betaFeedback

WorksPhysics and Astronomy

ViBra: Configuration Interaction for Anharmonic Vibrational Spectroscopy and Quantum-Sampled Configuration Spaces

Abstract

Quantum-centric workflows are a promising route to improving the accuracy of property predictions in computational chemistry and materials science. By integrating quantum sampling algorithms with classical solvers, electronic structure calculations have recently demonstrated their potential even on noisy intermediate-scale quantum devices. In principle, the method of Vibrational Configuration Interaction (VCI) is suitable for integration with quantum sampling algorithms as well. However, demonstrations of computational workflows for quantum-centric, vibrational property predictions are still lacking. Here, we introduce a methodology for performing anharmonic vibrational structure calculations that can be deployed in a hybrid, quantum-classical mode. Starting from a quartic force field, the approach combines a Vibrational Self-Consistent Field (VSCF) with VCI in either Full, Selected (S-VCI), or Symmetry-Adapted (SA-VCI) mode. In S-VCI, an Epstein-Nesbet perturbative screening significantly reduces the configuration space while retaining high predictive accuracy. A state-list input enables the integration of externally generated vibrational configurations as a seed space. As a proof-of-concept, we demonstrate a hybrid, quantum-classical computational workflow, in which a quantum sampling algorithm provides the seed. Our vibrational wave function analysis package ViBra, equipped with a graphical interface, is available at https://github.com/raphafe96/ViBra.

Introduction

Vibrational spectroscopy provides a direct window into molecular structure and dynamics. The positions and intensities of fundamental, overtone, and combination bands encode information about bond strengths, mode couplings, and the underlying potential energy surface (PES). Accurate theoretical prediction of these observables therefore constitutes a stringent test of electronic structure methods and is indispensable for the assignment of experimental spectra.

Within the Born–Oppenheimer approximation, the nuclear vibrational problem can be formulated as the solution of the vibrational Schrödinger equation in mass-weighted normal coordinates Q=(Q1,…,QM)\mathbf{Q} = (Q_1,\ldots,Q_M), where M=3Nat−6M = 3N_{\mathrm{at}} - 6 or M=3Nat−5M = 3N_{\mathrm{at}} - 5 for linear molecules is the number of vibrational degrees of freedom. The simplest harmonic approach approximation uses a quadratic PES and diagonalizes the mass-weighted Hessian to yield independent normal modes. However, this approximation fails to reproduce experimental fundamental frequencies, cannot describe overtones and combination bands, and gives vanishing intensities for these transitions within the double-harmonic (mechanical and electrical) approximation. [1] The accurate treatment of mechanical anharmonicity requires the inclusion of at least cubic and quartic terms in the PES expansion. [1, 2]

Two broad classes of methods have been developed for the anharmonic vibrational problem. Perturbative approaches, exemplified by second-order vibrational perturbation theory (VPT2), [1, 3] are computationally efficient and have been widely deployed in quantum chemistry codes such as Gaussian [1] and ORCA. [4, 5] However, VPT2 breaks down in the presence of Fermi and Darling–Dennison resonances, requiring ad hoc treatments that can compromise accuracy.

Variational methods, by contrast, are systematically improvable and are robust in the presence of resonances. The Vibrational Self-Consistent Field (VSCF) approach, introduced independently by Bowman [6] and by Gerber and Ratner, [7] provides a mean-field description in which each normal mode evolves in the average potential of all other modes. To recover the correlation missed by the mean-field treatment, VSCF is typically followed by Vibrational Configuration Interaction (VCI), where the wavefunction is expanded in a basis of Hartree products of VSCF-optimised one-mode functions (modals) and the Hamiltonian is diagonalised in this space.[2, 8, 9]

Over the past three decades, several powerful VCI implementations have been developed. The MULTIMODE code, developed by Bowman and co-workers,[9] uses an nn-mode representation of the potential in normal coordinates with a hierarchical coupling scheme and has been applied to molecules with up to dozens of atoms. The MOLPRO package features state-specific configuration-selective VCI (cs-VCI) algorithms pioneered by Rauhut and co-workers,[10, 11] which use perturbative estimates to efficiently screen the correlation space. More recently, Mathea and collaborators have introduced advanced screening techniques and efficient handling of rovibrational angular momentum terms that substantially accelerate these calculations.[12, 13]

The CRYSTAL program, developed by Erba, Maul, Dovesi, and co-workers, has extended VSCF and VCI methods to periodic systems (solids), implementing a fully automated scheme for constructing the anharmonic PES from density functional theory (DFT) calculations and computing anharmonic vibrational states of solids.[14, 15] A subsequent extension by Carbonnière et al. additionally enabled the calculation of anharmonic infrared and Raman intensities for periodic systems.[16]

In the open-source domain, the PyVCI package by Sibaev and Crittenden[17] provides a flexible Python-based VCI implementation with computationally intensive routines written in C via Cython. The DVCI (Dual Vibration Configuration Interaction) method by Garnier et al.[18] employs an efficient factorisation of the Hamiltonian to target specific spectral transitions with minimal memory usage, while the A-VCI (Adaptive VCI) algorithm[19, 20] iteratively constructs nested bases guided by a posteriori error estimates.

A significant recent advance has been the development of Vibrational Heat-Bath Configuration Interaction (VHCI). Introduced independently by Fetherolf and Berkelbach in 2021[21] and by Bhatty and Brorsen in the same year,[22] VHCI adapts the heat-bath CI algorithm from electronic structure theory to the vibrational domain. It uses a simple, pre-sorted list of anharmonic force constants to efficiently identify important basis states, achieving near-full-VCI accuracy at a fraction of the computational cost. A subsequent extension incorporated VSCF modals as the one-mode basis and semistochastic perturbation theory, further improving convergence.[23]

Despite this wealth of methodology, several challenges remain: the exponential scaling of the VCI basis with MM, the accurate treatment of transition intensities beyond the linear dipole surface, and the need for user-friendly software that bridges computation and visualisation. We therefore designed ViBra to address these challenges through a combination of:

  • Selected CI (S-VCI): an Epstein–Nesbet perturbation-theory-driven configuration selection that significantly reduces the VCI space while preserving accuracy for low-lying states (§ 2.5);

  • Symmetry-Adapted VCI: exact block-diagonalisation of the Hamiltonian for molecules in Abelian point groups with one-dimensional irreducible representations (C1C_1, CsC_s, CiC_i, C2C_2, C2hC_{2h}, C2vC_{2v}, D2D_2, D2hD_{2h}), providing irrep labels and reducing diagonalisation cost by up to a factor of nΓ2n_{\Gamma}^{2} (§ 2.6), where nΓn_{\Gamma} is the number of symmetry elements of a given Abelian point group;

  • Treatment of electrical anharmonicity: all VCI-level transition dipoles include both first- and second-order dipole moment derivatives, essential for overtone and combination band intensities and for the correct redistribution of oscillator strength through Fermi resonances (§ 2.7);

  • Graphical interface and interactive viewer: a Python/CustomTkinter GUI for input preparation and job execution, and a Matplotlib-based viewer with temperature-dependent spectra, experimental overlay, baseline correction, and 3D normal-mode animation via 3Dmol.js (§ 3.1).

A further motivation for the development of ViBra is the growing interest in quantum algorithms for vibrational structure and spectroscopy. In quantum chemistry and materials science, recent quantum-centric workflows such as sample-based quantum diagonalization (SQD) and sample-based Krylov quantum diagonalization (SKQD) have highlighted a practical near-term paradigm in which quantum processors are used as sampling devices, while compact Hamiltonian construction, diagonalization, and observable evaluation are performed classically. [24, 25]

More broadly, quantum-centric supercomputing has been proposed as a route for integrating quantum processors with high-performance classical resources in materials-science workflows. [26]

In contrast to electronic-structure applications, however, quantum algorithms for anharmonic vibrational dynamics remain comparatively underdeveloped, despite recent progress in quantum algorithms for vibronic spectra, vibrational structure calculations, and Trotterized simulation of vibrational Hamiltonians. [27, 28, 29]

The selected-VCI infrastructure implemented in ViBra is well aligned with this sampling-based perspective because it can serve as the classical VSCI engine after a quantum-generated seed space has been obtained. A quantum sampling algorithm proposes a compact set of vibrational configurations, while ViBra can either construct and diagonalize the Hamiltonian directly in that sampled space or enlarge it through EN-PT2-guided selection before variational rediagonalization. In both cases, ViBra evaluates transition observables and generates anharmonic vibrational spectra from the final configuration space. Thus, ViBra provides both the projected-VCI post-processing layer and an optional classical refinement stage for quantum-seeded vibrational configuration spaces.

Theoretical Framework

The potential energy surface (PES) governing nuclear motion can be represented in various forms, including numerical grids or analytical expansions. In ViBra, we have opted for a Taylor expansion of the PES in terms of the normal coordinates, truncated at fourth order. [15] This quartic force field representation provides a compact, analytically tractable form that enables efficient evaluation of all required matrix elements in closed form, since low-order polynomial terms in QiQ_i have simple analytic matrix elements in the harmonic-oscillator basis. Furthermore, it allows for a clean separation of the potential into harmonic, diagonal anharmonic, and coupling contributions, which is essential for the VSCF and VCI formulations described below. Grid-based representations such as the discrete variable representation (DVR), [30] while more flexible and not restricted to a low-order polynomial form, require the PES to be evaluated on a dense multidimensional grid whose size grows exponentially with the number of modes, and they do not offer a natural analytic separation into diagonal and coupling terms; this makes them less suitable for the sparsity-driven algorithms (§ 4) that ViBrA relies on.

Vibrational Hamiltonian and Quartic Force Field

For a non-linear molecule with NatN_{\mathrm{at}} atoms, mass-weighted normal coordinates Q=(Q1,…,QM)\mathbf{Q} = (Q_1, \ldots, Q_M) are obtained by diagonalizing the mass-weighted Cartesian Hessian, where M=3Nat−6M = 3N_{\mathrm{at}} - 6 is the number of vibrational degrees of freedom. Neglecting rotation–vibration coupling (J=0J = 0) and the Watson correction terms that are small for semi-rigid molecules, the vibrational Hamiltonian in atomic units reads

H^vib=−12∑i=1M∂2∂Qi2+V(Q).(1)\hat{H}_{\mathrm{vib}} = -\frac{1}{2}\sum_{i=1}^{M}\frac{\partial^2}{\partial Q_i^2} + V(\mathbf{Q}). \tag*{(1)}

The potential energy surface V(Q)V(\mathbf{Q}) is expanded in a Taylor series about the equilibrium geometry, here up to fourth order:

V(Q)=12∑i=1Mωi2Qi2+16∑i=1M∑j=1M∑k=1MϕijkQiQjQk+124∑i=1M∑j=1M∑k=1M∑l=1MϕijklQiQjQkQl.(2)V(\mathbf{Q}) = \frac{1}{2}\sum_{i=1}^{M}\omega_i^2 Q_i^2 + \frac{1}{6}\sum_{i=1}^{M}\sum_{j=1}^{M}\sum_{k=1}^{M}\phi_{ijk} Q_iQ_jQ_k + \frac{1}{24}\sum_{i=1}^{M}\sum_{j=1}^{M}\sum_{k=1}^{M}\sum_{l=1}^{M}\phi_{ijkl} Q_iQ_jQ_kQ_l. \tag*{(2)}

The harmonic frequencies ωi\omega_i and the cubic and quartic force constants ϕijk\phi_{ijk} and ϕijkl\phi_{ijkl} are provided exclusively in the ORCA VPT2 output format (.vpt2).[4, 5] This format was chosen because it provides all necessary information, including harmonic frequencies, cubic and quartic force constants with their permutational degeneracy factors, normal-coordinate transformation matrices, and first and second dipole derivatives, in a self-contained, well-documented text structure. Notably, ORCA is an free quantum chemistry package, making the entire computational workflow from force constant evaluation to anharmonic vibrational analysis fully freely accessible. Users of other quantum chemistry packages can convert their outputs to the equivalent format using as reference the ORCA output files provided in Support Information.

The tensors are totally symmetric under index permutation; only the upper-triangular components are stored, and each distinct term carries a permutational degeneracy factor. For cubic terms with i≤j≤ki \le j \le k, the weight gg is 1 (all indices equal), 3 (two equal), or 6 (all distinct). For quartic terms i≤j≤k≤li \le j \le k \le l, the weight is 1 (all equal), 4 (three equal), 6 (two doubles), 12 (one double, two singles), or 24 (all distinct). The code stores the effective potential in terms of these weighted coefficients as

V(Q)=∑i(12ωi2Qi2+ϕ~iiiQi3+ϕ~iiiiQi4)+∑i<j≤kϕ~ijkQiQjQk+∑i<j≤k≤lϕ~ijklQiQjQkQl.(3)V(\mathbf{Q}) = \sum_i \left(\frac{1}{2}\omega_i^2 Q_i^2 + \tilde{\phi}_{iii} Q_i^3 + \tilde{\phi}_{iiii} Q_i^4\right) + \sum_{i<j\le k} \tilde{\phi}_{ijk} Q_i Q_j Q_k + \sum_{i<j\le k\le l} \tilde{\phi}_{ijkl} Q_i Q_j Q_k Q_l. \tag*{(3)}

The coefficients ϕ~⋯\tilde{\phi}_{\cdots} incorporate the original force constants and the appropriate degeneracy factors. Internally, a separate diagonal potential Vidiag(Qi)=12ωi2Qi2+ϕ~iiiQi3+ϕ~iiiiQi4V_i^{\mathrm{diag}}(Q_i) = \frac{1}{2}\omega_i^2 Q_i^2 + \tilde{\phi}_{iii}Q_i^3 + \tilde{\phi}_{iiii}Q_i^4 is identified for each mode, and the coupling potential VcV_c groups all terms involving at least two distinct mode indices.

