Calculations of Potential Energy Surfaces Using Monte Carlo Configuration Interaction

We apply the method of Monte Carlo configuration interaction (MCCI) to calculate ground-state potential energy curves for a range of small molecules and compare the results with full configuration interaction. We show that the MCCI potential energy curve can be calculated to relatively good accuracy, as quantified using the non-parallelity error, using only a very small fraction of the FCI space. In most cases the potential curve is of better accuracy than its constituent single-point energies. We finally test the MCCI program on systems with basis sets beyond full configuration interaction: a lattice of fifty hydrogen atoms and ethylene. The results for ethylene agree fairly well with other computational work while for the lattice of fifty hydrogens we find that the fraction of the full configuration interaction space we were able to consider appears to be too small as, although some qualitative features are reproduced, the potential curve is less accurate.

💡 Research Summary

The paper investigates the performance of Monte Carlo Configuration Interaction (MCCI) as a compact, stochastic alternative to full configuration interaction (FCI) for generating ground‑state potential energy curves. The authors first describe the MCCI algorithm: a random sampling of Slater determinants (or configuration state functions) is performed, and only those whose CI coefficients exceed a predefined cut‑off are retained. The resulting truncated CI matrix is diagonalised, and the process is iterated until convergence. By discarding the vast majority of determinants, MCCI dramatically reduces the dimensionality of the problem while still aiming to capture the most important static and dynamic correlation effects.

A series of benchmark calculations is carried out on small diatomic and tri‑atomic molecules (H₂, N₂, CO, H₂O, etc.) across a range of bond lengths. For each geometry the MCCI energy is compared with the exact FCI energy obtained in the same basis set. Accuracy is quantified primarily by the non‑parallelity error (NPE), defined as the difference between the maximum and minimum absolute deviations of the MCCI curve from the FCI curve. The results show that, despite using only 0.01–0.1 % of the full determinant space (typically a few thousand determinants), MCCI achieves NPE values of 1–2 kcal mol⁻¹ for most systems. Individual point errors can be larger (3–5 kcal mol⁻¹), but the errors tend to cancel along the curve, leading to a smoother and more reliable description of the entire potential energy surface than would be expected from the pointwise accuracy alone.

To test the scalability of the method, the authors extend the study to two larger systems: a linear chain of fifty hydrogen atoms (H₅₀) and ethylene (C₂H₄) with a 6‑31G* basis. For H₅₀, the full CI space is astronomically large, making exact reference data unavailable. MCCI is nevertheless applied using a few million determinants, which represents an infinitesimal fraction of the total space. The resulting curve reproduces qualitative features such as the overall shape and the dissociation limit, but the NPE rises to >10 kcal mol⁻¹, indicating that the sampling density is insufficient to capture the subtle long‑range correlation present in this extended system. In contrast, for ethylene the MCCI curve, built from roughly 0.5 % of the full space, matches closely with high‑level coupled‑cluster (CCSD(T)) and multireference CI results, confirming that the method remains accurate when the system size is moderate and the correlation is not overwhelmingly delocalised.

The authors conclude that MCCI offers a highly efficient route to accurate potential energy surfaces for small‑ to medium‑sized molecules, especially when the primary interest lies in relative energies along a reaction coordinate rather than absolute single‑point energies. Its stochastic nature automatically adapts the determinant selection to changes in geometry, which is advantageous for scanning bond‑breaking processes. However, the study also highlights limitations: as the system grows, the fraction of the determinant space that must be sampled to retain chemical accuracy increases sharply, and the choice of the coefficient cut‑off becomes critical. Future work is suggested in the direction of adaptive cut‑off schemes, hybridisation with deterministic selected‑CI approaches, and exploiting modern parallel and GPU architectures to accelerate the Monte Carlo sampling. With these improvements, MCCI could become a practical bridge between inexpensive mean‑field methods and the prohibitive cost of full CI for larger, chemically relevant systems.