Sources of residual errors at chemical-accuracy regime for FiRE

Determine whether the residual errors in the finite-range embeddings (FiRE) neural-network variational Monte Carlo predictions for non-covalent interaction energies, singlet–triplet gaps, and ionization potentials primarily arise from (i) inaccuracies in high-level reference methods such as CCSD(T), (ii) discrepancies between 0 K gas-phase calculations and experimental measurement conditions, or (iii) unaccounted structural relaxation effects that can alter relative energies.

Background

The paper demonstrates that FiRE achieves high accuracy across several challenging benchmarks, including non-covalent interactions (S22), singlet–triplet gaps in acenes, and organometallic ionization potentials, often matching or improving upon CCSD(T), AFQMC, and other NN-VMC approaches.

Despite this accuracy, the authors note that small residual discrepancies persist and explicitly raise uncertainty about their origin—whether they stem from limitations of reference methods (e.g., CCSD(T)), mismatches between 0 K gas-phase theory and experimental conditions, or structural relaxations affecting relative energies. Clarifying these sources would refine benchmarking practices and guide further methodological improvements.

References

At this accuracy, it is unclear whether the remaining errors are due to errors in references, e.g., CCSD(T) errors, comparing 0 K gas phase to experimental conditions, or structural relaxations which may affect relative energies.

— Accurate Ab-initio Neural-network Solutions to Large-Scale Electronic Structure Problems (2504.06087 - Scherbela et al., 8 Apr 2025) in Results, Subsection 2.2 Accurate relative energies (end of subsection)