One-Mode Basis: Harmonic Oscillator

The primitive one-mode basis consists of harmonic oscillator eigenfunctions ϕn(Qi)\phi_n(Q_i) with frequency ωi\omega_i. Their matrix elements

mp(μ,ν)≡⟨ϕμ∣Qip∣ϕν⟩(4)m_p^{(\mu,\nu)} \equiv\langle\phi_\mu| Q_i^p | \phi_\nu\rangle \tag*{(4)}

for p=0,…,4p = 0,\ldots,4 are evaluated analytically using ladder-operator algebra. The kinetic energy matrix elements are computed from the second-derivative operator:

⟨ϕμ∣T^i∣ϕν⟩=−12⟨ϕμ|∂2∂Qi2|ϕν⟩.(5)\langle\phi_\mu| \hat{T}_i | \phi_\nu\rangle= -\frac{1}{2}\left\langle\phi_\mu\middle| \frac{\partial^2}{\partial Q_i^2} \middle| \phi_\nu\right\rangle. \tag*{(5)}

A fundamental selection rule governs these integrals: ⟨ϕμ∣Qip∣ϕν⟩=0\langle\phi_\mu| Q_i^p | \phi_\nu\rangle= 0 whenever ∣μ−ν∣>p|\mu-\nu| > p. All required matrix elements up to p=4p = 4 are pre-computed by the compute_integrals module and stored in memory for efficient look-up.

VSCF Method and Modal Basis

The VSCF wavefunction is the Hartree product

Ψn(Q)=∏i=1Mϕni(i)(Qi),(6)\Psi_{\mathbf{n}}(\mathbf{Q}) = \prod_{i=1}^{M} \phi_{n_i}^{(i)}(Q_i), \tag*{(6)}

where n=(n1,…,nM)\mathbf{n} = (n_1,\ldots,n_M) specifies the quantum state of each mode. Each one-mode function (modal) is expanded in the HO basis:

ϕni(i)(Qi)=∑μ=1NexpCμ,ni(i)ϕμ−1(Qi).(7)\phi_{n_i}^{(i)}(Q_i) = \sum_{\mu=1}^{N_{\mathrm{exp}}} C_{\mu,n_i}^{(i)} \phi_{\mu-1}(Q_i). \tag*{(7)}

Here NexpN_{\mathrm{exp}} is the number of HO basis functions per mode (user-supplied keyword NEXPAN; typical values range from 8 to 16).

Requiring the energy functional En=⟨Ψn∣H^vib∣Ψn⟩E_{\mathbf{n}} = \langle\Psi_{\mathbf{n} | \hat{H}_{\mathrm{vib}} | \Psi_{\mathbf{n}} \rangle} to be stationary with respect to variations of each modal under the constraint ⟨ϕni(i)∣ϕni(i)⟩=1\langle\phi_{n_i}^{(i)} | \phi_{n_i}^{(i)} \rangle= 1 yields MM coupled effective one-mode equations:

h^ieffϕni(i)(Qi)=εni(i)ϕni(i)(Qi),(8)\hat{h}_i^{\mathrm{eff}} \phi_{n_i}^{(i)}(Q_i) = \varepsilon_{n_i}^{(i)} \phi_{n_i}^{(i)}(Q_i), \tag*{(8)}

where εni(i)\varepsilon_{n_i}^{(i)} acts as the Lagrange multiplier enforcing normalisation and has the physical interpretation of a modal energy. The effective Hamiltonian

h^ieff=T^i+Vidiag(Qi)+⟨Vc⟩j≠i(9)\hat{h}^{\mathrm{eff}}_i = \hat{T}_i + V_i^{\mathrm{diag}}(Q_i) + \langle V_c\rangle_{j\ne i} \tag*{(9)}

contains the kinetic energy of mode ii, its diagonal potential, and the mean-field average of the coupling potential over all other modes:

⟨Vc⟩j≠i=⟨∏j≠iφnj(j)|Vc(Q)|∏j≠iφnj(j)⟩.(10)\langle V_c\rangle_{j\ne i} = \left\langle \prod_{j\ne i}\varphi_{n_j}^{(j)} \middle| V_c(\mathbf{Q}) \middle| \prod_{j\ne i}\varphi_{n_j}^{(j)} \right\rangle. \tag*{(10)}

Because VcV_c is a polynomial of degree four in the normal coordinates, the mean-field average is itself a polynomial in QiQ_i, so that

h^ieff=T^i+∑p=04Xi(p)Qip.(11)\hat{h}^{\mathrm{eff}}_i = \hat{T}_i + \sum_{p=0}^{4} X_i^{(p)} Q_i^p. \tag*{(11)}

The coefficients Xi(p)X_i^{(p)} are real numbers that depend on the force constants and on the current expectation values ⟨φnj(j)∣Qjp∣φnj(j)⟩\langle\varphi_{n_j}^{(j)}|Q_j^p|\varphi_{n_j}^{(j)}\rangle of all modes j≠ij\ne i. They are updated at each self-consistent iteration.

Substituting the modal expansion (7) into (8) and projecting onto ⟨ϕν−1∣\langle\phi_{\nu-1}| yields an Nexp×NexpN_{\mathrm{exp}}\times N_{\mathrm{exp}} real symmetric matrix eigenvalue problem:

∑μ=1NexpHνμ(i)Cμ,ni(i)=εni(i)Cν,ni(i),(12)\sum_{\mu=1}^{N_{\mathrm{exp}}} H_{\nu\mu}^{(i)} C_{\mu,n_i}^{(i)} = \varepsilon_{n_i}^{(i)} C_{\nu,n_i}^{(i)}, \tag*{(12)}

with matrix elements

Hνμ(i)=⟨ϕν−1∣T^i∣ϕμ−1⟩+∑p=04Xi(p)⟨ϕν−1∣Qip∣ϕμ−1⟩.(13)H_{\nu\mu}^{(i)} = \langle\phi_{\nu-1}|\hat{T}_i|\phi_{\mu-1}\rangle +\sum_{p=0}^{4} X_i^{(p)} \langle\phi_{\nu-1}|Q_i^p|\phi_{\mu-1}\rangle. \tag*{(13)}

The nin_i-th eigenvector gives the updated coefficients Cμ,ni(i)C_{\mu,n_i}^{(i)}. The VSCF equations are solved iteratively (Algorithm 1) until the energy change between successive iterations falls below 10−CVGSCF10^{-\mathrm{CVGSCF}} cm−1^{-1}. The ground-state modals are stored for use as the VCI one-mode basis.

Because each coupling term in VcV_c is counted in the effective Hamiltonian of every mode (either as an operator or as a constant shift Xi(0)X_i^{(0)}), the sum of modal energies ∑iEni(i)\sum_i \mathcal{E}_{n_i}^{(i)} overcounts the coupling energy. The correct VSCF total energy is therefore

EVSCF=∑i=1MEni(i)−(M−1)V‾c,(14)E_{\mathrm{VSCF}} = \sum_{i=1}^{M} \mathcal{E}_{n_i}^{(i)} - (M-1)\overline{V}_c, \tag*{(14)}

where the correlation potential

V‾c=⟨∏i=1Mφni(i)|Vc(Q)|∏i=1Mφni(i)⟩(15)\overline{V}_c = \left\langle\prod_{i=1}^{M}\varphi_{n_i}^{(i)} \middle| V_c(\mathbf{Q}) \middle| \prod_{i=1}^{M}\varphi_{n_i}^{(i)} \right\rangle \tag*{(15)}

is the fully averaged coupling potential. This correction ensures that each coupling term appears exactly once in the total energy.

Algorithm 1 VSCF Self-Consistent Field Iteration

  1. Initialize modal coefficients Cμ,ni(i)C_{\mu,n_i}^{(i)} to small uniform values for all modes

  2. repeat

  3. Precompute all one-mode expectation values ⟨φnj(j)∣Qjc∣φnj(j)⟩\langle\varphi_{n_j}^{(j)}|Q_j^c|\varphi_{n_j}^{(j)}\rangle for c=0,…,4c = 0,\ldots,4 (overlap cache)

  4. for i=1i = 1 to MM do

  5. Construct Xi(p)X_i^{(p)} (p=0,…,4p = 0,\ldots,4) from force constants and current overlaps

  6. Build Hνμ(i)H_{\nu\mu}^{(i)} via Eq. (13)

  7. Diagonalize H(i)\mathbf{H}^{(i)} (LAPACK DSYEVD); retain Cμ,ni(i)C_{\mu,n_i}^{(i)} from the nin_i-th eigenvector

  8. end for

  9. Compute V‾c\overline{V}_c as the fully contracted coupling potential

  10. EVSCF←∑iEni(i)−(M−1)V‾cE_{\mathrm{VSCF}} \leftarrow\sum_i \mathcal{E}_{n_i}^{(i)} - (M-1)\overline{V}_c

  11. until ∣EVSCF−Eold∣<10−CVGSCF cm−1|E_{\mathrm{VSCF}} - E_{\mathrm{old}}| < 10^{-\mathrm{CVGSCF}}\ \mathrm{cm}^{-1}

VCI Expansion and Precomputed Modal Integrals

The VCI expansion can be built using either the primitive harmonic oscillator basis or the ground-state VSCF modals. The VSCF modals yield faster variational convergence of the VCI energies with respect to increasing the maximum total quanta, as they already account for the average effect of anharmonic couplings. This is achieved without additional computational overhead, since all required modal integrals are precomputed and stored (see § 4). The VSCF-based VCI approach is therefore recommended for routine anharmonic vibrational calculations and thus employed in ViBrA. As a result, the wavefunction is a linear combination of Hartree products built from the ground-state VSCF modals:

∣ΨI⟩=∑n∈Scn(I)∣Φn⟩,∣Φn⟩=∏i=1M∣φni(i)⟩.(16)|\Psi_I\rangle= \sum_{\mathbf{n}\in\mathcal{S}} c_{\mathbf{n}}^{(I)}|\Phi_{\mathbf{n}}\rangle,\qquad|\Phi_{\mathbf{n}}\rangle= \prod_{i=1}^{M}|\varphi_{n_i}^{(i)}\rangle. \tag*{(16)}

The configuration space S\mathcal{S} is truncated by a maximum total quantum number NqN_q (keyword NQUANT) such that ∑i=1Mni≤Nq\sum_{i=1}^{M} n_i \le N_q. Using a single fixed set of ground-state modals ensures orthonormality of the configuration basis. A critical component enabling efficient VCI is the precomputed modal integral table:

Ii(v,v′;p)≡⟨φv(i)∣Qip∣φv′(i)⟩=∑μ,ν=1NexpCμ,v(i)Cν,v′(i)⟨ϕμ−1∣Qip∣ϕν−1⟩,(17)\mathcal{I}_i(v,v';p) \equiv\langle\varphi_v^{(i)}|Q_i^p|\varphi_{v'}^{(i)}\rangle= \sum_{\mu,\nu=1}^{N_{\mathrm{exp}}} C_{\mu,v}^{(i)} C_{\nu,v'}^{(i)}\langle\phi_{\mu-1}|Q_i^p|\phi_{\nu-1}\rangle, \tag*{(17)}

with p=0,…,5p=0,\ldots,5 (where p=5p=5 stores kinetic energy matrix elements in the modal basis). The table is symmetric in its first two arguments and occupies M×(Nq+1)2×6M\times(N_q+1)^2\times6 double-precision entries—typically a few tens of megabytes. Once computed, every one-mode integral required for the VCI Hamiltonian construction becomes an O(1)O(1) look-up, eliminating the need for the O(Nexp2)O(N_{\mathrm{exp}}^2) double sum otherwise required for each evaluation.

The Hamiltonian matrix element Hmn≡⟨Φm∣H^vib∣Φn⟩H_{mn}\equiv\langle\Phi_m|\hat{H}_{\mathrm{vib}}|\Phi_n\rangle is assembled from three types of contribution:

One-body terms collect all diagonal anharmonic corrections for each mode ii:

H(1)=∑i=1M[Ii(mi,ni;5)+ϕ~iiiiIi(mi,ni;4)+ϕ~iiiIi(mi,ni;3)+ωi2Ii(mi,ni;2)]Sall∖i.(18)H^{(1)}=\sum_{i=1}^{M}\left[\mathcal{I}_i(m_i,n_i;5)+\tilde{\phi}_{iiii}\mathcal{I}_i(m_i,n_i;4)+\tilde{\phi}_{iii}\mathcal{I}_i(m_i,n_i;3)+\frac{\omega_i}{2}\mathcal{I}_i(m_i,n_i;2)\right]\mathcal{S}_{\mathrm{all}\setminus i}. \tag*{(18)}

Cubic coupling terms with unique modes {s1,s2,s3}\{s_1,s_2,s_3\} and multiplicities {c1,c2,c3}\{c_1,c_2,c_3\} contribute

H(3,term)=ϕ~term∏r=13Isr(msr,nsr;cr) Sall∖{s1,s2,s3}.(19)H^{(3,\mathrm{term})}=\tilde{\phi}_{\mathrm{term}}\prod_{r=1}^{3}\mathcal{I}_{s_r}(m_{s_r},n_{s_r};c_r)\,\mathcal{S}_{\mathrm{all}\setminus\{s_1,s_2,s_3\}}. \tag*{(19)}

Quartic coupling terms are treated analogously with up to four unique modes.

