Extending density functional theory with near chemical accuracy beyond pure water

The study addresses the long-standing challenge of achieving chemically accurate, first-principles simulations of water and aqueous systems using density functional theory (DFT). While Kohn-Sham DFT offers an excellent accuracy-to-cost ratio, it has historically failed to reproduce experimental data for water. A recent breakthrough demonstrated that using the SCAN meta-GGA with density-corrected DFT (DC-DFT) greatly improves accuracy for pure water by mitigating density-driven errors. However, HF-SCAN without dispersion corrections cannot capture long-range noncovalent interactions, and naïve addition of dispersion destroys its accuracy for water. Additionally, SCAN suffers from slow grid convergence and numerical issues. The authors therefore aim to construct a practical, efficient functional that (1) preserves or improves chemical accuracy for pure water, (2) correctly describes dispersion-dominated noncovalent interactions (NCIs), and (3) avoids SCAN’s grid convergence issues. They propose HF-r²SCAN-DC4, combining HF densities (to reduce density-driven errors), r²SCAN (to regularize SCAN and fix grid issues), and a dispersion correction parameterized under DC-DFT principles.

Background includes: (i) SCAN meta-GGA’s strong performance across molecules and materials but numerical grid issues; (ii) density-corrected DFT (DC-DFT) framework separating density-driven from functional-driven errors and showing HF-DFT can improve energetics in density-sensitive cases (e.g., water clusters, barriers, halogen bonds); (iii) prior success of HF-SCAN for water clusters and liquid water when paired with many-body potentials (e.g., MB-pol), but its lack of explicit dispersion leads to poor performance on standard NCI datasets; (iv) r²SCAN regularizes SCAN and restores exact constraints, improving numerical robustness, but HF-r²SCAN alone is less accurate for water than HF-SCAN; (v) empirical dispersion corrections (e.g., Grimme’s D3/D4) are crucial for NCIs but can worsen results when improperly combined with density-sensitive self-consistent DFT; (vi) benchmark datasets WATER27 and GMTKN55 are standard for assessing water clusters and general main-group chemistry/NCIs; (vii) prior parameterizations (e.g., SM21) that fit D4 without respecting DC-DFT principles underperform on water clusters compared to a DC-aware fitting.

• Framework: Employ HF-DFT (evaluate the functional on Hartree-Fock densities and orbitals) to reduce density-driven errors. Use r²SCAN as the base meta-GGA to avoid SCAN’s grid convergence issues. Add an empirical D4-type dispersion term with parameters determined using DC-DFT principles, yielding HF-r²SCAN-DC4.
• DC-DFT principles: Separate density-driven from functional-driven errors. Fit empirical parameters only on density-insensitive (DI) reactions so that the fit targets functional errors, not density errors. Exclude density-sensitive (DS) cases from training; in water, DS errors can dominate.
• Parameterization of D4 (DC4): Train D4 parameters on the DI subset of GMTKN55 (weighted by MAE minimization) and use water–water interaction energies as a validation set. Exclude DS WATER27 clusters (sensitivity S > 2 kcal/mol) from fitting (retain them for unbiased testing). Minimize MAE over DI cases; details in Supplementary Note 4. The acronym DC4 indicates D4 parameters obtained via DC-DFT-informed fitting.
• Density sensitivity assessment: Compute the density sensitivity S (how much a DFT energy changes when the density is changed) to classify reactions as DS or DI. Use S to guide exclusion from training and to analyze performance trends (e.g., error growth with S for self-consistent methods).
• Datasets and test problems: WATER27 (binding energies of water clusters), water hexamer and 20-mer isomer energies, water dimer interaction energies (Smith stationary points and MD-sampled dimers), water–organic interactions (water–aspirin), stacked nucleobase dimers (cytosine), and broader NCI datasets from GMTKN55 (inter- and intramolecular).
• Many-body analysis: For water hexamers, perform many-body expansion (K-body, K=2–6) to dissect sources of errors and assess systematic behavior beyond 2-body terms.
• References and sampling: Use CCSD(T)/CBS or DLPNO-CCSD(T)-F12 as reference energies. Generate MD structures at T=298.15 K for water dimers and water–aspirin interactions. Compare HF-r²SCAN-DC4 against HF-SCAN, HF-r²SCAN, SC-r²SCAN, and SC-r²SCAN-D4.
• Implementation: A PySCF script for HF-r²SCAN-DC4 is provided. Numerical stability is improved relative to SCAN due to r²SCAN’s regularization, enabling routine use without extreme grids.

