Realistic atomistic structure of amorphous silicon from machine-learning-driven molecular dynamics (1803.02802v1)

Abstract: Amorphous silicon (a-Si) is a widely studied non-crystalline material, and yet the subtle details of its atomistic structure are still unclear. Here, we show that accurate structural models of a-Si can be obtained by harnessing the power of machine-learning algorithms to create interatomic potentials. Our best a-Si network is obtained by cooling from the melt in molecular-dynamics simulations, at a rate of 10$^{11}$ K/s (that is, on the 10 ns timescale). This structure shows a defect concentration of below 2% and agrees with experiments regarding excess energies, diffraction data, as well as $^{29}$Si solid-state NMR chemical shifts. We show that this level of quality is impossible to achieve with faster quench simulations. We then generate a 4,096-atom system which correctly reproduces the magnitude of the first sharp diffraction peak (FSDP) in the structure factor, achieving the closest agreement with experiments to date. Our study demonstrates the broader impact of machine-learning interatomic potentials for elucidating accurate structures and properties of amorphous functional materials.

Realistic Atomistic Structure of Amorphous Silicon from Machine-Learning-Driven Molecular Dynamics

The research presented investigates the atomistic structure of amorphous silicon (a-Si) utilizing advanced machine-learning-driven molecular dynamics simulations. This paper introduces structural models of a-Si derived from machine-learning (ML) based interatomic potentials using the Gaussian approximation potential (GAP) and Smooth Overlap of Atomic Positions (SOAP) frameworks. The methodology developed in this research facilitates the creation of a-Si models by quenching from the melt at a cooling rate of $10^{11}$ K/s, a process which aligns more closely with experimental conditions when compared to traditional quantum mechanical simulations.

The most notable contribution of the paper is the achievement of defect concentrations below 2% in the simulated a-Si structures, demonstrating high fidelity to experimental data regarding excess energies and diffraction measurements. Additionally, the research highlights that achieving such an agreement is not feasible through faster quench simulations, thereby underscoring the significance of the slower ML-driven quenching process.

The simulations were conducted on a system of 4,096 atoms, which successfully replicated the magnitude of the first sharp diffraction peak in the structure factor, matching to a significant degree the empirical observations available. This advancement affirms the capability of ML-driven simulations to address the inherent limitations of traditional density-functional theory (DFT) or classical force fields, which either face restrictions in computational cost or lack the necessary accuracy to predict the structural nuances of amorphous materials.

The implications of this research extend to practical and theoretical realms, suggesting enhanced insights into the structure-property relationships of disordered materials like a-Si. This advancement is particularly pertinent for applications in photovoltaics, thin-film transistors, and battery electrodes where comprehension of local atomic environments and defect structures can profoundly impact electronic and macroscopic properties.

Experimentally validating the structures, the paper’s results align closely with calorimetry, the magnitude of chemical shifts in silicon solid-state NMR, and the structural insights offered by X-ray diffraction patterns. For instance, the NMR data from the structures quenched at $10^{11}$ K/s show a chemical shift consistent with well-annealed a-Si samples known from experimental data.

By scaling to larger systems using ML potentials, the research asserts the feasibility of simulating structures on the nanometer scale, thus bridging the substantial gap between small-scale quantum simulations and large-scale empirical modeling. This simulation approach further represents a qualitative leap in providing cost-effective simulations that mirror experimental conditions more closely, as demonstrated by significant cost reductions from millions of core-hours in DFT to tens of thousands in the ML-driven predictions.

Future directions in this domain suggest an expansion of such ML methodologies to other complex disordered systems, where traditional simulation methods struggle with size and cost constraints. Extending this ML-driven framework could offer new insights and predictive capabilities for various amorphous materials, potentially revolutionizing our understanding and engineering of such materials for advanced technological applications.