In the above, Sall\SS_{\mathrm{all}\backslash S} denotes the product of spectator-mode overlaps:

Sall\S=∏k∉SIk(mk,nk;0).(20)S_{\mathrm{all}\backslash S} = \prod_{k \notin S} I_k(m_k,n_k;0). \tag*{(20)}

A fundamental selection rule emerges: because Ik(mk,nk;0)=δmk,nkI_k(m_k,n_k;0)=\delta_{m_k,n_k} (modal orthonormality), a potential term can couple configurations m\mathbf{m} and n\mathbf{n} only if it involves all modes in which the two configurations differ. Consequently, two VCI configurations can be coupled only if they differ in at most four modes (the maximum number of distinct indices in a quartic term). This sparsity is exploited through a precomputed pair list and an inverted potential index.

The Hamiltonian is diagonalised using LAPACK routines: DSYEVD (divide-and-conquer) for full eigenvalue spectra, and DSYEVR (relatively robust representations) for partial spectra when only the lowest NstateN_{\mathrm{state}} eigenvalues are required.

Selected Vibrational CI (S-VCI)

Full VCI becomes prohibitive for large MM and NqN_q since ∣S∣∼(M+NqNq)|\mathcal{S}| \sim\binom{M+N_q}{N_q}. ViBra’s SCI algorithm addresses this by constructing a compact active space through Epstein–Nesbet perturbation theory[31, 32] screening (EN-PT).

Reference Space

The reference space Sref\mathcal{S}_{\mathrm{ref}} is set by the user through the MAXSCI keyword,

MAXSCI Nsel [auto | list] [nref]

where NselN_{\mathrm{sel}} is the number of external configurations retained per reference state (§ 2.5.3), the second argument selects how the reference space is built, and the third, optional, argument sets the truncation order nrefn_{\mathrm{ref}}, used only when the auto option is chosen.

In auto mode the reference space comprises all configurations whose total excitation does not exceed nrefn_{\mathrm{ref}},

Sref={n∈S:∑ini≤nref}.(21)\mathcal{S}_{\mathrm{ref}} = \left\{ \mathbf{n} \in\mathcal{S} : \sum_i n_i \le n_{\mathrm{ref}} \right\}. \tag*{(21)}

For nref=2n_{\mathrm{ref}} = 2 (ground state, one-mode singles ni=1n_i = 1, and doubles ni=2n_i = 2 or ni=nj=1n_i = n_j = 1 pairs) this has dimension 1+2M+(M2)1 + 2M + \binom{M}{2}, linear in MM for the singles and quadratic for the doubles; larger values of nrefn_{\mathrm{ref}} extend the reference to include triples, quadruples, and so on, at the cost of a larger reference Hamiltonian.

In list mode the reference space is instead supplied explicitly by the user through a plain-text file, list_states.txt, whose first line gives the number of reference configurations, followed by one configuration (a length-MM vector of mode quanta nin_i) per line. This allows the user to target configurations of specific chemical or spectroscopic interest, independent of any excitation-order truncation. In either case, the Hamiltonian in the resulting reference space Sref\mathcal{S}_{\mathrm{ref}} is built and diagonalised exactly, yielding reference energies EI(0)E_I^{(0)} and eigenvectors ∣ΨI(0)⟩=∑n∈Srefcn(I)∣Φn⟩|\Psi_I^{(0)}\rangle= \sum_{n \in\mathcal{S}_{\mathrm{ref}}} c_n^{(I)}|\Phi_n\rangle, for I=1,…,NstateI = 1, \ldots, N_{\mathrm{state}}, where NstateN_{\mathrm{state}} here is the number of vibrational reference states.

Epstein–Nesbet PT2 Screening

For each external configuration ∣Φα⟩∉Sref|\Phi_\alpha\rangle\notin\mathcal{S}_{\mathrm{ref}} and each reference state II to be converged, the Epstein–Nesbet second-order energy correction is computed:

eα(I)=∣∑n∈Srefcn(I)Hαn∣2EI(0)−Hαα.(22)e_\alpha^{(I)} = \frac{\left|\sum_{n \in\mathcal{S}_{\mathrm{ref}}} c_n^{(I)} H_{\alpha n}\right|^2}{E_I^{(0)} - H_{\alpha\alpha}} . \tag*{(22)}

The denominator HααH_{\alpha\alpha} is the diagonal Hamiltonian element of the external configuration itself, Hαα=⟨Φα∣H∣Φα⟩H_{\alpha\alpha} = \langle\Phi_\alpha|H|\Phi_\alpha\rangle. This is the defining feature of the Epstein–Nesbet partitioning of the Hamiltonian: rather than approximating a configuration’s zeroth-order energy as a sum of independent-mode (harmonic) energies, as would be done in a Møller–Plesset-type partitioning, the exact diagonal matrix element—already including diagonal anharmonic corrections—is used directly. This makes the energy denominator in Eq. (22) a tighter and more physically faithful estimate for each configuration, at the cost of no additional computation beyond evaluating HααH_{\alpha\alpha}. The numerator in Eq. (22) instead couples ∣Φα⟩|\Phi_\alpha\rangle to every reference configuration nn that differs by at most four quanta; this coupling is evaluated efficiently using the inverted potential index (Section 4).

Top-NN Selection and Active Space Construction

For each reference state II, the NselN_{\mathrm{sel}} external configurations with the largest ∣eα(I)∣\lvert e_{\alpha}^{(I)} \rvert are retained, where NselN_{\mathrm{sel}} is the first argument of the MAXSCI keyword. A sorted insertion procedure maintains the top-NN candidates for each state as the external space is scanned once, avoiding a full sort of all PT2 contributions. The final active space is the union

Sactive=Sref ∪ ⋃I=1NstateSsel(I).(23)\mathcal{S}_{\mathrm{active}} = \mathcal{S}_{\mathrm{ref}} \,\cup\, \bigcup_{I=1}^{N_{\mathrm{state}}} \mathcal{S}_{\mathrm{sel}}^{(I)}. \tag*{(23)}

Because a single external configuration may be selected by multiple reference states, duplicate insertions are automatically avoided when the union is formed. The full VCI Hamiltonian is then rebuilt in the final active space and diagonalised using the same optimised kernel used for the reference space and for the conventional full-VCI calculation.

For example, for the ethylene molecule, with 12 vibrational modes, a choice of 6 quanta, using Nsel=100N_{\mathrm{sel}} = 100 yields an active space corresponding to only 32% of the full VCI dimension, while the lowest eigenvalues agree with the full VCI results within 1–5 cm−1^{-1}. S-VCI is activated automatically when MAXSCI Nsel>0N_{\mathrm{sel}} > 0 and no molecular symmetry is specified (i.e., PGROUP is C1 or absent). If symmetry is active, the code defaults to the symmetry-adapted full VCI.

Symmetry-Adapted VCI (SA-VCI)

Motivation and Scope

The vibrational Hamiltonian H^vib\hat{H}_{\mathrm{vib}} is invariant under every symmetry operation of the molecular point group G\mathcal{G}: [H^vib,R^]=0[\hat{H}_{\mathrm{vib}}, \hat{R}] = 0 for all R^∈G\hat{R} \in\mathcal{G}. Consequently, its matrix representation in a symmetry-adapted basis is block-diagonal: matrix elements between configurations of different irreducible representations (irreps) are identically zero.

ViBra exploits this for the eight Abelian point groups with exclusively one-dimensional irreps: C1C_1, CsC_s, CiC_i, C2C_2, C2hC_{2h}, C2vC_{2v}, D2D_2, D2hD_{2h}. In these groups, all characters are ±1\pm1 and the irrep of a Hartree product factorises as a direct product of one-mode irreps.

Determination of Normal-Mode Irreps

The irrep of each normal mode kk is determined by projecting its Cartesian displacement vector L(k)\mathbf{L}^{(k)} (a 3Nat3N_{\mathrm{at}}-dimensional vector) onto the symmetry operations in the principal-axis frame (PAF). The character of mode kk under operation R^\hat{R} is

χ(k)(R^)=1∥L(k)∥2∑a=1Nat∑α,βLα,R^(a)(k)RαβLβ,a(k).(24)\chi^{(k)}(\hat{R}) = \frac{1}{\lVert\mathbf{L}^{(k)} \rVert^2} \sum_{a=1}^{N_{\mathrm{at}}} \sum_{\alpha,\beta} L_{\alpha,\hat{R}(a)}^{(k)} R_{\alpha\beta} L_{\beta,a}^{(k)}. \tag*{(24)}

Here R\mathbf{R} is the 3×33 \times3 rotation/reflection matrix of the operation, R^(a)\hat{R}(a) is the atom to which atom aa is mapped, and the double sum runs over all three Cartesian components and all atoms. The sum includes both “diagonal” contributions (atoms mapped to themselves) and “off-diagonal” (permuted) ones, both of which are essential for obtaining the correct ±1\pm1 character when the mode displacement is concentrated on permuted atoms.

The PAF is obtained by translating the molecule to its centre of mass, computing the inertia tensor Ipq=∑ama(δpq∣ra∣2−ra,pra,q)I_{pq} = \sum_a m_a(\delta_{pq}\lvert\mathbf{r}_a\rvert^2-r_{a,p}r_{a,q}), and diagonalising it (LAPACK dsyev). The rotated coordinates and displacement vectors are used in Eq. (24).

The irrep Γk\Gamma_k of mode kk is identified by minimal L1L^1 distance between the computed character vector χ(k)\boldsymbol{\chi}^{(k)} and the rows of the point-group character table.

Configuration Irreps and Block Diagonalisation

The VSCF modals inherit the parity structure of the HO basis: R^ φni(i)(Qi)=[χ(Γi,R^)]niφni(i)(Qi)\hat{R}\,\varphi_{n_i}^{(i)}(Q_i) = [\chi(\Gamma_i,\hat{R})]^{n_i}\varphi_{n_i}^{(i)}(Q_i). Therefore the contribution of mode ii to the irrep of a configuration is Γi\Gamma_i when nin_i is odd and Γ1\Gamma_1 (the totally symmetric irrep) when nin_i is even. The irrep of a configuration is the direct product of all contributing odd-quanta mode irreps:

Γ(∣Φn⟩)=⨂i=1mi oddMΓ(Qi).(25)\Gamma(\lvert\Phi_n\rangle)=\bigotimes_{\substack{i=1\\m_i\ \mathrm{odd}}}^{M}\Gamma(Q_i). \tag*{(25)}

The configuration space partitions into nΓn_{\Gamma} blocks: S=⋃ΓBΓ\mathcal{S}=\bigcup_{\Gamma}\mathcal{B}_{\Gamma}. Each block is diagonalised independently, reducing the cost from O(∣S∣3)O(\lvert\mathcal{S}\rvert^3) to ∑ΓO(∣BΓ∣3)\sum_{\Gamma}O(\lvert\mathcal{B}_{\Gamma}\rvert^3). No approximation is introduced: the symmetry-adapted VCI is exactly equivalent to the full VCI for the correct group assignment.

Diagnostic Output and User Verification

The user specifies the point group via the PGROUP keyword. Because ViBrA cannot auto-detect the true symmetry, extensive diagnostic output is printed to stdout: PAF coordinates, atom permutation tables, per-atom contributions to each mode character, and the assigned irrep. A complete discussion of the consequences of incorrect group assignment is provided in the Supporting Information.

Transition Dipoles and Infrared Intensities

Infrared intensities are proportional to the squared transition dipole moment:

I0I∝∑α=x,y,z∣⟨Ψ0∣μ^α∣ΨI⟩∣2.(26)I_{0I}\propto\sum_{\alpha=x,y,z}\left\lvert\langle\Psi_0\lvert\hat{\mu}_{\alpha}\rvert\Psi_I\rangle\right\rvert^2. \tag*{(26)}

The dipole operator is expanded to second order in normal coordinates:

μ^α=μα(0)+∑i=1Mμα(i)Qi+12∑i=1M∑j=1Mμα(ij)QiQj.(27)\hat{\mu}_{\alpha}=\mu_{\alpha}^{(0)}+\sum_{i=1}^{M}\mu_{\alpha}^{(i)}Q_i+\frac{1}{2}\sum_{i=1}^{M}\sum_{j=1}^{M}\mu_{\alpha}^{(ij)}Q_iQ_j. \tag*{(27)}

where μα(i)≡∂μα/∂Qi∣0\mu_{\alpha}^{(i)}\equiv\left.\partial\mu_{\alpha}/\partial Q_i\right|_0 and μα(ij)≡∂2μα/∂Qi∂Qj∣0\mu_{\alpha}^{(ij)}\equiv\left.\partial^2\mu_{\alpha}/\partial Q_i\partial Q_j\right|_0.

Transformation of Dipole Derivatives from ORCA

The ORCA VPT2 output provides the first derivatives in Cartesian coordinates, ∂μα/∂xk\partial\mu_{\alpha}/\partial x_k. They are transformed to normal coordinates via

∂μα∂Qi=∑k=13Nat∂μα∂xkLkima,(28)\frac{\partial\mu_\alpha}{\partial Q_i} = \sum_{k=1}^{3N_{\mathrm{at}}} \frac{\partial\mu_\alpha}{\partial x_k} \frac{L_{ki}}{\sqrt{m_a}}, \tag*{(28)}

where LkiL_{ki} are the mass-weighted Hessian eigenvectors and a=⌈k/3⌉a = \lceil k/3 \rceil is the atom index. The result is a (3Nat×3)(3N_{\mathrm{at}} \times3) matrix; only the MM vibrational rows are retained.