• Construction outcome: HF-r²SCAN-DC4 combines HF densities, r²SCAN, and DC-DFT-parameterized D4. Each component is essential; dropping any element degrades at least one target (water accuracy, NCIs, or numerical robustness).
• Water clusters and dimers:
  • Water hexamers: Recovers correct isomer ordering and improves relative energies vs HF-SCAN by up to ~0.7 kcal/mol; MAE in total interaction energies 0.19 kcal/mol vs 0.22 kcal/mol (HF-SCAN). Systematic, small overestimation (~0.2 kcal/mol) across isomers; gets ordering right where HF-SCAN fails (e.g., bag vs chair).
  • Water 20-mers: For relative isomer energies (density-insensitive), SC-r²SCAN-D4 beats HF-r²SCAN-DC4; however, HF-r²SCAN-DC4 still has much smaller errors than HF-SCAN and remains small on a per-molecule basis. Improvements over HF-SCAN reach up to 2.4 kcal/mol for 20-mers.
  • WATER27: Errors of SC-r²SCAN-D4 grow with density sensitivity S and can be worse than SC-r²SCAN. HF-r²SCAN-DC4 outperforms HF-SCAN by ~4 kcal/mol for the four most density-sensitive clusters. DI clusters (S ≤ 2 kcal/mol) were not used in fitting if DS; HF-r²SCAN-DC4 makes genuinely accurate predictions for most clusters.
  • Water dimers: HF-r²SCAN-DC4 essentially matches HF-SCAN’s near-chemical-accuracy (MAEs ≈ 0.08–0.11 kcal/mol depending on set), while HF-r²SCAN alone underbinds and SC-r²SCAN-D4 overbinds. Improvement over SC-r²SCAN-D4 diminishes with increasing O–O distance as sensitivity decreases.
• Noncovalent interactions (NCIs) and biomolecular relevance:
  • Stacked cytosine dimers: HF-SCAN underbinds by ~2–3 kcal/mol; HF-r²SCAN-DC4 reduces errors substantially (MAE ~0.4 kcal/mol), though B3LYP-D3(BJ) remains slightly better (<0.2 kcal/mol) but is poorer for water due to density-driven errors.
  • Water–aspirin interactions (MD structures): HF-r²SCAN-DC4 errors are much smaller than HF-SCAN and also smaller than SC-r²SCAN-D4, highlighting the necessity of both dispersion and density correction near organic molecules.
  • Broader NCI datasets (GMTKN55 inter-/intra-molecular): HF-r²SCAN-DC4 is highly accurate and on average outperforms SC-r²SCAN-D4, whereas HF-SCAN lacks long-range dispersion and performs poorly here.
• Global benchmarks (Fig. 5): HF-r²SCAN-DC4 reduces both water-metric errors and GMTKN55 WTMAD-2 by roughly half compared to HF-SCAN, achieving a balance unmatched by other comparably efficient functionals. While ωB97M-V can be even more accurate, it is substantially more expensive and less practical for DFT-MD.
• Numerical robustness: r²SCAN resolves SCAN’s grid convergence issues; HF-r²SCAN-DC4 can be used routinely and efficiently.
• DC-aware fitting matters: A prior HF-r²SCAN-D4 parameterization (SM21) that ignored DC-DFT principles performed noticeably worse on water clusters and related metrics than HF-r²SCAN-DC4, despite similar WTMAD-2.

The work addresses the key question of how to extend near-chemical accuracy beyond pure water within a practical DFT framework. By combining HF densities (to mitigate density-driven errors), r²SCAN (to maintain SCAN-level accuracy while avoiding numerical grid issues), and a dispersion correction fit exclusively on density-insensitive problems, HF-r²SCAN-DC4 achieves high accuracy for pure water while adding robust, long-range dispersion needed for NCIs. Analyses across water dimers, hexamers, 20-mers, and WATER27 show consistent improvements over HF-SCAN where density-driven errors are significant and maintain competitive accuracy in low-sensitivity regimes. The method substantially improves interactions of water with organic/biological molecules and general NCI benchmarks, demonstrating relevance for biomolecular and solution-phase simulations. The many-body expansion indicates that improvements are systematic and not merely due to error cancellations among K-body terms. A key insight is that fitting empirical parameters on DI reactions avoids contaminating the fit with density-driven errors, preserving accuracy for water and enabling transferability to broader chemistries.

The authors present HF-r²SCAN-DC4, a density-corrected, dispersion-inclusive meta-GGA framework that: (i) matches or surpasses HF-SCAN for pure water (improving hexamer and 20-mer energetics and correcting isomer ordering), (ii) accurately captures noncovalent interactions, including dispersion-dominated cases (e.g., stacked nucleobases, water–organic interactions), and (iii) eliminates SCAN’s grid convergence issues via r²SCAN. The DC-DFT-informed parameterization of the D4 dispersion term is crucial to these outcomes. The approach offers a practical, efficient alternative to high-level methods for generating accurate reference energetics, suitable for simulations of solutions and moderately large biomolecular systems, and for training ML potentials. Future work could include extensive DFT-MD simulations in complex solutions and biomolecular environments, exploration of adaptive DC(HF)-DFT switching strategies where beneficial, and broader validation across ionic and heterogeneous aqueous systems.

• Density correction is not universally beneficial: in density-insensitive cases (e.g., relative isomer energies of some water 20-mers), self-consistent r²SCAN-D4 can outperform HF-r²SCAN-DC4, though the latter still significantly outperforms HF-SCAN.
• The method evaluates functionals on HF densities universally (HF-DFT) rather than using a sensitivity-based DC(HF)-DFT switch, potentially sacrificing small gains in DI regimes for simplicity.
• While performance on GMTKN55 improves markedly over HF-SCAN, it is not as low-error as more expensive functionals like ωB97M-V.
• Reliance on HF densities and empirical dispersion parameters introduces additional computational steps and parameter dependencies; spin contamination considerations apply to HF-DFT in general (handled in DC(HF)-DFT but not explicitly toggled here).