SpecSwap RMC: A novel reverse Monte Carlo approach using a discrete configuration space and pre-computed properties

We present a novel approach to reverse Monte Carlo (RMC) modeling, SpecSwap-RMC, which makes use of pre-computed property data from a discrete configuration space replacing atomistic moves with swap moves of contributions to a sample-set representing the average, or summed property. The approach is particularly suitable for disordered systems and properties which require significant computer time to compute. We demonstrate the approach by fitting jointly and separately the EXAFS signal and x-ray absorption spectrum (XAS) of ice Ih using as SpecSwap sample-set 80 configurations from a space of 1382 local structures with associated pre-computed spectra. As an additional demonstration we compare SpecSwap and FEFFIT fits of EXAFS data on crystalline copper finding excellent agreement.

💡 Research Summary

The paper introduces SpecSwap‑RMC, a novel variant of the reverse Monte Carlo (RMC) method that replaces conventional atom‑by‑atom moves with “swap” operations performed on a pre‑computed discrete configuration space. Traditional RMC iteratively perturbs atomic coordinates and, after each perturbation, recomputes the observable (e.g., EXAFS, XAS) using an electronic‑structure code. When the observable is computationally expensive, this approach becomes prohibitive, especially for disordered or large systems. SpecSwap‑RMC circumvents this bottleneck by first generating a finite set of representative local structures (the configuration space) and calculating the target property for every member of this set in advance. In the Monte Carlo sampling stage, a “sample‑set” of a fixed number of configurations is maintained; at each step a configuration in the sample‑set is swapped with another candidate from the pool. The average property of the sample‑set is compared to the experimental data, and the swap is accepted or rejected according to the Metropolis criterion. Because the property values are already known, each Monte Carlo step requires only a simple arithmetic update of the average, giving an O(1) computational cost per step.

The authors demonstrate the method on two distinct problems. First, they model the combined EXAFS and X‑ray absorption near‑edge structure (XANES) of hexagonal ice (Ih). A pool of 1 382 local ice configurations was generated, each with a FEFF‑computed spectrum. From this pool a sample‑set of 80 configurations was used in the RMC simulation. The resulting averaged spectra reproduce the experimental EXAFS oscillations and the XANES edge shape with high fidelity, capturing subtle variations in the hydrogen‑bond network that would be difficult to resolve with standard RMC due to the high cost of repeated FEFF calculations. Second, they apply SpecSwap‑RMC to crystalline copper, fitting its EXAFS data and comparing the outcome with the widely used FEFFIT program. The fitted nearest‑neighbor distance, Debye‑Waller factor, and overall R‑factor from SpecSwap‑RMC match those obtained by FEFFIT, while the total computational time is markedly lower because the method never recalculates the scattering paths during the fit.

Key advantages of SpecSwap‑RMC are: (1) drastic reduction of computational expense, since the expensive forward calculation is performed only once per configuration; (2) exact reproducibility of the forward model because the same pre‑computed spectra are used throughout the fitting; (3) natural suitability for disordered systems where a diverse pool of local motifs can be assembled beforehand. The paper also discusses limitations. The quality of the final model depends critically on the completeness and representativeness of the initial configuration pool; constructing such a pool may require expert intuition or automated sampling strategies. Moreover, the pre‑computed spectra are generated at specific thermodynamic conditions; mismatches with experimental temperature, pressure, or electronic state may necessitate additional correction terms. Finally, the current objective function focuses on matching the average property, so higher‑order correlations (e.g., multi‑atom clustering) are not directly constrained.

In conclusion, SpecSwap‑RMC offers a powerful framework for fitting spectroscopic data that are costly to compute, opening the door to routine RMC‑style modeling of XAS, EXAFS, and potentially other observables such as neutron scattering or vibrational spectra. Future work could integrate machine‑learning‑driven generation of configuration pools, simultaneous fitting of multiple observables, and extensions to incorporate variance‑based constraints, thereby broadening the applicability of the method to complex materials, nanostructures, and soft‑matter systems.