The second derivatives are supplied in a mixed representation: ∂2μα/∂Qi∂xk\partial^2 \mu_\alpha/\partial Q_i \partial x_k, with the first index already in normal coordinates. Only the second Cartesian index requires transformation:

∂2μα∂Qi∂Qj=∑k=13Nat∂2μα∂Qi∂xkLkjma.(29)\frac{\partial^2 \mu_\alpha}{\partial Q_i \partial Q_j} = \sum_{k=1}^{3N_{\mathrm{at}}} \frac{\partial^2 \mu_\alpha}{\partial Q_i \partial x_k} \frac{L_{kj}}{\sqrt{m_a}}. \tag*{(29)}

The ORCA-provided array has dimensions M×3Nat×3M \times3N_{\mathrm{at}} \times3. Internally, the resulting M×M×3M \times M \times3 array is divided by 2 to absorb the factor of 1/21/2 from the Taylor expansion, so that the VCI dipole routine multiplies directly by the stored value.

Dipole Matrix Elements in the VCI Basis

The full dipole matrix element between configurations m\mathbf{m} and n\mathbf{n} consists of three contributions:

Linear dipole (one-body):

Dα(1)(m,n)=∑i=1Mμα(i)Ii(mi,ni;1)Sall\i.(30)D_\alpha^{(1)}(\mathbf{m},\mathbf{n}) = \sum_{i=1}^{M} \mu_\alpha^{(i)} \mathcal{I}_i(m_i,n_i;1) S_{\mathrm{all}\backslash i}. \tag*{(30)}

Diagonal second derivative (one-body):

Dα(2a)(m,n)=∑i=1Mμα(ii)Ii(mi,ni;2)Sall\i.(31)D_\alpha^{(2a)}(\mathbf{m},\mathbf{n}) = \sum_{i=1}^{M} \mu_\alpha^{(ii)} \mathcal{I}_i(m_i,n_i;2) S_{\mathrm{all}\backslash i}. \tag*{(31)}

Off-diagonal second derivative (two-body):

Dα(2b)(m,n)=∑i<j2μα(ij)Ii(mi,ni;1)Ij(mj,nj;1)Sall\{i,j}.(32)D_\alpha^{(2b)}(\mathbf{m},\mathbf{n}) = \sum_{i<j} 2\mu_\alpha^{(ij)} \mathcal{I}_i(m_i,n_i;1) \mathcal{I}_j(m_j,n_j;1) S_{\mathrm{all}\backslash\{i,j\}}. \tag*{(32)}

At the HO and VSCF levels, only Dα(1)D_{\alpha}^{(1)} is used, i.e., the linear dipole surface. At the VCI, S-VCI, and SA-VCI levels, all three contributions are evaluated. The quadratic terms Dα(2a)D_{\alpha}^{(2a)} and Dα(2b)D_{\alpha}^{(2b)} are essential for overtone and combination band intensities, which are strictly zero in the double-harmonic approximation, and for correctly redistributing oscillator strength among Fermi-resonant states.

The configuration-space dipole matrix is transformed to the eigenstate basis using BLAS: first tα=μα⋅c0t_{\alpha} = \boldsymbol{\mu}_{\alpha} \cdot\mathbf{c}_{0} (DGEMV), then ⟨ΨI∣μ^α∣Ψ0⟩=cIT⋅tα\langle\Psi_{I} \lvert\hat{\mu}_{\alpha} \rvert\Psi_{0} \rangle= \mathbf{c}_{I}^{T} \cdot t_{\alpha} (DDOT). Intensities are globally normalised by default for convenience in spectral visualisation and peak assignment. The normalisation is performed according to:

I0Inorm=100I~0I∑ΔE≤EcutoffI~0J2(33)I_{0I}^{\mathrm{norm}} = 100 \frac{\widetilde{I}_{0I}}{\sqrt{\sum_{\Delta E \le E_{\mathrm{cutoff}}} \widetilde{I}_{0J}^{2}}} \tag*{(33)}

where I~0I=∑α∣μα,0I∣2\widetilde{I}_{0I} = \sum_{\alpha} \lvert\mu_{\alpha,0I} \rvert^{2}, ΔE=EI−E0\Delta E = E_{I} - E_{0} is the transition energy from the ground state, and EcutoffE_{\mathrm{cutoff}} is a user-specified energy threshold with a default value of 4500 cm−1^{-1}. This ensures that the relative intensities are independent of higher-energy states that are often irrelevant for practical comparisons. Transitions with energy above EcutoffE_{\mathrm{cutoff}} are not displayed in the spectrum. The normalisation is explicitly defined so that the ratios between peaks are invariant with respect to the total number of calculated states. Absolute intensities, e.g. in units of km mol−1^{-1} can be obtained by modifying the source code to bypass this normalisation; as ViBRa is open-source, such modifications are straightforward for users requiring absolute values.

Program Architecture and User Utilities

ViBRa consists of a Fortran 90/95 computational engine and a Python 3 graphical interface. The computational engine is organized into modular components, each responsible for a distinct stage of the anharmonic vibrational spectroscopy workflow. The principal modules are summarized in Table 1.

ModuleRole
main.f90Entry point and orchestration layer for the execution workflow.
read_input.f90Keyword-based input parser with validation of user-specified options.
read_orca.f90Parser for ORCA .vpt2 output files. This module performs Hessian diagonalization, handles dipole transformations, and writes normal_mode.txt.
get_combination.f90Configuration-enumeration routines, including recursive stars-and-bars combinatorial counting, unique-element decomposition, and degeneracy-factor evaluation.
integrals.f90Analytical harmonic-oscillator matrix elements used in the construction of the vibrational Hamiltonian.
one_mode_operation.f90VSCF mean-field routines, including OpenMP-parallelized one-mode operations.
vci.f90Central VCI kernel. This module contains the shared compute_H_element routine and supports full VCI, selected VCI with EN-PT2 screening, symmetry-adapted VCI, and dipole-property calculations.
jacobi.f90LAPACK-based eigensolver wrappers, including DSYEVD and DSYEVR, with automatic workspace queries and eigenvector normalization.
symmetry.f90Symmetry-handling routines, including point-group character tables, symmetry operations, direct-product tables, normal-mode irrep assignment by principal-axis-frame projection, and block-diagonal VCI dispatch.

Table 1. Main computational modules in the ViBrA Fortran engine.

The engine produces three output files, which can be utilized by the visualisation interface:

  • vscf.out — complete log: VSCF convergence histories, modal coefficient matrices, VCI configuration list, eigenstate energies, and the three leading CI coefficients with quantum-number assignments;

  • intensities.txt — transition frequencies (cm−1\mathrm{cm}^{-1}) and normalised intensities for HO, VSCF, and VCI/SCI methods;

  • normal_mode.txt — equilibrium geometry (Å) and Cartesian displacement vectors for all 3Nat3N_{\mathrm{at}} modes.

Graphical User Interface

A CustomTkinter application, Figure 1 (VSCFVCIApp) simplifies job preparation. The user navigates to the ORCA .vpt2 file; the working directory is set automatically. A modal dialog collects all keywords (Table 2), validates their ranges with tooltip guidance, and writes input_vscf.txt. The Fortran executable is launched as a subprocess in a background thread, with real-time output streaming to a scrollable panel. A stop button terminates a running job, and the panel supports save and clear operations.

Graphical interface of ViBra, showing the capability to directly load ORCA VPT2 output files, generate input files, and execute the code.

Figure 1. Graphical interface of ViBra, showing the capability to directly load ORCA VPT2 output files, generate input files, and execute the code.

KeywordTypeDescription
NMODESintNumber of vibrational modes (MM).
NEXPANintHO basis size per mode (NexpN_{\mathrm{exp}}).
FILECTstrPath to ORCA .vpt2 file.
CTEMODstrFormat (orca_vpt2).
NQUANTintMaximum total quanta (NqN_q); ≤0\le0 disables VCI.
NSTATEintNumber of eigenstates (≤0=\le0 = all).
CVGSCFintVSCF convergence exponent (10−C cm−110^{-C}\ \mathrm{cm}^{-1}).
THREADintNumber of OpenMP threads.
PGROUPstrPoint group (C1, Cs, Ci, C2, C2h, C2v, D2, D2h).
PROJCTrealProjection cutoff for symmetry detection (Å).
MAXSCIintNselN_{\mathrm{sel}} for S-VCI. Default: 100.
Usage: MAXSCI [N] [mode] [ref]
N>0N > 0, mode = auto: iterative selection with CI reference (ref = s, d, t, q).
N>0N > 0, mode = list: iterative selection with list reference, no ref needed.
N=0N = 0: full VCI (no S-VCI), no mode and no ref needed.
N=0N = 0, mode = list: full VCI using user-provided state list, no ref needed.

Table 2. Principal input keywords of ViBra.

Interactive Spectral Viewer

The viewer, Figure 2 (MainApplication, Python 3, Matplotlib, 3Dmol.js) reads the three output files and provides a rich environment for analysis and comparison:

Dedicated ViBra spectrum visualizer

Figure 2. Dedicated ViBra spectrum visualizer, highlighting the capability to directly compare HO, VSCF, and VCI calculations with experimental data, while seamlessly analyzing the peak composition associated with each assigned mode. The displayed results correspond to the ethylene molecule, where the full VCI method was applied including up to 6 quanta, and 10 HO functions per mode were used in the VSCF calculation. Experimental data were retrieved from the National Institute of Standards and Technology database. [33]

  • Gaussian broadening: each transition at frequency νk\nu_k with intensity IkI_k is represented as

S(ν)=∑Ikexp⁡[−(ν−νk)2/(2σ2)]S(\nu) = \sum I_k \exp[-(\nu-\nu_k)^2/(2\sigma^2)]

with σ=FWHM/(22ln⁡2)\sigma= \mathrm{FWHM}/(2\sqrt{2\ln2}). FWHM is adjustable via stepper and direct text entry.

  • Temperature-dependent spectra: for T>0T > 0 K, VCI intensities are Boltzmann-weighted:

Ikeff=Ik(p0−pk)I_k^{\mathrm{eff}} = I_k\left(p_0-p_k\right), where pk=exp⁡(−hcνk/kBT)/Zp_k=\exp\left(-hc\nu_k/k_{\mathrm{B}}T\right)/Z and hc/kB=1.438777 cm⋅Khc/k_{\mathrm{B}}=1.438777\ \mathrm{cm\cdot K}. Temperature is adjustable from 0 to 950 K.

  • JCAMP-DX overlay: experimental spectra in JCAMP-DX format (.jdx/.dx) can be loaded. Transmittance is automatically converted to absorbance, negative values are clipped, and the result is normalised.

  • ALS baseline correction: for experimental spectra exhibiting a sloping baseline, an Asymmetric Least Squares (ALS) algorithm[34] estimates and subtracts a smooth baseline. The cost function Q(z)=∑iwi(yi−zi)2+λ∥Dz∥2Q(\mathbf{z})=\sum_i w_i(y_i-z_i)^2+\lambda\lVert\mathbf{D}\mathbf{z}\rVert^2 is minimised iteratively, where D\mathbf{D} is the second-difference matrix. The smoothness parameter λ\lambda and asymmetry parameter pp are adjustable via steppers with real-time redisplay.

  • Peak inspection: clicking a VCI peak label displays the frequency, intensity, and three leading CI coefficients with quantum-number assignments (e.g., “ν1+2ν3\nu_1+2\nu_3”). Buttons launch 3D normal-mode animations for any involved mode.

  • 3D normal-mode animation: sinusoidal displacement trajectories (60 frames, adjustable amplitude) are generated from normal_mode.txt, saved as JSON, and rendered via an auto-generated HTML page with embedded 3Dmol.js viewer supporting play/pause, frame slider, speed control, and adjustable rendering styles and atomic radii. See Figure 3.

Normal mode animator displaying force vectors and symmetry elements

Figure 3. Normal mode animator, featuring the ability to display force vectors and symmetry elements whenever a point group other than C1 is employed.

  • Data export: the displayed spectrum can be saved as a text file and as a 300 dpi PNG image.

Computational Optimisation Strategies

ViBra employs several targeted algorithmic optimisations that are essential for making VCI calculations on molecules with M≈30M\approx30–4040 and larger tractable. The complete rationale for each strategy is given in the manual, provided as supporting information; the most impactful are summarised here.

Overlap Cache in VSCF. At the beginning of each VSCF iteration, all scalar overlaps ⟨φn(j)∣Qcj∣φn(j)⟩\langle\varphi_n^{(j)} \vert Q_c^{j} \vert\varphi_n^{(j)} \rangle (c=0,…,4)(c = 0, \ldots, 4) are precomputed and cached. The Xi(p)X_i^{(p)} coefficients are then assembled via O(1)O(1) look-ups rather than repeated O(Nexp2)O(N_{\mathrm{exp}}^2) double sums.

