[Paper Review] Grassmann extrapolation of density matrices for Born-Oppenheimer molecular dynamics

Étienne Polack, Geneviève Dusson+2 more|arXiv (Cornell University)|Jul 28, 2021

Machine Learning in Materials ScienceMaterials Science42 references22 citations

TL;DR

This paper introduces Grassmann extrapolation (g-Ext), a differential geometry-based method that enables linear extrapolation of density matrices in Born-Oppenheimer molecular dynamics by mapping them to a tangent space, ensuring idempotency while preserving physical consistency. The method significantly reduces SCF iterations—especially under tight convergence thresholds—outperforming the established XLBO approach in accuracy and efficiency for DFT and post-HF simulations.

ABSTRACT

Born-Oppenheimer Molecular Dynamics (BOMD) is a powerful but expensive technique. The main bottleneck in a density functional theory bomd calculation is the solution to the Kohn-Sham (KS) equations, that requires an iterative procedure that starts from a guess for the density matrix. Converged densities from previous points in the trajectory can be used to extrapolate a new guess, however, the non-linear constraint that an idempotent density needs to satisfy make the direct use of standard linear extrapolation techniques not possible. In this contribution, we introduce a locally bijective map between the manifold where the density is defined and its tangent space, so that linear extrapolation can be performed in a vector space while, at the same time, retaining the correct physical properties of the extrapolated density using molecular descriptors. We apply the method to real-life, multiscale polarizable QM/MM.

Motivation & Objective

• To address the high computational cost of SCF iterations in Born-Oppenheimer molecular dynamics (BOMD), especially under tight convergence criteria.
• To overcome the non-linearity of density matrix idempotency, which prevents direct use of standard linear extrapolation techniques.
• To develop a geometric framework that enables linear extrapolation in a vector space while preserving the physical constraints of the density matrix manifold.
• To improve convergence speed and stability in multiscale polarizable QM/MM BOMD simulations, particularly for excited-state and post-HF calculations.
• To provide a time-reversible alternative to existing extrapolation schemes, with potential for integration into machine learning-enhanced descriptors.

Proposed method

• The method uses a locally bijective map from the manifold of idempotent density matrices to its tangent space, enabling linear extrapolation in a vector space.
• A molecular descriptor—specifically the Coulomb matrix—is used to represent molecular geometry in a way that respects invariance under rotation and translation.
• Previous descriptors at earlier MD steps are used to fit a least-squares extrapolation of the descriptor coefficients in the tangent space.
• The extrapolated vector in the tangent space is then mapped back to the manifold of physical density matrices via the exponential map, ensuring idempotency.
• The resulting density matrix is used as an initial guess for the SCF procedure, reducing the number of iterations required for convergence.
• The approach is general and can be combined with any linear extrapolation technique and any choice of molecular descriptor, including advanced ML-based descriptors.

Experimental results

Research questions

• RQ1 Can a geometric framework be developed to enable linear extrapolation of density matrices while preserving their idempotent and physical constraints?
• RQ2 How does Grassmann extrapolation compare to the extended Lagrangian Born-Oppenheimer (XLBO) method in terms of convergence speed and energy conservation under tight SCF thresholds?
• RQ3 Can the method be effectively applied to multiscale polarizable QM/MM simulations, particularly for excited-state potential energy surfaces?
• RQ4 To what extent does the choice of molecular descriptor influence the accuracy and robustness of the extrapolation scheme?
• RQ5 Is it possible to extend the g-Ext method to geometry optimization, where XLBO is not applicable?

Key findings

• The g-Ext method outperforms the XLBO method in reducing SCF iterations, especially under tight convergence thresholds (e.g., 10⁻⁷), where the XLBO method shows diminished performance.
• For a 3-HF molecule, g-Ext achieved convergence with fewer than 4 SCF iterations at a 10⁻⁶ threshold, demonstrating high efficiency.
• The method maintains good energy conservation over short to medium trajectories, though a small long-term energy drift was observed due to lack of time reversibility.
• The use of the Coulomb matrix as a molecular descriptor provided a stable and effective representation for extrapolation, though more advanced descriptors may further improve performance.
• The method is applicable beyond DFT, including Hartree-Fock and semiempirical methods, broadening its potential impact.
• A Julia implementation of the g-Ext algorithm is publicly available, facilitating adoption and further development.