and homogeneous nonequilibrium MD methods and the related spectral decomposition techniques, we calculate the temperature-and thickness-dependent thermal conductivity values of a-Si and show that they agree well with available experimental results from 10 K to room temperature. Our results also highlight the importance of quantum effects in the calculated thermal conductivity and support the quantum-correction method based on the spectral thermal conductivity.