Vectorised potentials and nonzero flags. Force constants VijkV_{ijk} and VijklV_{ijkl} are permutationally symmetric, so a dense M3M^3 or M4M^4 array wastes memory and access time on redundant duplicates. Instead, each is flattened into a 1D array of length T3T_3 (respectively T4T_4), containing one entry per sorted index combination i≤j≤ki \le j \le k (respectively i≤j≤k≤li \le j \le k \le l), together with precomputed metadata: the distinct mode indices appearing in the term, their multiplicities (e.g. ViijV_{iij} has mode ii with multiplicity 2 and mode jj with multiplicity 1), and a boolean flag marking whether the term exceeds a numerical threshold and is therefore treated as nonzero. This flattening removes redundant storage and lets the Hamiltonian-construction loop iterate directly over the T3+T4T_3 + T_4 distinct terms rather than the full M3+M4M^3 + M^4 index space, with the nonzero flag skipping negligible terms at no extra cost.

Sparse pair list. Before building HH, the code makes a single pass over all configuration pairs and records only those with d≤4d \le4 differing modes, storing for each such pair the indices of the modes that differ. The Hamiltonian-construction loop then iterates directly over this precomputed list rather than over all ∣S∣(∣S∣+1)/2|\mathcal{S}|(|\mathcal{S}| + 1)/2 pairs, so pairs that are known in advance to give zero coupling are never visited.

Inverted potential index. For a pair differing in dd modes, a cubic or quartic term contributes only if it involves all dd differing modes; otherwise modal orthogonality forces at least one factor to vanish. Scanning the full list of T3+T4T_3 + T_4 potential terms for every pair would waste most of the work rejecting non-matching terms. Instead, each mode is pre-indexed with the list of potential terms it appears in. For a given pair, only the terms attached to the first differing mode are examined, and each is quickly checked for the remaining d−1d - 1 differing modes. This reduces the average per-pair potential scan from O(T3+T4)O(T_3 + T_4) to O(T3+T4Md−1)O\left(\frac{T_3 + T_4}{M^{d-1}}\right) since a term touches a given mode with probability ∼1/M\sim1/M.

Prefix–Suffix Overlap Products. The spectator product Sall\SS_{\mathrm{all}\backslash S} (Eq. (20)) is needed for every Hamiltonian and dipole contribution. Prefix and suffix arrays of ok=Ik(mk,nk;0)o_k = I_k(m_k,n_k;0) are computed once per matrix element in O(M)O(M) time, enabling O(1)O(1) evaluation of any excluded-mode product. For multi-mode exclusion, the total product is divided by the excluded factors; a fallback explicit product handles the case where any overlap vanishes.

OpenMP Parallelism. The Hamiltonian construction loop over the sparse pair list, the EN-PT2 screening in SCI, and the dipole matrix construction are parallelised with dynamic scheduling and empirically tuned chunk sizes (16–256). Dynamic scheduling is essential because pairs with ndiff=0n_{\mathrm{diff}} = 0 (diagonal) evaluate all nonzero potential terms while high-ndiffn_{\mathrm{diff}} pairs evaluate far fewer, creating significant load imbalance. The VSCF stage runs in serial since, with the overlap cache, it completes significantly faster.

LAPACK/BLAS Integration. The VSCF modal matrices are diagonalised with DSYEVD. For VCI/S-VCI, DSYEVD is used when all eigenvalues are requested; DSYEVR computes only the lowest NstateN_{\mathrm{state}} eigenvalues (if point group is C1 and no external list is given) for partial spectra, saving both memory and time. Dipole transformations use DGEMV and DDOT.

Memory Considerations. The full VCI Hamiltonian is stored as a dense symmetric matrix, which limits the practical configuration space size to ∣S∣∼104|\mathcal{S}| \sim10^{4}–10510^{5} depending on available memory (the matrix requires ∣S∣2×8|\mathcal{S}|^{2} \times8 bytes). The SCI approach directly alleviates this by reducing the matrix dimension. The modal integral table is compact (M×(Nq+1)2×6M \times(N_q+1)^{2} \times6 entries). The sparse pair list and inverted index use integer arrays whose size is data-dependent but typically modest compared to the Hamiltonian itself.

Validating and using ViBra

Validation of the ViBra implementation

We carried out a systematic validation of the ViBra code through four complementary benchmarks, progressing from the simplest VSCF level to fully correlated VCI and symmetry-adapted formulations. All geometry files, ORCA inputs and outputs, ViBra inputs and outputs, as well as the Crystal23 output for the water, are provided as Supporting Information for numerical comparison. For the water benchmark (step 1 and 2), the agreement with reference data is within numerical precision; therefore, we do not report the numerical outputs in the main text, and the reader is referred to the SI for full details. For the validation of the SA-VCI and S-VCI implementations (step 3 and 4), the complete dataset, comprising all vibrational modes, VCI coefficients, and energy contributions for each molecule is extensive and is provided in its entirety in the Supporting Information and in a public Zenodo dataset. [35] In the following § 5.4, we present the ethylene molecule as a representative case study to illustrate the key findings.

Step 1: Validation of the VSCF implementation

Goal: to verify the correctness of the VSCF module in ViBra. Action: we first validated the VCI@VSCF implementation by benchmarking VSCF ground-state energies of a water molecule computed with Crystal23 at the B3LYP/6-31G* level of theory. VSCF calculations were performed using a basis set of 10 harmonic oscillator (HO) functions per vibrational mode. We obtained the reference data (i.e. cubic and quartic force constants and normal mode frequencies) from the Crystal23 website, where they are provided as part of the anharmonic vibrational tutorial. [14, 36] Therefore, we performed no computations at this level of theory, nor did we make use of Crystal23. Comparison: we compared ground-state energies, fundamental transition energies, and optimized modal coefficients. The computed energies show excellent agreement with the reference values, with deviations in the range of 0.01–0.001 cm−1\mathrm{cm}^{-1}. The modal coefficients differ on the order of 10−310^{-3} or lower, also indicating excellent agreement. Conclusion: the VSCF implementation in ViBra is correct and reproduces reference data with numerical accuracy.

Validation of the VCI implementation

Goal: to verify the correctness of the VCI diagonalization and eigenstate analysis in ViBra.

Action: we analyzed VCI energies and eigenstate coefficients using the same water molecule benchmark. A total excitation truncation of up to 5 quanta was utilized. Comparison: the observed discrepancies between ViBra and Crystal23 are of the same magnitude as those obtained at the VSCF level and are attributed to numerical noise. The most important configurations, together with their numerical contributions, for each VCI state were correctly identified. Conclusion: These results demonstrate that the entire workflow, from the initial VSCF step to the final diagonalization of the VCI Hamiltonian matrix, is correctly implemented in ViBra. In particular, the computed eigenvalues (energies) and eigenvectors (VCI coefficients) are in full agreement with reference results, confirming the correctness of the implementation.

Validation of the SA-VCI implementation

Goal: To validate the symmetry-adapted VCI (SA-VCI) formulation against full VCI calculations. Methodology: We employed semi-quartic force fields, defined here as cubic force constants together with quartic contributions in which at least two vibrational modes are identical. Geometry optimizations, energy minimizations, and force constant computations were carried out using ORCA 6.1 at the wB97X-D4/aug-cc-pVTZ level of theory. The test set comprised molecules spanning several point-group symmetries: HFCO\mathrm{HFCO} (CsC_s), N2H4\mathrm{N_2H_4} (C2C_2), H2O\mathrm{H_2O} (C2vC_{2v}), trans-N2H2\mathrm{N_2H_2} (C2hC_{2h}), and C2H4\mathrm{C_2H_4} (D2hD_{2h}). Comparison: For all molecules examined, the low lying state energies obtained from full VCI and SA-VCI agree to within numerical noise, on the order of 10−2 cm−110^{-2}\,\mathrm{cm}^{-1} or better. Conclusion: The SA-VCI formulation is correctly implemented and reproduces full VCI results to numerical precision. Computational details: Our choice of functional and basis set was driven not by the pursuit of highly accurate vibrational frequencies, but by considerations of balanced computational cost and methodological robustness. The range separated hybrid functional wB97X-D4 combined with the aug-cc-pVTZ basis set is widely used in benchmark vibrational studies, including several entries at the CCCBDB/NIST frequency database.[37] In the VSCF step, we adopted a basis of 10 harmonic oscillator functions per vibrational mode, a choice motivated by the rapid convergence of VSCF energies with respect to the number of HO functions and by the fact that this basis provides a balanced description of vibrational states with few quanta of excitation per mode.[14] For both the full VCI and SA-VCI expansions, we included configurations with up to 4 quanta. Important note on symmetry handling: Despite the tight convergence criteria for minimizing the geometries, no symmetry constraints were imposed during the ORCA computations; geometries, Hessians, and force constants were evaluated entirely without point-group restrictions. This choice was deliberate for the present validation study: our objective is also to probe the robustness of the SA-VCI implementation against minute numerical asymmetries that may arise in the quartic force field, we caution that for production level applications users are strongly advised to enforce appropriate symmetry constraints during the electronic structure calculations. Even tiny deviations introduced by a non symmetrized quartic force field can accumulate, leading to larger errors, specially for higher energy states.

Step 4: Three-tier benchmark of intensities and variational selection

Goal: To validate the intensity calculations and the selected-state VCI (S-VCI) procedure. Action: We carried out a three tier benchmark that begins from the standard double harmonic approximation and progressively includes higher order effects. The electronic structure calculations were the same as those described in Step 3, and all molecules in the test set were benchmarked. Tier 1 (DHA validation): We benchmarked intensities obtained with ORCA 6.1 and ViBrA within the harmonic approximation using linear dipole derivatives, finding identical results for all molecules, noting that ViBrA intensities are internally normalized. This step corresponds to the full double harmonic approximation (harmonic frequencies + linear dipole derivatives). Tier 2 (Electrical anharmonicity): We analyzed ViBrA energies and intensities including also second-order dipole derivatives, both within the full VCI framework and within the symmetry-adapted VCI (SA-VCI) formulation. For all molecules, the resulting energies agree within numerical noise, on the order of 10−2 cm−110^{-2}\ \mathrm{cm}^{-1} or lower. The intensities obtained from SA-VCI and full VCI are identical within numerical precision. Tier 3 (S-VCI variational validation): To validate the S-VCI (selected states via EN-PT2), we analyzed the variational behavior of the method by systematically increasing the number of included states. As expected for a variational approach, all state energies decreased monotonically upon enlarging the configuration space by increasing MAXSCI. Conclusion: The intensity implementations (both DHA and with second-order derivatives) are correct, and the S-VCI selection procedure behaves variationally as expected.

Using ViBra

To demonstrate the efficiency of the software, particularly for VCI and its variants, we consider the ethylene (C2H4\mathrm{C}_2\mathrm{H}_4) molecule. Its D2hD_{2h} symmetry provides a clear example of the computational advantages gained from block diagonalization. Since VSCF calculations are computationally inexpensive, we do not provide a detailed performance analysis for this step. We employ a fixed basis of 10 harmonic oscillator functions per vibrational mode and an energy convergence threshold of 10−6 cm−110^{-6}\,\mathrm{cm}^{-1} in the VSCF procedure preceding all VCI and VCI-variant calculations.

It is important to note that all frequency calculations performed with ViBra were carried out on an 11th Gen Intel® Core™ i9-11900K processor, with the CPU frequency capped at 3.5 GHz. The machine was equipped with 2x32 GB of RAM running at 3200 MHz, and cache sizes of 640 KB (L1), 4.0 MB (L2), and 16.0 MB (L3). The code is compiled with ifx, Intel’s next-generation LLVM-based Fortran compiler, using -O3 for aggressive optimisation including loop transformations and vectorisation. Parallelism is enabled at multiple levels: /Qopenmp activates explicit OpenMP directives, /threads enables compiler auto-parallelisation of loops, and /Qmkl:parallel links the threaded Intel Math Kernel Library for parallel BLAS/LAPACK routines. The /Qm64 flag targets the 64-bit instruction set, while /heap-arrays allocates temporary arrays on the heap to prevent stack overflow from large automatic arrays. Finally, /fpscomp:logicals enforces compatibility with older Intel Fortran logical behaviour. A pre-compiled executable is provided alongside the source code in the repository.

Using this computational setup, we examine the dependence of basis size, energy convergence, and diagonalization cost on the maximum number of quanta included in the VCI expansion, as summarized in Table 3. Note that a single core was set for this benchmark. The Hamiltonian dimension increases rapidly as the maximum excitation level is raised. Increasing the maximum number of allowed quanta from 4 to 5, and subsequently from 5 to 6, approximately triples the number of vibrational configurations at each stage, resulting in an approximately ninefold increase in the number of Hamiltonian matrix elements (and hence memory footprint) for every additional maximum quanta.

# Quanta# Statestdiagt_{\mathrm{diag}}ttott_{\mathrm{tot}}ZPEE1E_1ΔE\Delta E
2910011191.405612036.0327844.6271
34550011129.722612034.6711904.9484
418201211127.002811959.2915832.2886
56188384411126.773711953.6805826.9069
6185641400143811125.515911952.9510827.4351
6(20)1856463267011125.515911952.9510827.4351

Table 3. Convergence of VCI energies and computational cost with increasing excitation quanta for ethylene. The final row corresponds to partial diagonalization of the 6-quanta Hamiltonian requesting only the lowest 20 eigenvalues. All energies are reported in cm−1^{-1} and timings in seconds.

The zero point energy (ZPE) and the first excited state energy (E1E_1) both decrease monotonically as the maximum excitation level is increased, as expected from the systematic enlargement of the VCI configuration space. However, their convergence rates differ significantly. The ZPE is already effectively converged when a maximum of 3 quanta is included, whereas E1E_1 requires at least 4 quanta to reach a practically stable value. This behavior is consistent with the semi quartic character of the force field. Because the reductions in the ZPE and E1E_1 are not proportional, the transition energy, ΔE=E1−ZPE\Delta E = E_1 - \mathrm{ZPE}, does not exhibit monotonic convergence. Instead, it increases by approximately 60 cm−1\mathrm{cm}^{-1} when the maximum excitation level is increased from 2 to 3 quanta before decreasing to a value below that obtained in the calculation with 2 quanta.

