Abstract
Variational approaches for the calculation of vibrational wave functions and energies are a natural route to obtain highly accurate results with controllable errors. Here, we demonstrate how the density matrix renormalization group (DMRG) can be exploited to optimize vibrational wave functions (vDMRG) expressed as matrix product states. We study the convergence of these calculations with respect to the size of the local basis of each mode, the number of renormalized block states, and the number of DMRG sweeps required. We demonstrate the high accuracy achieved by vDMRG for small molecules that were intensively studied in the literature. We then proceed to show that the complete fingerprint region of the sarcosyn-glycin dipeptide can be calculated with vDMRG.
Introduction
Predicting a molecule's "infrared fingerprint" requires solving the vibrational Schrödinger equation to find wave functions and energies. Traditional approaches face a significant trade-off. Variational methods like Vibrational Configuration Interaction (VCI) provide high accuracy but suffer from exponential scaling, limiting their use to small molecules (typically under 20 atoms). Conversely, perturbative methods like Second-Order Vibrational Perturbation Theory (VPT2) scale better but often fail for "floppy" molecules with strong anharmonicity or shallow potential energy surfaces.
This paper introduces vDMRG, a framework that bridges this gap. By adapting the Density Matrix Renormalization Group (DMRG) algorithm—originally developed for condensed matter physics—to molecular vibrations, the authors provide a variational method that scales polynomially, enabling high-accuracy anharmonic calculations for medium-to-large molecular systems.
Methodology: Vibrational DMRG
The core of vDMRG lies in representing the vibrational wave function as a Matrix Product State (MPS). In this representation, the wave function is decomposed into a chain of tensors, where each site corresponds to a vibrational mode. The vibrational Hamiltonian (Watson Hamiltonian) is expressed in a second-quantized bosonic form and encoded as a Matrix Product Operator (MPO).
The optimization follows a Sweep Algorithm. Instead of a global diagonalization, vDMRG optimizes the wave function site-by-site. This iterative process focuses computational resources on the most physically relevant "entangled" configurations. Accuracy is controlled by two primary parameters:
- $N_{max}$: The size of the local basis for each vibrational mode.
- $m$ (Bond Dimension): The number of renormalized block states kept during the sweep, which determines the amount of correlation captured.
Results and Discussion
The vDMRG method was validated across several benchmarks, including $ClO_2$, $CH_3CN$, and $C_2H_4$. For acetonitrile ($CH_3CN$), a standard benchmark for quartic force fields, vDMRG demonstrated exceptional precision.
As shown in the table above, vDMRG results for $CH_3CN$ converge rapidly toward reference values as the bond dimension $m$ and local basis size $N_{max}$ increase. Most states converged to within $0.2 \text{ cm}^{-1}$ of the reference, significantly outperforming the harmonic approximation.
The method's scalability was further tested on the $SarGly^+$ dipeptide (35 modes). vDMRG successfully reproduced the experimental infrared spectrum in the complex "fingerprint region" (900–1700 $\text{cm}^{-1}$), a task typically unfeasible for standard variational methods. The calculations captured essential red-shifts and band splittings that the harmonic model missed, completing in approximately 4.8 hours per state.
Conclusion and Future Directions
vDMRG represents a significant step forward in vibrational spectroscopy, offering a variational approach that avoids exponential scaling. While highly effective, the authors note several areas for optimization:
- Mode Ordering: The efficiency of the MPS representation depends on the 1D ordering of modes; using quantum entanglement measures could optimize this.
- Basis Optimization: Moving from a harmonic oscillator basis to a Vibrational Self-Consistent Field (VSCF) basis might reduce the required $N_{max}$.
- Hybrid Approaches: Future work may combine vDMRG for strongly coupled modes with perturbation theory for the remainder, further extending the method's reach to even larger biomolecular systems.
Summary updated