Data-driven construction of machine-learning-based interatomic potentials for gas-surface scattering dynamics: the case of NO on graphite

Technical Deep Dive: Data-driven Interatomic Potentials for NO Scattering on Graphite

Overview

This work presents a data-driven workflow for constructing a machine-learning interatomic potential (MLIP) tailored to gas-surface scattering dynamics, using nitric oxide (NO) scattering from highly oriented pyrolytic graphite (HOPG) as a benchmark system. The key components are:

• Controlled sampling of ab initio molecular dynamics (AIMD) data using descriptor-based methods to build a compact, representative training set.
• Active learning to iteratively refine the MLIP by selectively introducing new DFT calculations in underrepresented regions.
• Extensive classical molecular dynamics (MD) simulations using the MLIP to investigate NO scattering from graphite over broad ranges of collision energies and surface temperatures.

The resulting MLIP achieves high-fidelity agreement with DFT while enabling efficient large-scale trajectory sampling to provide detailed atomistic insight into the NO-graphite scattering mechanisms.

Methodology

• Performed AIMD simulations of NO scattering on HOPG at 100 K and 300 K, with incident energies of 0.1 eV and 0.3 eV.
• Represented local atomic environments using SOAP descriptors and performed dimensionality reduction via PCA.
• Employed farthest point sampling (FPS) in the reduced descriptor space to build a compact, diverse training set (dataset A) from the AIMD data.
• Trained an initial committee of Deep Potential MLIP models on dataset A.
• Conducted classical MD simulations using one MLIP model to generate additional candidate configurations.
• Selectively labeled new configurations exhibiting high model uncertainty via DFT calculations, and refined the MLIP (dataset B).

Results

• The FPS-sampled dataset A captured the essential configurational diversity, with only 6,671 configurations (0.9% of full AIMD data).
• The final MLIP trained on dataset B achieved high accuracy, with energy RMSE of 0.0601 eV and force RMSE of 0.0334 eV/Å on the validation set.
• MD simulations using the MLIP provided detailed insights into NO scattering on graphite:
  • Scattering probability increases from ~7% at 0.05 eV to near-unity above 1 eV.
  • Scattered molecules lose 50-82% of their initial kinetic energy, with higher losses at lower incident energies.
  • Polar angular distributions evolve from broad to strongly forward-peaked as incident energy increases.
  • Rotational excitation increases with incident energy, exhibiting a transition from near-thermal to non-Boltzmann distributions.

Interpretation

• The transition from trapping-mediated to direct scattering regimes reflects the interplay between the shallow physisorption well depth and the incident kinetic energy.
• The substantial translational energy loss is attributed to efficient momentum transfer during the collision, leading to partial thermalization of the scattered molecules.
• The progressive narrowing and forward focusing of the angular distributions indicates a transition toward more impulsive, specular-like scattering at higher energies.
• Rotational excitation dynamics reveal a combination of moderate torques leading to broad, thermal-like distributions and highly impulsive collisions resulting in rotational rainbow scattering.

Limitations & Uncertainties

• The simulations were limited to normal incidence and did not fully capture the influence of surface temperature on the relative contributions of diffuse and specular scattering components.
• Longer simulation times may be required to fully resolve the trapping-desorption dynamics, especially at lower surface temperatures.
• While the MLIP demonstrates excellent accuracy, it was trained and validated against a finite AIMD dataset and may not fully capture rare or unexplored configurations.

What Comes Next

• Extend the workflow to other gas-surface systems, including more complex adsorbates and reactive processes.
• Investigate the impact of surface defects, steps, and other heterogeneities on the scattering dynamics.
• Explore the coupling between rotational, vibrational, and translational degrees of freedom in the energy transfer mechanisms.
• Incorporate the MLIP into multiscale modeling approaches to bridge the gap between atomistic simulations and macroscopic experimental observables.