The apparent stabilization at 4 quanta should not be interpreted as strict numerical convergence. Rather, it indicates that the remaining changes are sufficiently small for the practical applications considered here and do not justify the substantial additional computational expense associated with including higher excitation levels. Therefore, a maximum of 4 quanta provides an appropriate compromise between accuracy and computational efficiency for routine calculations. When higher numerical accuracy is required, however, calculations including 5 or 6 quanta may still be necessary.

This need for increased accuracy must be balanced against the rapidly increasing computational cost. Both the memory requirements and the CPU time increase dramatically as the maximum number of allowed quanta is raised from 4 to 6, with matrix diagonalization becoming the dominant computational bottleneck. For the largest calculation, corresponding to a maximum of 6 quanta and a Hamiltonian containing 18564 vibrational configurations, diagonalization with DSYEVD required 1400 s (23.3 min) out of a total runtime of 1438 s.

To reduce this computational expense, a partial diagonalization was also performed using DSYEVR, requesting only the lowest 20 eigenvalues. This approach avoids constructing and storing the complete set of eigenvectors, reducing the diagonalization time to 632 s (10.5 min) and the total runtime to 670 s, as reported in the final column of Table 3. Although the eigenvectors are no longer available for subsequent intensity calculations, the reduction in computational cost is substantial and becomes increasingly advantageous when only a limited number of state energies is required. Consequently, for applications that require only the ZPE, such as the calculation of zero point energy corrections to electronic energies in chemical reaction transition barrier calculations, partial diagonalization provides a particularly efficient strategy.

The full diagonalization of the 6-quanta Hamiltonian, although manageable for ethylene, remains computationally demanding. This motivates the use of selected VCI (S-VCI) and symmetry-adapted VCI (SA-VCI) to further reduce the cost while preserving accuracy. Table 4 compares the performance of both approaches using 8 OpenMP threads; the reported real time refers to the total wall-clock time encompassing state selection (S-VCI) or symmetrization (SA-VCI), Hamiltonian construction, diagonalization, and intensity evaluation. In the S-VCI columns, the number following the slash denotes the value of the MAXSCI keyword, which controls how many states are retained by the perturbative selection criterion. Increasing MAXSCI systematically enlarges the selected space and drives the energies toward the full-VCI limit. Already at MAXSCI=50, the ZPE and E1E_1 lie within 1.6 and 2.3 cm−1^{-1} of the full-VCI result, respectively, and further increasing the keyword from 100 to 300 yields only marginal improvement. This indicates that low MAXSCI values may filter too aggressively, discarding states that contribute significantly to the correlation energy (see § 6.1); a moderate threshold (here 50) already captures the essential physics. The optimal value is molecule-dependent and should be calibrated, when possible, by comparing a small-basis S-VCI calculation against full VCI, as done here. Notably, within the S-VCI kernel, we used the default VCISD reference, with external states generated up to NQUANT=6.

Method# StatesCPU time (s)Real time (s)ZPE (cm−1^{-1})E1E_1 (cm−1^{-1})
S-VCI/50319331411127.082011955.2475
S-VCI/10059211381811126.708311953.8296
S-VCI/15078952082811126.537411953.6018
S-VCI/20092413034011126.505611953.4682
S-VCI/250103394145511126.426511953.4365
S-VCI/300111404946611126.374911953.4153
Full VCI18564200526511125.515911952.9510
SA-VCI (D2hD_{2h})18564891311125.520211952.9591
SA-VCI (CiC_i)185646018111125.517711952.9550

Table 4. Performance comparison of selected VCI (S-VCI) and symmetry-adapted VCI (SA-VCI) against full VCI for ethylene at the 6-quanta level. S-VCI rows are labelled by the MAXSCI keyword value. All calculations used 8 OpenMP threads.

Turning to symmetry-adapted VCI, the D2hD_{2h} point group of ethylene partitions the Hamiltonian into eight irreducible representation blocks. In the ideal scenario where all blocks are of equal size, each block has dimension N/8N/8, and the cubic scaling of diagonalization (e.g., using the DSYEVD subroutine) reduces the total cost by a factor of 8×(1/8)3=1/648 \times(1/8)^3 = 1/64, yielding a theoretical maximum speedup of 64×64 \times. In our implementation, the symmetry blocks are processed sequentially, but each block diagonalization is executed using all available OpenMP threads, rather than distributing blocks across threads and subdividing the thread team. This design ensures that the full multi-threaded capability of the diagonalizer is applied to every block, avoiding load imbalance when the number of symmetry operations does not evenly divide the thread count.

In practice, the observed reduction in real time is approximately 20×20 \times (from 265 s for full VCI to 13 s for SA-VCI/D2hD_{2h}). The primary reason for 20×20 \times speed up is simple: not all operations scale cubically, so a 64×64 \times speedup is never expected, it will always be less. Moreover, it should be noted that SA-VCI and full VCI employ different computational kernels, and the timings reported in Table 4 reflect the total execution time from initialization to completion, not solely the diagonalization step. Yet, the irreducible blocks are not perfectly equal in size, so the sum of cubed block dimensions exceeds 8×(N/8)38 \times(N/8)^3, reducing the theoretical gain. Also, the non-diagonalization overhead (symmetrization, integral transformation, etc.) accounts for a larger fraction of the total time in the SA-VCI run, further diluting the speedup. Finally, the per-block matrices (around 2300×23002300 \times2300) may be too small to keep all 16 threads fully utilized, leading to higher parallel overhead than in the full VCI diagonalization of the much larger unblocked matrix.

These effects are partially offset by improved cache efficiency when blocks or specific operations requires arrays that fit into lower cache levels, but the net result is a reduction from the ideal 64×64 \times to ∼20×\sim20 \times.

Notably, the gap between ideal and observed speedup narrows as the number of symmetry blocks decreases. Using a smaller symmetry group, such as CiC_i, yields only two blocks, a theoretical speedup of 4×4 \times, and an observed reduction to 81 s—an efficiency of 82% relative to the ideal, compared to 31% for the D2hD_{2h} case. The larger block size in CiC_i restores diagonalization as the dominant cost, and the ∼2300×2300\sim2300 \times2300 blocks are replaced by matrices roughly 16 times larger, ∼9200×9200\sim9200 \times9200, which make better use of the available threads. This trend indicates that our sequential-block parallelization strategy performs best when blocks are large enough to saturate the multi-threaded overhead.

Importantly, the energies obtained with SA-VCI/D2h\mathrm{SA\text{-}VCI}/D_{2h}, SA-VCI/Ci\mathrm{SA\text{-}VCI}/C_i, and full VCI are numerically equivalent to within a few hundredths of a cm−1\mathrm{cm}^{-1}, confirming that symmetry adaptation introduces no approximation beyond block diagonalization. A word of caution is warranted: assigning normal modes to irreps requires the correct point group. Using a lower symmetry than the true molecular symmetry (e.g., D2hD_{2h} for benzene instead of D6hD_{6h}) can lead to incorrect irrep assignments, particularly for degenerate modes. It is therefore advisable to run a low-quanta SA-VCI calculation and verify that the resulting energies match the corresponding full-VCI values before proceeding to larger bases.

Selected VCI as a bridge to quantum simulation algorithms

The selected-VCI machinery in ViBra can also be used in a quantum-centric workflow for vibrational problems. In a conventional ViBra calculation, the initial and selected spaces are generated classically using the procedures described in § 2.5. In the quantum-centric setting considered here, a quantum sampling routine instead provides an externally generated seed set of vibrational configurations. ViBra then acts as the classical selected-VCI engine: it can construct and diagonalize the projected anharmonic Hamiltonian directly in that seed or optionally enlarge it through EN-PT2-guided configuration selection before evaluating energies, eigenvectors, transition moments, and spectra.

This division of labour is closely related to recent sample-based quantum diagonalisation strategies in electronic-structure theory. In sample-based quantum diagonalisation (SQD) and sample-based Krylov quantum diagonalisation (SKQD), a quantum processor is used to generate basis-state samples, while the final Hamiltonian projection, diagonalisation, and observable evaluation are performed classically.[24, 25, 38] The same principle can be applied to vibrational structure. Instead of electronic occupation strings, the samples correspond to vibrational configurations

n=(n1,…,nM),\mathbf{n} = (n_1, \ldots, n_M),

where nin_i labels the occupied modal or harmonic-oscillator level of mode ii. The sampled configurations define a compact subspace SQ\mathcal{S}_{\mathrm{Q}}, and ViBra evaluates

Hαβ(SQ)=⟨Φα∣H^vib∣Φβ⟩,Φα,Φβ∈SQ,(34)H_{\alpha\beta}^{(\mathcal{S}_{\mathrm{Q}})} = \langle\Phi_\alpha| \hat{H}_{\mathrm{vib}} | \Phi_\beta\rangle,\qquad\Phi_\alpha, \Phi_\beta\in\mathcal{S}_{\mathrm{Q}}, \tag*{(34)}

using the same modal-integral and Hamiltonian-construction kernels employed in the classical VCI and S-VCI calculations. This distinction is central: ViBra does not replace the quantum sampling step. The quantum-facing routine supplies the initial seed configurations, while ViBra converts that seed into a projected VCI calculation and, when requested, systematically enriches it through classical EN-PT2-guided selection.

Figure 4 summarises the proposed workflow. The starting point is the same as in a conventional anharmonic vibrational calculation: electronic-structure calculations provide normal-mode frequencies, cubic and quartic force constants, and dipole derivatives. From these data, ViBra constructs the VSCF modal basis and the corresponding VCI Hamiltonian.

Quantum–classical workflow connecting ViBra to quantum vibrational spectroscopy algorithms

Figure 4. Quantum–classical workflow connecting ViBra (green) to quantum vibrational spectroscopy algorithms. The quantum-facing part of the workflow (yellow) maps the vibrational Hamiltonian to qubit operators and uses fragmentation and Trotterized time evolution to generate samples or time-domain data. The sampled configuration space is then passed back to ViBra, which constructs and diagonalises the projected VCI Hamiltonian and evaluates spectroscopic observables (blue). These can then be validated against full VCI or even experimental data when available.

For the quantum-facing part of the workflow, the same vibrational Hamiltonian is mapped to qubit operators using a boson-to-qubit or modal-to-qubit representation. In the compact configuration-index encoding employed here, the D(M,Nq)=(Nq+MM)D(M,N_q)=\binom{N_q+M}{M} configurations satisfying ∑i=1Mni≤Nq\sum_{i=1}^{M} n_i \le N_q are mapped bijectively to computational-basis states of an nqb=⌈log⁡2D(M,Nq)⌉n_{\mathrm{qb}}=\lceil\log_2 D(M,N_q)\rceil-qubit register, with any remaining basis states treated as unused padding. After mapping, the Hamiltonian takes the Pauli form

H^vib⟶H^q=∑kckPk,(35)\widehat{H}_{\mathrm{vib}} \longrightarrow\widehat{H}_{q}=\sum_k c_k P_k, \tag*{(35)}

where PkP_k are tensor products of single-qubit Pauli operators. The mapped Hamiltonian can then be partitioned into mutually commuting blocks of Pauli operators,

H^q=∑νH^ν,(36)\widehat{H}_{q}=\sum_{\nu}\widehat{H}_{\nu}, \tag*{(36)}

so that a real-time propagator can be approximated using a product formula, such as the first-order Suzuki-Trotter decomposition

U^Trot(t)=[∏νexp⁡(−iH^νt/r)]r+O(t2/r),(37)\widehat{U}_{\mathrm{Trot}}(t)=\left[\prod_{\nu}\exp\left(-i\widehat{H}_{\nu}t/r\right)\right]^r+O(t^2/r), \tag*{(37)}

with rr Trotter steps. For the initial state of the evolution, we have used a normalized superposition of the vacuum and all first single-mode excitations. The resulting circuits can be used either to generate samples from states relevant to the low-energy vibrational spectrum, or to estimate a time-domain correlation function directly.

Given a set of measured bit strings {b}\{\mathbf{b}\}, the classical post-processing stage decodes each valid string into a vibrational configuration n=D(b)\mathbf{n}=\mathcal{D}(\mathbf{b}). In a binary modal encoding, invalid strings that do not contain one occupied modal per mode can be removed by post-selection or, when appropriate, processed using a configuration-recovery procedure analogous to that introduced in the original SQD work. [24] The quantum-sampled subspace is then

SQ={n∣n=D(b), b∈samples, c(b)≥cmin⁡},(38)S_{\mathrm{Q}} = \left\{ \mathbf{n} \mid\mathbf{n} = \mathcal{D}(\mathbf{b}),\ \mathbf{b} \in\mathrm{samples},\ c(\mathbf{b}) \ge c_{\min} \right\}, \tag*{(38)}

where c(b)c(\mathbf{b}) is the observed count of bit string b\mathbf{b} and cmin⁡c_{\min} is an optional count threshold. In practice, the sampled space can be formed as the union of configurations obtained from multiple state preparations, time steps, or Krylov-like states. Once SQS_{\mathrm{Q}} is defined, no additional quantum measurements are required to evaluate the projected Hamiltonian matrix elements: they are computed classically by ViBra from the force field and modal integrals.

