[论文解读] Initial Guesses for Multicomponent Mean-Field Methods: Assessment and New Developments

Denis G. Artiukhin|arXiv (Cornell University)|Feb 11, 2026

Advanced Chemical Physics Studies被引用 0

一句话总结

The paper introduces new protonic initial guesses for multicomponent mean-field (NEO) calculations, notably HO-based approaches, and benchmarks them against existing guesses for electron-proton systems. It also analyzes computational trade-offs and implementation aspects.

ABSTRACT

The convergence of self-consistent field equations in mean-field nuclear-electronic orbital methods strongly depends on the choice of initial guesses for quantum nuclei. Although several such guesses have been proposed in the literature, a systematic comparison of their performance as well as attempts of constructing novel approximations based on model tasks of quantum mechanics were not reported to date. In this work, we address both issues by introducing novel nuclear initial guesses derived from the analytical solutions of the three-dimensional quantum harmonic oscillator and benchmarking them against existing approaches. We demonstrate that the isotropic variant of our guess outperforms existing approximations in nuclear-electronic orbital density functional theory calculations employing a simultaneous self-consistent field convergence protocol. Although our guess requires the computation of partial Hessians, we demonstrate that these can be evaluated with low-cost methods without affecting the accuracy of resulting protonic density matrices. Our results demonstrate that the proposed guess is robust and efficient and could provide a route to improved convergence in mean-field nuclear-electronic orbital computations.

研究动机与目标

• Develop and analyze new protonic initial guesses for multicomponent mean-field (NEO) methods.
• Benchmark the new HO-based guesses against existing core, 1s, SAD, and SND approaches.
• Evaluate convergence performance and robustness across single and multiple quantum protons.
• Assess practical computational costs and projection techniques for integrating HO-based guesses into proton basis sets.

提出的方法

• Construct two initial guesses from the 3D harmonic oscillator (HO) analytically for the quantum proton: HOa (anisotropic) and HOi (isotropic) variants.
• Compute partial Hessians and diagonalize to obtain HO frequencies along nuclear displacements.
• Project HO wave functions onto chosen proton basis sets using a basis-set projection operator to form initial proton density matrices.
• Derive and implement analytical overlap expressions between anisotropic HO functions and isotropic Gaussian AOs to enable HOa projection.
• Compare HOa, HOi, 1s-type, core, and SAD/SND-like guesses within NEO-HF and NEO-DFT using stepwise and simultaneous SCF convergence protocols.
• Explore lower-cost alternatives for Hessian computation and discuss practical implementation considerations.

Figure 1: $f$ -values computed with NEO-HF and NEO-DFT using the Cartesian PB4-F2 basis set and four different initial guesses for molecules containing a single quantum proton. Bars correspond to $f$ -values averaged over all molecules in the set containing a single quantum proton, whereas range bar

实验结果

研究问题

• RQ1 Do HO-based protonic initial guesses improve SCF convergence and accuracy for NEO-HF/NEO-DFT relative to existing guesses?
• RQ2 How do anisotropic (HOa) and isotropic (HOi) HO-derived guesses perform across different molecules and basis sets in terms of f-rank scores and SCF iterations?
• RQ3 What are the computational costs and feasibility of HO-based guesses, considering partial Hessians and projection steps?
• RQ4 Can HO-based guesses be effectively generalized to multiple quantum protons and heavier nuclei beyond protons?

主要发现

• HOi (isotropic HO) typically yields the best performance in NEO-DFT across most test cases.
• HOa (anisotropic HO) can be very effective for certain molecules but is more complex and shows molecule-dependent results.
• For single-quantum-proton systems, HOi often outperforms 1s and core guesses in NEO-DFT, while in NEO-HF the core/1s guesses can perform better due to proton density localization.
• The core guess generally provides the smallest number of SCF iterations in NEO-HF under simultaneous convergence, whereas HOi tends to require fewer iterations in many NEO-DFT cases.
• A grid search indicated that for NEO-DFT with PB4-F2, using zeta ≈ 10 for the 1s-like initial guess yields high f-rank scores (~0.97).
• HOa requires partial Hessians, which can be costly, but lower-cost Hessians can be used without compromising results.
• HOi can be implemented without changing the code base beyond the projection step and provides a practical alternative to 1s or core guesses.

Figure 2: FHF - protonic density plotted along the F–H–F direction. Results generated with the NEO-HF approach are shown on the left, whereas those from NEO-DFT are on the right. Cartesian PB4-F2 basis set was employed in both cases.