Skip to content

Experiment-Driven Atomistic Materials Modeling: A Case Study Combining X-Ray Photoelectron Spectroscopy and Machine Learning Potentials to Infer the Structure of Oxygen-Rich Amorphous Carbon

Scope

Experiment-driven structure generation: grand-canonical Monte Carlo with a modified Hamiltonian couples an ML interatomic potential and an ML XPS model (trained on GW/DFT) to build oxygen-rich amorphous carbon (a-COₓ) models that match XPS and stay low in energy.

Summary

The authors address atomistic structure determination for disordered materials by jointly satisfying experimental X-ray photoelectron spectroscopy (XPS) and DFT-level energetics, instead of post hoc picking structures from generic MD/MC sampling. They introduce a workflow that combines machine-learned potentials (MLPs) with ML-predicted XPS (trained on GW and DFT references) inside grand-canonical Monte Carlo using a modified Hamiltonian / generalized dynamics formalism so that sampling steers toward experimental observables while remaining chemically reasonable (low energy). The case study is oxygenated amorphous carbon, a-COₓ, motivated by questions such as how much oxygen can be incorporated before decomposition to CO/CO₂, and by applications/tribology contexts noted in the introduction.

Methods

1 — MD application — N/A as primary engine: the workflow centers on sampling driven by MLIPs and ML XPS inside GCMC, not a classical production MD run described in the short extract.

2 — Force-field training / MLIP — ML interatomic potential (MLIP) supplies energies and forces; ML XPS maps atomic structure to spectra; both trained on DFT/GW-level references per the article (see PDF for architecture and data volume).

3 — Static QM / DFT (reference for ML). GW and DFT data underpin XPS training and energy labels; the extract stops early in XPS interpretation—N/A for a full DFT parameter table on this page.

4 — Sampling (grand-canonical Monte Carlo). GCMC with a modified Hamiltonian (generalized-dynamics / sampling formalism) couples ML energy models and ML XPS so that drawn a-CO\(_x\) configurations track XPS while remaining low in DFT-consistent energy (abstract / §2). The paper contrasts this with “generate many candidates then pick closest to experiment” pipelines as slow and not guaranteed. The extract in-repo ends early in XPS interpretation; move sets and full validation are in the PDF.

Findings

Outcomes, comparisons, and generality (abstract / introduction). a-CO\(_x\) models that fit experimental XPS while remaining favorable vs DFT-level energetics when using the ML XPS + MLIP pair. Network-based XPS deconvolution into structural motifs highlights limits of standard peak assignment and yields atomic-scale insight for oxygenated amorphous carbon. The workflow is framed as portable to other experimental observables beyond XPS.

Corpus honesty — see ## Limitations; RMSE/range numbers require the full JACS PDF.

Limitations

The checked-in extract is short relative to the full JACS paper; numerical benchmarks (RMSE on XPS, DFT energy windows, composition ranges) should be taken from the PDF. ML XPS and MLIPs inherit data and functional errors from their training sets.

Relevance to group

This is not a van Duin-group paper; it is included as corpus context on ML-driven structure inference and spectroscopy-informed sampling, adjacent to broader reactive MD / materials informatics themes.

Citations and evidence anchors