Atomic cluster expansion potential for the Si-H system

TL;DR

This work develops SiH-ACE-25, a nonlinear ACE-based interatomic potential for the Si--H system, enabling accurate, large-scale simulations of crystalline, amorphous, surface, and molecular configurations in $a$-Si:H. An active-learning-driven training dataset spanning diverse Si--H environments is used to parameterize a custom asymmetric ACE model, with rigorous validation against DFT and selective experimental observables. SiH-ACE-25 shows strong performance for bulk amorphous/crystalline phases, reasonable defect energetics, and reliable surface and molecular predictions, while illustrating limitations for unseen pristine surfaces. The model enables efficient, device-scale MD studies of hydrogenated silicon, including structure–property relationships under mechanical deformation, and the accompanying dataset provides a resource for benchmarking and further MLIP development.

Abstract

The silicon-hydrogen system is of key interest for solar-cell devices, including both crystalline and amorphous modifications. Elemental amorphous Si is now well understood, but the atomic-scale effects of hydrogenating the silicon matrix remain to be fully explored. Here, we present a machine-learned interatomic potential model based on the atomic cluster expansion (ACE) framework that can describe a wide range of Si-H phases, from crystalline and amorphous bulk structures to surfaces and molecules. We perform numerical and physical validation across a range of hydrogen concentrations and compare our results to experimental findings. Our work constitutes an advancement toward the exploration of large structural models of a-Si:H at realistic device scales.

Figures (10)

• Figure 1: Simplified overview of the approach for MLIP training dataset construction, including iterative exploration and ACE model fitting, adopted in the present study.
• Figure 2: The SiH-25 dataset. (a) Training dataset evolution over active learning iterations, color-coded by hydrogen content, visualized using UMAP mcinnes_umap_2018 and SOAP descriptors Bartok-13-05. (b) Projected map of the full dataset, where each point represents a structure in the training set and is color-coded by structure type. Representative structures of each configurational type, visualized using OVITO ovito, are shown as insets.
• Figure 3: (a) Combined energy and force loss against number of elemental basis functions for Si (top), H (middle) and SiH (bottom), where we plot the average of five randomly initiated models and the standard deviation in gray. Colored vertical dashed line indicate the basis sizes for three models of increasing cost. (b) Energy and force RMSE as a function of dataset size for three models of increasing cost, where we systematically pick the best of 5 randomly initiated models, presented as log--log plots. (c) Evolution of box size against simulation time for three models of increasing cost, plotting the average and the standard deviation of five repeats for the "symmetric" models (top) and "asymmetric" models (bottom).
• Figure 4: Average partial structure factors of annealed $a$-Si:H structures, showing Si--Si (top), Si--H (middle), and H--H (bottom) contributions. Structure factors are averaged over ten repeats at each concentration. Experimental measurements from Ref. bellisent_structure_1989 for a sample with $\sim$16 at.-% H are plotted as dashed black lines, normalized for direct comparison.
• Figure 5: (a) Structure visualization of a pore in an $a$-Si:H sample with 10 at.-% H, showing the pore edges shaded in gray using the OVITO alpha-shape mesh modifier ovitostukowski_computational_2014. (b) Fractional pore volume in the $a$-Si:H structures as a function of hydrogen content. (c) Count of the occurrence of SiH$_{n}$ fragments in the $a$-Si:H structures relative to the total number of Si atoms, for $n$ taking values from 0 to 4, as a function of hydrogen content in the structures. (d) Count of the occurrence of three different H fragments relative to the total number of H atoms as a function of hydrogen content in the structures, showing the occurrence of H atoms bonded to one Si atom (top, yellow), H atoms bonded to one H atom to form a H$_2$ molecule (middle, pink), and bond-centered H atoms, bonded to two Si atoms (bottom, blue). Visualizations are added to illustrate the structural motifs. We report averaged results over ten repeats at each concentration; shading indicates the standard deviations across repeats.