The quantum-facing components used in this proof of concept are implemented outside the main ViBra code base. In particular, the routines for modal-to-qubit mapping, Pauli-operator construction, and Trotterized time evolution follow the implementation provided in [39, 40] These external routines are used only to generate the qubit Hamiltonian, fragmented evolution operators, quantum circuits, and simulated samples. The subsequent construction of the projected vibrational Hamiltonian, diagonalization of the sampled configuration space, and evaluation of transition intensities are performed by ViBra.

Workflow Validation for H2OH_2O

As a proof-of-concept demonstration, we applied this workflow to the water molecule. Water is a convenient initial test system: its three vibrational modes keep the full, truncated VCI problem exactly tractable, while still incorporating nontrivial anharmonic couplings and infrared-active fundamentals. We performed the calculation for several values of truncation of the configuration space NqN_q, varying it from 2 to 12, with the full VCI(12) calculation serving as our reference. All ViBra calculations in this section were performed using: NEXPAN=15, CVGSCF=6. We used the same semiquartic force-field data (wB97X-D4/aug-cc-pVTZ) as in the validation section (Steps 3 and 4).

All quantum-facing calculations in this study were performed using a simulator; no quantum hardware was used. For each value of NqN_q, we prepared a different quantum circuit and we evolved them with 4 Trotter steps separated by a dimensionless time δ=0.1\delta= 0.1, or a physical time step of 0.531 ps0.531\,ps, and each circuit was sampled 1000 times. The number of qubits, as well as other relevant information on the circuits is provided in Table 5.

NqN_qQubits1 qubit gates2 qubit gates
241421494
3558872142
46239028894
56238958894
679627436222
779628336222
88386446146174
98386414146174
1091548396587262
1191548438587262
1291548446587262

Table 5. Gate counts for the bosonic SKQD circuits

The sampled bit strings were decoded into vibrational configurations, and the union of the configurations over all time steps defines the quantum-generated seed space SQ(Nq)\mathcal{S}_{\mathrm{Q}}(N_q). The results reported below correspond to \plot_water_qpu_bench\bosonic_selected_states_water_Nq_quanta in the database.1

The sampled configurations were supplied to ViBRA in list-reference mode. With MAXSCI 0 list, ViBRA constructs and diagonalizes the Hamiltonian directly in SQ\mathcal{S}_{\mathrm{Q}}. With MAXSCI mm list, where m=1m = 1, 2, or 3, the sampled list is instead used as the reference for an optional classical enlargement. External configurations are ranked for each targeted root using the Epstein–Nesbet second-order perturbative importance measure described in § 2.5. The highest-ranked configurations are added to the quantum seed, duplicates are removed, and the Hamiltonian is variationally rediagonalized in the resulting space,

S(m)=SQ∪SEN(m).\mathcal{S}^{(m)} = \mathcal{S}_{\mathrm{Q}} \cup\mathcal{S}_{\mathrm{EN}}^{(m)}.

Table 6 reports the total number of states obtained in each case.

NqN_q56789101112
VCI5684120165220286364455
Quantum seed, MAXSCI=05165827474616253
ViBra-enlarged, MAXSCI=18410512511811810010186
ViBra-enlarged, MAXSCI=2107125149136136120122106
ViBra-enlarged, MAXSCI=3115142163155155136140118

Table 6. Dimensions of the quantum-sampled seed spaces and of the spaces obtained after optional EN-PT2-guided enlargement by ViBra. The row with MAXSCI=0 corresponds to the direct use of the externally sampled seed. The remaining rows are quantum-seeded spaces that have been enlarged by ViBra and variationally rediagonalized. The VCI row gives the maximum number of states for a given NqN_q, which is (Nq+33)\binom{N_q+3}{3}, where 3 is the number of modes for water. Note that the enlarged spaces exceed their respective counts because the enlargement was performed, for all NqN_q (in the qubit mapping), using NQUANT=12 in ViBra.

Thus, the quantum-facing routine generates the initial seed configurations, while ViBRA can either analyze that seed directly or enlarge it through classical EN-PT2-guided selection. The EN-PT2 quantities are used only to rank additional configurations; they are not added as perturbative corrections to the reported energies, which are eigenvalues obtained by variational rediagonalization of the final projected Hamiltonian.

For Nq=2N_q = 2–4, the sampling procedure retained every configuration in the corresponding VCI space. Those calculations therefore do not test configuration filtering and are omitted from the following discussion. At Nq=5N_q = 5, 51 of the 56 available configurations were retained, and we consequently discuss the results from Nq=5N_q = 5 onward.

Table 6 shows that the absolute size of the sampled seed is nonmonotonic with NqN_q: the directly sampled lists contain between 51 and 82 configurations. Relative to the available space, however, the retained fraction decreases from 51/56 (91.1%) at Nq=5N_q = 5 to 53/455 (11.6%) at

Nq=12N_q = 12. Increasing the sampling cutoff therefore does not, under the present sampling protocol, guarantee a larger seed or a more complete representation of a target eigenstate. The identity and connectivity of the sampled configurations are at least as important as their number.

To assess the physical quality of the resulting spaces, we considered the ground state together with the nine states whose dominant reference configurations span the fundamentals, overtones, and combination bands through two quanta. Figure 5 reports the signed total-energy deviation. For each indicated reference configuration, we identified the projected-space root to which that configuration contributes the largest wavefunction coefficient.

Figure 5. Signed energy deviation, for the water molecule as a function of the maximum total quanta Nq=5N_q = 5–1212 used to generate the quantum-sampled seed. Panel (a) shows the ground state; panels (b)–(j) show the nine states dominated by the indicated configurations. MAXSCI=0 denotes direct variational diagonalization of the sampled seed, while MAXSCI=1–3 denote EN-PT2-guided enlargement of that seed by ViBra followed by variational rediagonalization. EN-PT2 is used only for configuration ranking; no perturbative correction is added to the reported energies.

Direct diagonalization of the quantum seed reproduces the zero-point energy to approximately 0.1 cm−1^{-1} across the cutoff range and generally performs well for the low-energy one-quantum states. Larger and more variable deviations occur for the two-quantum manifold. In particular, the direct-space deviations reach approximately 70–80 cm−1^{-1} for the configurations shown in Figure 5i and Figure 5j. These large deviations persist at both the lower and upper ends of the sampling-cutoff range.

The agreement with the VCI benchmarks improves markedly after EN-PT2-guided configuration-space enlargement by ViBra, at the cost of larger projected Hamiltonians and correspondingly longer calculations. Even in the worst case, the MAXSCI=3 list (163 of 455 states) comprises only 36% of the full VCI(12) space, a substantial reduction. Already at MAXSCI=1, errors that previously exceeded 80 cm−1^{-1} drop below 5 cm−1^{-1} for all ten states considered. Moving from MAXSCI=2 to 3 tightens this further, to roughly 1 cm−1^{-1}, showing an excellent agreement with the reference. We note that Figure 5h is the only state for which the Nq=5N_q = 5 and 12 results differ significantly once NqN_q reaches 12, with the energy discrepancy dropping from roughly 30 to 10 cm−1^{-1}. This might suggest that the unenlarged space is of higher quality at larger quanta limits; however, this is not observed as a general trend, and no broader conclusion should be drawn from this single case.

An important question is whether the improved energetic agreement alone is sufficient to validate the SKQD+PT2 approach. Figure 6 is instructive here. The wavefunction composition analysis in Figure 6 helps to elucidate a classic truncated selected CI problem. In the VCI(12) eigenvector, (0, 2, 0) is dominant and (0, 1, 2) is its largest secondary contribution. In the directly sampled space, (5, 0, 0) instead becomes the largest secondary contribution. Even with MAXSCI=1, it was sufficient to recover the composition of the wavefunction. It is important to note that the state (0, 1, 2) is present in all calculations, spanning MAXSCI from 0 to 3. This does not mean that ViBra evaluates the retained off-diagonal Hamiltonian elements incorrectly. Rather, the incomplete configuration space omits contributions needed to reproduce the effective mixing pattern of the full calculation. With MAXSCI=1–3, the dominant coefficients closely recover the VCI(12) composition.

Figure 6. Absolute dominant configuration coefficients for the (0,2,0)(0, 2, 0) state at sampling cutoff Nq=8N_q = 8. Panel (a) shows the full VCI(12) reference; panels (b-d) show the ViBra-enlarged spaces with MAXSCI=1 to 3, respectively. (e) shows direct diagonalization of the quantum-sampled seed with MAXSCI=0, highlighting in red the wrong coupling of states. The recovery of the reference composition at MAXSCI=1–3 illustrates the effect of classical EN-PT2-guided configuration-space enlargement after quantum sampling.

The immediate conclusion is that the unenlarged quantum-sampled seed should be used with caution: while it may yield deceptively favorable energies, it can also introduce spurious state mixing or apparent resonances. We would also expect an incomplete space to underrepresent peak splitting where it should occur. We therefore emphasize the value of inspecting wavefunction composition, peak intensities, and splittings alongside the energies.

These results establish the intended role of ViBra in a hybrid quantum–classical vibrational workflow. The simulated quantum-facing routine supplies an initial set of vibrational configurations, SQS_Q. With MAXSCI 0 list, ViBra constructs and diagonalizes the projected Hamiltonian directly in this externally sampled space. With MAXSCI>0>0 in list-reference mode, the same sampled set is used as the reference for ViBRA’s EN-PT2-guided configuration selection, producing a quantum-seeded but classically enlarged subspace. In both cases, the reported energies and eigenvectors are obtained through variational rediagonalization in the final configuration space.

The present H₂O calculations are intended as workflow and interface validation. They show that quantum-generated seed configurations can be decoded, supplied to ViBRA, analyzed directly, or systematically enlarged through its selected-VCI machinery. All quantum-facing calculations were performed using a simulator, and no execution on quantum hardware was undertaken. Accordingly, these results do not constitute a benchmark of SQD or SKQD, a study of shot complexity or scaling, or a claim of quantum advantage. Future work should examine configuration-recovery strategies, [24, 41, 42, 43] alternative SQD/SKQD sampling protocols, circuit and fragmentation optimization, [29] alternative initial states, hardware execution, shot requirements, scaling with molecular size and spectral congestion, and the conditions under which a quantum advantage might eventually become possible.

Conclusion

The ViBra package delivers a complete, efficient, and user friendly toolchain for anharmonic vibrational spectroscopy at the quartic force field level. By integrating configuration selection via S-VCI with symmetry exploitation via SA-VCI, it extends the reach of accurate VCI calculations to larger molecular systems without sacrificing physical rigour. The inclusion of second order dipole derivatives guarantees reliable intensities across the full vibrational spectrum, while the accompanying GUI and viewer make the results immediately interpretable and directly comparable with experiment. The numerical correctness of ViBra was established through a four-tier benchmark. The VSCF and full VCI modules were validated against Crystal23 for water, achieving agreement within 0.001 cm−10.001\ \mathrm{cm}^{-1} and 0.01 cm−10.01\ \mathrm{cm}^{-1}, respectively. The SA-VCI and S-VCI modules were validated against ViBra’s own full VCI calculations across a test set of five molecular point groups; SA-VCI reproduces full VCI to numerical precision, while S-VCI converges variationally to sub 5 cm−1\mathrm{cm}^{-1} accuracy at substantially reduced cost. Taken together, these benchmarks confirm that ViBra is numerically correct and physically faithful.

The selected-VCI infrastructure in ViBra also provides a natural interface to quantum-centric algorithms for anharmonic vibrational spectroscopy. In such workflows, a quantum-facing routine supplies samples that define a compact seed space, while ViBra either analyzes that space directly or enlarges it through EN-PT2-guided configuration selection before constructing and diagonalizing the projected Hamiltonian and evaluating transition observables. The H2OH_{2}O proof of concept validates this two-stage interface by combining qubit mapping, Hamiltonian fragmentation, and simulated time-evolution sampling with ViBra configuration-space refinement and projected diagonalization. This simulator-based demonstration does not constitute a hardware benchmark or imply quantum advantage, but it serves as motivation for ViBra and its capabilities, as well as for the development of future quantum-centric workflows focused on vibrational problems.

In terms of implementation, there are several natural directions in which the current framework can be extended. Although the current focus is on quartic force fields, the code architecture is general and higher order expansions of the potential energy surface can be added straightforwardly. Similarly, the method can in principle be pushed toward still larger systems, though such applications have not yet been tested. On the algorithmic side, the current SA-VCI implementation is restricted to Abelian point groups with one dimensional irreducible representations, and extending it to non-Abelian groups with degenerate irreps is planned for a future release.

Moreover, a key bottleneck in VCI calculations is the dramatic growth of the Hamiltonian matrix dimension with the number of modes and excitation quanta, which rapidly makes memory the limiting factor. Because the Hamiltonian matrix becomes increasingly sparse as the vibrational space grows, on the fly diagonalizers that exploit this structure could reduce the memory footprint substantially. At present we use full dense matrix diagonalization via the DSYEVR and DSYEVD subroutines from intel-lapack MKL. We retain this choice for now owing to the outstanding performance of these routines, even though it is not the optimal strategy for the sparse matrices generated by our pair list approach. Implementing on the fly diagonalizers adapted to this sparsity pattern is a natural next step.

Acknowledgement

The authors thank Rodrigo Neumann (IBM), Sumathy Raman (EMTech), Ramon Cardias (CBPF), Ivan S. Oliveira (CBPF) and Gustavo Petronilo (CBPF) for support. This research was supported by ExxonMobil Technology and Engineering Company. Any opinions, findings, conclusions, or recommendations are those of the authors and not of the sponsors.

Code Availability

The ViBra source code, documentation, and example input files are available at Ref. 44 The external proof-of-concept routines used for the quantum vibrational spectroscopy workflow, including the modal-to-qubit mapping, Pauli-operator construction, Hamiltonian fragmentation, Trotterized time-evolution, and simulator-based sampling scripts, are available at Refs. [39, 40] The calculations reported in § 6 were performed using the code versions identified by the repository release tags or commit hashes reported in the corresponding references.

Data Availability

The numerical data necessary to reproduce the validation tests, benchmark calculations, tables, and figures reported in this work are provided in a separate Zenodo data archive. [35] The archive includes the quartic force-field inputs, ViBra input files, reference spectra, selected VCI and symmetry-adapted VCI benchmark outputs, timing data, selected configuration lists, and processed numerical data used to generate the figures and tables. Data associated with the exploratory H2_2O quantum-vibrational proof-of-concept calculation, including sampled vibrational configuration spaces and spectra generated from the external mapping, fragmentation, and time-evolution routines, are also included. Source code and analysis scripts are distributed separately, as described in the Code Availability statement.

Supporting Information Available

A comprehensive manual of ViBra is provided as Supporting Information. It contains the detailed derivation of all equations, detailed descriptions of every Fortran module and Python class, a complete input-keyword reference, and an explanation of the output file formats.

References

TOC Graphic

ViBra logo

Figure 7 (PDF p. 55). ViBra – a program for VSCF/VCI. The logo is inspired on the sunset at Morro Dois Irmãos, a classic landscape postcard from Rio de Janeiro, resembling two overlapping Gaussian peaks.

References

  1. [1]Barone, V. Anharmonic vibrational properties by a fully automated second-order perturbative approach. J. Chem. Phys. 2005, 122, 014108.pubmed.ncbi.nlm.nih.gov/15638643
  2. [2]Christiansen, O. Selected new developments in vibrational structure theory: potential con- struction and vibrational wave function calculations. Phys. Chem. Chem. Phys. 2007, 9, 2942–2953.
  3. [3]Bloino, J. A VPT2 route to near-infrared spectroscopy: The role of mechanical and electrical anharmonicity. J. Phys. Chem. A 2015, 119, 5269.DOI
  4. [4]Neese, F.; Wennmohs, F.; Becker, U.; Riplinger, C. The ORCA quantum chemistry program package. J. Chem. Phys. 2020, 152, 224108.DOI
  5. [5]Neese, F. Software update: The ORCA program system — Version 5.0. WIREs Comput. Mol. Sci. 2022, 12, e1606.DOI
  6. [6]Bowman, J. M. Self-consistent field energies and wavefunctions for coupled oscillators. J. Chem. Phys. 1978, 68, 608.DOI
  7. [7]Gerber, R. B.; Ratner, M. A. A semiclassical self-consistent field (SC SCF) approximation for eigenvalues of coupled-vibration systems. Chem. Phys. Lett. 1979, 68, 195–198.DOI
  8. [8]Christiansen, O. Vibrational coupled cluster theory. J. Chem. Phys. 2004, 120, 2149.pubmed.ncbi.nlm.nih.gov/15268353
  9. [9]Bowman, J. M.; Carrington, T.; Meyer, H.-D. Variational quantum approaches for computing vibrational energies of polyatomic molecules. Mol. Phys. 2008, 106, 2145–2182.DOI
  10. [10]Rauhut, G. Configuration selection as a route towards efficient vibrational configuration interaction calculations. J. Chem. Phys. 2007, 127, 184109.pubmed.ncbi.nlm.nih.gov/18020632
  11. [11]Neff, M.; Rauhut, G. Toward large scale vibrational configuration interaction calculations. J. Chem. Phys. 2009, 131, 124129.DOI
  12. [12]Mathea, T.; Rauhut, G. Advances in vibrational configuration interaction theory — part 1: Efficient calculation of vibrational angular momentum terms. J. Comput. Chem. 2021, 42, 2321–2330.DOI
  13. [13]Mathea, T.; Petrenko, T.; Rauhut, G. Advances in vibrational configuration interaction theory — part 2: Fast screening of the correlation space. J. Comput. Chem. 2022, 43, 6–18.DOI
  14. [14]Erba, A.; Maul, J.; Ferrabone, M.; Dovesi, R.; Rérat, M.; Carbonnière, P. Anharmonic vibrational states of solids from DFT calculations. Part II: Implementation of the VSCF and VCI methods. J. Chem. Theory Comput. 2019, 15, 3766–3777.DOI
  15. [15]Erba, A.; Maul, J.; Ferrabone, M.; Carbonnière, P.; Rérat, M.; Dovesi, R. Anharmonic Vibrational States of Solids from DFT Calculations. Part I: Description of the Potential Energy Surface. J. Chem. Theory Comput. 2019, 15, 3755–3765.DOI
  16. [16]Carbonnière, P.; Erba, A.; Richter, F.; Dovesi, R.; Rérat, M. Calculation of anharmonic IR and Raman intensities for periodic systems from DFT calculations: Implementation and validation. J. Chem. Theory Comput. 2020, 16, 3343–3351.DOI
  17. [17]Sibaev, M.; Crittenden, D. L. PyVCI: A flexible open-source code for calculating accurate molecular infrared spectra. Comput. Phys. Commun. 2016, 203, 290–300.
  18. [18]Garnier, R.; Odunlami, M.; Bris, V. L.; Bégué, D.; Baraille, I.; Coulaud, O. Dual vibration configuration interaction (DVCI). An efficient factorization of molecular Hamiltonian for high performance infrared spectrum computation. Comput. Phys. Commun. 2019, 234, 263–273.
  19. [19]Garnier, R.; Odunlami, M.; Le Bris, V.; Bégué, D.; Baraille, I.; Coulaud, O. Adaptive vibrational configuration interaction (A-VCI): <i>A posteriori</i> error estimation to efficiently compute anharmonic IR spectra. J. Chem. Phys. 2016, 144.DOI
  20. [20]Odunlami, M.; Le Bris, V.; Bégué, D.; Baraille, I.; Coulaud, O. A-VCI: A flexible method to efficiently compute vibrational spectra. J. Chem. Phys. 2017, 146.DOI
  21. [21]Fetherolf, J. H.; Berkelbach, T. C. Vibrational heat-bath configuration interaction. J. Chem. Phys. 2021, 154, 074104.DOI
  22. [22]Bhatty, A. U.; Brorsen, K. R. An alternative formulation of vibrational heat-bath configuration interaction. Mol. Phys. 2021, 119, e1936250.
  23. [23]Tran, H. K.; Berkelbach, T. C. Vibrational heat-bath configuration interaction with semistochastic perturbation theory using harmonic oscillator or VSCF models. J. Chem. Phys. 2023, 159, 194101.arxiv.org/abs/2307.13246
  24. [24]Robledo-Moreno, J.; Motta, M.; Haas, H.; Javadi-Abhari, A.; Jurcevic, P.; Kirby, W.; Martiel, S.; Sharma, K.; Sharma, S.; Shirakawa, T.; Sitdikov, I.; Sun, R.-Y.; Sung, K. J.; Takita, M.; Tran, M. C.; Yunoki, S.; Mezzacapo, A. Chemistry Beyond the Scale of Exact Diagonalization on a Quantum-Centric Supercomputer. Science Advances 2025, 11, eadu9991.DOI
  25. [25]Yu, J.; Robledo Moreno, J.; Iosue, J. T.; Bertels, L.; Claudino, D.; Fuller, B.; Groszkowski, P.; Humble, T. S.; Jurcevic, P.; Kirby, W.; Maier, T. A.; Motta, M.; Pokharel, B.; Seif, A.; Shehata, A.; Sung, K. J.; Tran, M. C.; Tripathi, V.; Mezzacapo, A.; Sharma, K. Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization. arXiv preprint arXiv:2501.09702 2025,arxiv.org/abs/2501.09702
  26. [26]Alexeev, Y.; Amsler, M.; Barroca, M. A.; Bassini, S.; Battelle, T.; Camps, D.; Casanova, D.; Choi, Y. J.; Chong, F. T.; Chung, C.; Codella, C.; Córcoles, A. D.; Cruise, J.; Di Meglio, A.; Duran, I.; Eckl, T.; Economou, S.; Eidenbenz, S.; Elmegreen, B.; others Quantum-Centric Supercomputing for Materials Science: A Perspective on Challenges and Future Directions. Future Generation Computer Systems 2024, 160, 666–710.
  27. [27]Sawaya, N. P. D.; Huh, J. Quantum Algorithm for Calculating Molecular Vibrionic Spectra. The Journal of Physical Chemistry Letters 2019, 10, 3586–3591.DOI
  28. [28]Ollitrault, P. J.; Baiardi, A.; Reiher, M.; Tavernelli, I. Hardware Efficient Quantum Algorithms for Vibrational Structure Calculations. Chemical Science 2020, 11, 6842–6855.DOI
  29. [29]Malpathak, S.; Das Kallullathil, S.; Loaiza, I.; Fomichev, S.; Arrazola, J. M.; Izmaylov, A. F. Trotter Simulation of Vibrational Hamiltonians on a Quantum Computer. arXiv preprint arXiv:2508.11865 2025,arxiv.org/abs/2508.11865
  30. [30]Light, J. C.; Carrington, T. In Adv. Chem. Phys.; Prigogine, I., Rice, S. A., Eds.; John Wiley & Sons, Ltd.: Hoboken, NJ, 2000; Vol. 114; pp 263–310.
  31. [31]Epstein, P. S. The Stark effect from the point of view of Schroedinger’s quantum theory. Phys. Rev. 1926, 28, 695–710.DOI
  32. [32]Nesbet, R. K. Configuration interaction in orbital theories. Proc. R. Soc. London, Ser. A 1955, 230, 312–321.DOI
  33. [33]Coblentz Society, Inc. Infrared Spectrum of Ethylene (C2H4). Coblentz Collection, COBLENTZ NO. 8869. Digitized by NIST from hard copy. Dow Chemical Company, spectrometer with KBr foreprism., 1964; CAS Registry No. 74-85-1. Gas (150 mmHg diluted to a total pressure of 600 mmHg with N2), path length 5 cm, resolution 4 cm-1. XUNITS = 1/cm, YUNITS = transmittance. FIRSTX = 454.005, LASTX = 3795.69, NPOINTS = 3561.
  34. [34]Eilers, P. H. C. A perfect smoother. Anal. Chem. 2003, 75, 3631–3636.pubmed.ncbi.nlm.nih.gov/14570219
  35. [35]Ligorio, R.; Duriez, A. C.; Barroca, M. A. Numerical Data for ViBra: A Selected and Symmetry-Adapted VCI Platform for Anharmonic Vibrational Spectroscopy and Quantum-Sampled Configuration Spaces. Zenodo, 2026; https://doi.org/10.5281/zenodo.21537182, quartic force-field inputs, ViBrA input files, benchmark outputs, spectra, timing data, selected configuration lists, and numerical data used to generate the tables and figures.
  36. [36]Erba, A.; Mitoli, D. Modelling Phonon Couplings: Anharmonic Lattice Dynamics. CRYS-TAL Tutorials Project, edited by B. Civalleri and C. Roetti, https://tutorials.crysta lsolutions.eu/tutorial.html?td=anharmonicity&tf=anharm, Accessed: 2026-05-26.
  37. [37]National Institute of Standards and Technology NIST Computational Chemistry Com-parison and Benchmark Database, NIST Standard Reference Database 101, Release 22. https://cccbdb.nist.gov, 2022; Accessed: 2026-05-26.
  38. [38]Kanno, K.; Kohda, M.; Imai, R.; Koh, S.; Mitarai, K.; Mizukami, W.; Nakagawa, Y. O. Quantum-selected configuration interaction: Classical diagonalization of Hamiltonians in subspaces selected by quantum computers. Physical Review Research 2026, 8.DOI
  39. [39]Barroca, M. A. Ham-Vib. https://github.com/MarcoBarroca/Ham_vib, 2026; Research prototype used for the H₂O quantum vibrational hamiltonian mapping; commit 2a66a37, accessed 22 July 2026.
  40. [40]Duriez, A. C. AlgorithmQ. https://github.com/ACDuriez/algorithmq, 2026; Research prototype implementing the fragmentation and time-evolution routines used in the H₂O proof of concept; commit 3e5e6bc, accessed 22 July 2026.
  41. [41]Barison, S.; Robledo Moreno, J.; Motta, M. Quantum-centric computation of molecular excited states with extended sample-based quantum diagonalization. Quantum Science and Technology 2025, 10, 025034.DOI
  42. [42]Barroca, M. A.; Gujarati, T. P.; Sharma, V.; Neumann Barros Ferreira, R.; Na, Y.-H.; Giammona, M.; Mezzacapo, A.; Wunsch, B.; Steiner, M. B. Scaling Active Spaces in Sim-ulations of Surface Reactions through Sample-Based Quantum Diagonalization. Scientific Reports 2026,DOI
  43. [43]Duriez, A.; Carvalho, P. C.; Barroca, M. A.; Zipoli, F.; Jaderberg, B.; Neumann Bar-ros Ferreira, R.; Sharma, K.; Mezzacapo, A.; Wunsch, B.; Steiner, M. Computing Band Gaps of Periodic Materials via Sample-Based Quantum Diagonalization. npj Computational Materials 2026,DOI

Paper details

vlnc:W_tbs6h89r
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents