betaFeedback

WorksPhysics and Astronomy

Distributed Quantum-Centric Supercomputing with Multi-Layered Generative Learning for Intermolecular Interactions

Abstract

Quantum-centric supercomputing (QCSC) approaches, such as sample-based quantum diagonalization (SQD), combine quantum processors with high-performance computers (HPC), offering a promising route toward solving challenging electronic structure problems. However, implementation of QCSC approaches to large active space calculations is constrained by the associated number of qubits, deep quantum circuits, sampling cost, and large-scale Hamiltonian diagonalization. To overcome these limitations, we introduce a distributed QCSC framework that decomposes complex molecular problems into parallel sub-problems requiring fewer qubits, shallower circuits, and reduced shot budgets. A physically motivated, two-step aggregation scheme then reconstructs the global solution: first, a computational bit assignment strategy optimizes a cost function to appropriately map local bit indices to global ones, yielding a mapped tensor product state (MTPS); second, an inter-fragment correlation recovery module restores missing cross-fragment interactions. Herein, we employ a multi-layered generative machine learning architecture, starting with utilizing the recently proposed physics-informed generative SQD (PIGen-SQD) as a distributed QCSC solver, followed by a secondary generative module to model configuration distributions and identify dominant sectors of the global Hilbert space. For water and formamide dimers, our method achieves up to 95%95\% reduction in the diagonalization dimension while enhancing energy accuracy by an order of magnitude over conventional SQD. Furthermore, application to the water hexamer ((H2O)6\mathrm{(H_2O)_6}) cage successfully decomposes a 120-qubit problem into six 20-qubit monomeric subsystems, and stitches the monomers under a dimeric approximation, demonstrating excellent agreement with the reference classical methods.

Introduction

Accurate quantum-mechanical treatment of larger molecular systems with nontrivial electronic correlation effects is beyond the scope of classical mean-field theories and requires sophisticated high-level classical treatments with theories like coupled-cluster (CC)[1, 2, 3], configuration-interaction (CI)[4], and the density matrix renormalization group (DMRG)[5]. However, the computational cost of such methods scales steeply with system size and thus they are not scalable. Quantum computing algorithms show immense promise to efficiently and accurately solve such electronic structure problems. Nevertheless, hardware noise in currently available quantum processing units (QPUs) affects computations, leaving quantum algorithms like quantum phase estimation (QPE) infeasible. For near-term applications, the community relies upon hybrid quantum-classical approaches such as the variational quantum eigensolver (VQE) [6, 7, 8], projective quantum eigensolver (PQE) [9, 10, 11], or the more recent quantum-centric supercomputing (QCSC)/ quantum selected configuration interaction (QSCI) frameworks [12]. VQE- or PQE-like algorithms suffer from impractically long parameter optimization cycles and local traps [7, 13] in the cost-function landscapes, and active research is ongoing in this field to mitigate the effects of such challenges [10, 11, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34]. QCSC takes a different approach by performing state preparation and measurement to obtain electronic configuration samples from quantum hardware, followed by diagonalization on a classical processor to compute the desired observable.

In the context of quantum chemistry, QCSC is a class of selected configuration interaction (SCI) problem that leverages a quantum computer as a sampling engine to identify the dominant sector of the Hilbert space that contain the solution of the Schrödinger equation. However, present day quantum processors suffer from noise due to the interaction with the environment and often produce large number of error-prone samples that affect the computation of accurate energetics. The recently proposed sample-based quantum diagonalization (SQD)[36] addresses this issues by introducing a self-consistent configuration recovery (S-CORE) protocol to denoise the error-prone samples based on the average orbital occupancy and it subsequently expands the subspace to include configurations from the unexplored sector of the Hilbert space followed by diagonalization.

The efficiency and accuracy of any generic configuration recovery depend on hardware-generated sample quality, which is governed by hardware noise, quantum-state expressivity, and the available shot budget. For a system with MM orbitals and PP particles, the configuration space scales as (MP)\binom{M}{P}, implying combinatorial growth with system size. Even under ideal noiseless conditions, statistically meaningful sampling according to the Born’s interpretation may require number of shots that roughly scales as O((MP))O\left(\binom{M}{P}\right), limiting the scalability of sample-based quantum algorithms under restricted circuit-execution budgets. Furthermore, hardware noise, the challenges associated with preparing highly expressive quantum states, along with restricted measurement budgets collectively increase the computational burden on classical high-performance computing (HPC) resources within the QCSC workflow. These limitations motivate a fundamental question: can we avoid performing quantum state preparation and sampling for the entire system by partitioning it into chemically relevant subsystems that retain nontrivial electronic correlation effects? Such an approach would enable the application of distributed QCSC solvers to individual sub-problems, followed by classical post-processing to aggregate the sub-systems for accurate solution of the super-system. Specifically, partitioning an MM-qubit system into NN fragments, with MkM_k qubits and PkP_k electrons in kk-th fragment, reduces the sampling cost to approximately O((MkPk))O\left(\binom{M_k}{P_k}\right) per fragment. Fragmentation also reduces noisy entangling-gate counts and circuit depth, thereby enhancing the quality of samples per fragment. Toward this, there are many recent quantum-classical hybrid algorithmic developments that focus on combining QCSC, particularly SQD with existing classical embedding strategies to treat extended systems such as SQD with density matrix embedding theory (DMET-SQD)[37, 38], SQD with localized active space approximation (LASSQD)[39], and several other works[40, 41, 42]. The above-mentioned methods uses SQD as an active-space sub-system solver and are recombined by embedding them within each other’s environment at a lower level of theory.

In this work, we take a complementary approach to enable larger active space QCSC simulations even with limited shot budget and reduced diagonalization cost. Here we propose a distributed QCSC approach that involves decomposition of an extended molecular system into chemically motivated fragments that are solved using parallel fragment QCSC solvers followed by a physically motivated aggregation step. To this end, the recently developed physics-informed generative machine-learning driven SQD (PIGen-SQD)[43] is explored as the parallel fragment QCSC solver which proposes a concerted use of quantum hardware samples, low order perturbation theory and generative machine learning (Gen-ML) to explore the dominant sectors of the Hilbert space. PIGen-SQD has been demonstrated to provide extremely accurate description of the electronic strong correlation effects with the associated diagonalization performed in drastically compact subspaces compared to SQD. While distributed PIGen-SQD solvers provide accurate description of the intra-fragment correlations, the solution of the super-system requires accurate integration of these fragment solutions along with the recovery of inter-fragment correlations. This is done with a two-fold aggregation stage: first, a product state construction from the dominant fragment configurations is performed, which is assumed to provide non-trivial support to the eigenstate of the super-system Hamiltonian. However, the product state requires appropriate one-to-one assignment of the local computational bits to that of the super-system such that dominant local configurations are properly mapped to the canonical fock space. Toward this, we have proposed a computational bit mapping strategy here which solves a linear assignment problem (LAP) for the local-to-global computational bit correspondence. With this mapping strategy, one can construct a mapped tensor product state (MTPS) directly from the local QCSC solutions. MTPS provides a good initialization to the second stage of aggregation which involves an inter-fragment correlation recovery stage. In this paper we have explored an additional layer of Gen-ML to entangle two or more subsystems and reconstruct the bitstrings in the nn-meric computational basis. The algorithm is referred to as PIGen-SQD with mapped aggregation of partitioned subspace or PIGen-SQD(MAPS). Additionally, to provide a fair comparison, we have considered the traditional s-C0RE mechanism initialized with MTPS for the inter-fragment recovery as well. PIGen-SQD(MAPS) provides a few key advantages over traditional SQD performed over the entire system: (1) it can be performed with limited shot budget since quantum circuits are only run over fragments, (2) the diagonalization dimension is significantly low while energy accuracy with the reference is substantially better in comparison, and (3) depending on the complexity of the problem and the computational resources available, one can choose the number of monomeric fragments (nn) that needs to be entangled explicitly during the inter-fragment correlation recovery.

Here we have used PIGen-SQD(MAPS) to study non-covalent inter-molecular interactions which plays critical roles in biological processes such as protein-folding, membrane assembly, and is absolutely essential for drug discovery pipelines. Particularly, we have studied water dimer, which is the prototypical system for studying hydrogen bonds, and formamide (FMA) dimer which is biochemically relevant as the smallest system containing the peptide bond. Our numerical studies on the water and formamide dimers demonstrate that PIGen-SQD(MAPS) achieves up to 95% reduction in the diagonalization space and an order-of-magnitude improvement in energy accuracy, while requiring only a fraction of the shots compared to conventional SQD performed on the dimeric system. We further benchmark PIGen-SQD(MAPS) on cage-shaped (H2_2O)6_6 using a fragment-pair (n=2n=2) approximation, partitioning the 120-qubit active-space problem into subsystems requiring hardware samples from 20-qubit quantum circuits. Inter-fragment correlations are subsequently recovered at the dimeric supersystem level, followed by a many-body expansion to estimate the total energy of the hexamer. Importantly, in our algorithm, the quantum hardware is employed to sample only for the monomeric units that significantly improves the probability of obtaining high quality samples with limited shot budget.

Treatment of Extended Systems within QCSC Framework: General Considerations

QCSC algorithms such as SQD, proposes a quantum hardware as an efficient engine to sample from dominant sector of the Hilbert space for accurate solution of the fermionic Hamiltonian

H^=∑pqhpqa^p†a^q+∑pqrsvpqrsa^p†a^q†a^sa^r(1)\hat{H} = \sum_{pq} h_{pq}\hat{a}^{\dagger}_{p}\hat{a}_{q} + \sum_{pqrs} v_{pqrs}\hat{a}^{\dagger}_{p}\hat{a}^{\dagger}_{q}\hat{a}_{s}\hat{a}_{r} \tag*{(1)}

where, p,q,r,sp, q, r, s are molecular orbital indices, hpqh_{pq} and vpqrsv_{pqrs} are respectively the one and two electron integrals in the molecular orbital basis and a^p†,a^q\hat{a}^{\dagger}_{p},\hat{a}_{q} are the creation and annihilation operators. This is achieved by mapping the fermionic electronic-structure problem into the quantum-computing framework by encoding the MM molecular spinorbitals to MM qubits, so that each electronic configuration corresponds to a computational basis state (or configuration) ∣ϕ⟩=∣xM…x2x1⟩|\phi\rangle= |x_M \ldots x_2x_1\rangle with (xM…x2x1)∈{0,1}M(x_M \ldots x_2x_1) \in\{0,1\}^{M}, where the single-qubit states ∣0⟩|0\rangle and ∣1⟩|1\rangle represent unoccupied and occupied spinorbitals respectively. The Jordan-Wigner [44] transformation maps the second-quantized fermionic operators to qubit operators. A trial state on MM qubits is prepared by a physically motivated ansatz which is expressed by a parameterized quantum circuit (PQC). Ideally the state preparation should be done in a way that it has a substantial overlap with the target state. In this work, we have used the local unitary cluster jastrow (LUCJ) [45] ansatz in its single-layer form,

∣Ψ⟩LUCJ=e−K^2eK^1eiJ^1e−K^1∣Φ0⟩(2)|\Psi\rangle_{\mathrm{LUCJ}} = e^{-\hat{K}_{2}}e^{\hat{K}_{1}}e^{i\hat{J}_{1}}e^{-\hat{K}_{1}}|\Phi_{0}\rangle \tag*{(2)}

where K^1\hat{K}_{1} and K^2\hat{K}_{2} are one-body operators and J^1=∑pσrτJpσ,rτn^pσn^rτ\hat{J}_{1} = \sum_{p\sigma r\tau} J_{p\sigma,r\tau}\hat{n}_{p\sigma}\hat{n}_{r\tau} is a density-density operator with n^\hat{n} being the number operator, and ∣Φ0⟩|\Phi_{0}\rangle is the Hartree-Fock (HF) configuration. Repeated measurement of this circuit in the computational basis produces the distribution of outcomes

X~={ϕ∣ϕ∼P~(ϕ)}(3)\tilde{\mathcal{X}} = \{\phi\mid\phi\sim\tilde{P}(\phi)\} \tag*{(3)}

with the associated noisy probabilities

P~(ϕ)=NϕNshots(4)\tilde{P}(\phi) = \frac{N_{\phi}}{N_{\mathrm{shots}}} \tag*{(4)}

where NϕN_{\phi} is the frequency of configuration ∣ϕ⟩|\phi\rangle and NshotsN_{\mathrm{shots}} is the total number of shots. Due to the noise present in present-day quantum processors, the sampled distribution is contaminated with particle number and zz component of spin (S^z\hat{S}_{z}) non-conserving configurations and requires subsequent classical post-processing to correct the noise-affected distribution. The original SQD formulation involves the S-CORE post-processing module that is based on average occupancy of the associated spinorbitals and selected CI loops to identify dominant configurations for the Hamiltonian diagonalization. The performance of the S-CORE module, in terms of both computational efficiency and accuracy, depends heavily on the quality of the hardware-generated samples, which is influenced by the prevailing hardware noise, the expressivity of the prepared quantum state, and the available shot budget. These parameters pose significant challenges to extend QCSC algorithms to extended systems of practical interest.

From a quantitative point of view, the Schrödinger’s equation of an extended system (with MM orbitals and PP electrons) has the general CI solution

∣Ψ⟩=∑gcg∣Φg⟩(5)|\Psi\rangle= \sum_{g} c_{g}|\Phi_{g}\rangle \tag*{(5)}

where, {∣Φg⟩}∈{0,1}M\{|\Phi_{g}\rangle\} \in\{0,1\}^{M} are computational bases and cgc_{g} are the associated coefficients in the linear expansion. Solving the corresponding QCSC problem involves PQC with MM qubits (assuming one-to-one spinorbital-to-qubit mapping), O(M2)\mathcal{O}(M^{2}) two qubit gate complexity per layer (considering UCJ circuit) [46], and at least O((MP))\mathcal{O}\left(\binom{M}{P}\right) shot budget for relevant sampling statistics, which scales unfavorably with system size. Alternatively, a more tractable approach would be to partition the entire Hamiltonian into

H^=H^(0)+H^′(6)\hat{H} = \hat{H}^{(0)} + \hat{H}' \tag*{(6)}

where, H^(0)\hat{H}^{(0)} has an additively separable form as it contains only intra-fragment terms and mean-field inter-fragment effects, and is exactly solvable. The residual non-local inter-fragment interaction terms are considered into H^′\hat{H}'. In general, if the extended system can be partitioned into NN fragments, the zeroth order Hamiltonian has the generic form

H^(0)=∑k=1NH^k(0)(7)\hat{H}^{(0)} = \sum_{k=1}^{N} \hat{H}^{(0)}_{k} \tag*{(7)}

where kk is the fragment index. The local fragment Hamiltonian is defined as

H^k(0)=∑I∈kh^k(0)(I)+∑IJ∈kV^k(0)(IJ)+V^k(0)(eff)(8)\hat{H}^{(0)}_{k} = \sum_{I\in k}\hat{h}^{(0)}_{k}(I) + \sum_{IJ\in k}\hat{V}^{(0)}_{k}(IJ) + \hat{V}^{(0)(\mathrm{eff})}_{k} \tag*{(8)}

consisting of one-body (h^k(0)(I)\hat{h}^{(0)}_{k}(I)), two-body (V^k(0)(IJ)\hat{V}^{(0)}_{k}(IJ)) intra-fragment operators and an effective potential V^k(0)(eff)\hat{V}^{(0)}_{k}(\mathrm{eff}) that encodes the mean-field effects of all other fragments on kk-th fragment. The dimension of local Hilbert space Hk\mathcal{H}_k for kk-th fragment having MkM_k spin-orbitals and PkP_k number of particles is (MkPk)\binom{M_k}{P_k}. Assuming the fragmentation is done in such a way that the local Hilbert spaces are computationally challenging and involves non-trivial electronic strong correlation effects, one may solve each fragment with efficient QCSC solvers that provide highly accurate solutions for the fragments. The ground state of the fragment CI problems H^k(0)∣Ψk(0)⟩=Ek(0)∣Ψk(0)⟩\hat{H}^{(0)}_k\left|\Psi^{(0)}_k\right\rangle= E^{(0)}_k\left|\Psi^{(0)}_k\right\rangle obtained from QCSC solvers are represented as a linear combination

∣Ψk(0)⟩=∑αcαk(0)∣ϕαk(0)⟩(9)\left|\Psi^{(0)}_k\right\rangle= \sum_{\alpha} c^{(0)}_{\alpha k}\left|\phi^{(0)}_{\alpha k}\right\rangle \tag*{(9)}

where, ∣ϕαk(0)⟩∈{0,1}Mk\left|\phi^{(0)}_{\alpha k}\right\rangle\in\{0,1\}^{M_k} are the computational bases (configurations) of the kk-th fragment. This eigenvalue equation for each fragment can be solved by QCSC solvers in parallel QPU-HPC platforms to provide the associated eigenstate solutions which consist of a set of configurations and the associated CI coefficients. Under an nn-meric approximation with n≤Nn \leq N, the product-state solution of the zeroth order nn-meric Hamiltonian is

∣Ψ(0)⟩=∏k=1n∣Ψk(0)⟩(10)\left|\Psi^{(0)}\right\rangle= \prod_{k=1}^{n}\left|\Psi^{(0)}_k\right\rangle \tag*{(10)}

Note that this product-state solution is obtained by enforcing spin (zz component of the spin angular momentum S^z\hat{S}_z) and particle number conservation over each fragment as this constraint is necessary (at least in the zeroth order solution level) for efficient QCSC implementation.

At this stage one can start with the product-state solution ∣Ψ(0)⟩\left|\Psi^{(0)}\right\rangle (Eq. (10)) which neglects a substantial part of the inter-fragment correlation effects, and then treat the effects of H^′\hat{H}^{\prime} perturbatively to get the solution of the full extended system. The recently proposed FragPT2 method [47] takes such an approach in a classical computing platform, where the zeroth order solution is supplemented by a perturbative treatment of the inter-fragment correlation effects. While such a perturbative treatment is efficient and provides quantitative description of charge transfer, dispersion and triplet-triplet coupling effects, its accuracy might get limited due to the perturbative treatment of the inter-fragment effects. In order to obtain chemically meaningful accuracy, one may take a non-perturbative approach like (Full) CI over a suitably constructed nn-meric basis. This way, the quality of the solution primarily depends on the fragmentation scheme and not on the insufficient treatment of correlation. To proceed along this line within a QCSC/QSCI framework, one needs to project the nn-meric Hamiltonian onto a compact subspace of the corresponding nn-meric Hilbert space and solve the associated CI diagonalization problem. In such case, the computational bottleneck primarily comes from the large nn-meric diagonalization and one can account for only low order nn-meric approximation among the fragments. Such an approach presumes that the solutions of the local monomeric CI problems have substantial support to the solution of the composite nn-meric system. Consequently, one can construct a dominant subspace in the canonical Hilbert space directly from the information of the set of (dominant) local computational bases over various fragments. This assumption requires appropriate mapping of local computational basis originating from fragment QCSC solvers to the global spin-particle conserved fock space of the canonical system. This directly transfers the local fragment characteristics to the global Hilbert space. However, such mapping is a non-trivial task, particularly if there are near degenerate orbitals originating from different fragments, as it requires appropriate one-to-one assignment of local (fragment) molecular orbitals to the canonical molecular orbitals of the entire system. In this context, it is important to clarify that, the terms supersystem, canonical system, and global subspace are used interchangeably throughout this work to denote the nn-meric system, and the associated Hilbert space.

Here we address this by presenting a computational bit assignment strategy that leads to appropriate reconstruction of the global basis from local configurations. Since for chemical systems computational bits are directly associated with molecular spinorbitals, in this case the mapping is done by analyzing orbital overlap and spectral characters of super-system and sub-system orbitals. The assignment information are subsequently used to map the global basis from the dominant local basis and such a reconstruction is referred to as mapped tensor product state (MTPS). Computationally, MTPS forms a set of bitstrings that provides us with the optimal basis with appropriate distribution of occupancy and spin properties to initiate inter-fragment correlation recovery for the nn-meric super-system. One may note that the mapped product state, as discussed above, is structurally different from the direct product state ∣Ψ(0)⟩\left|\Psi^{(0)}\right\rangle in Eq. (10). With an accurately designed bit mapping, this modified product space forms a substantially compact subspace in the composite Hilbert space which is purely constructed based on the knowledge of the local CI solutions. Such an approach would drastically reduce the complexity of the QCSC problem if the fragmentation is done in a way such that (MkPk)≪(MP)\binom{M_k}{P_k} \ll\binom{M}{P} for all kk. This workflow enables one to solve a set of tractable monomeric QCSC problems and reconstruct a dominant subspace of the canonical Hilbert space that has non-trivial support to the target eigenroot.

In this aspect, it is important to mention that, there is no unique way to construct the monomeric Hamiltonian. Furthermore, the choice of the many-body expansion order nn may typically be user defined depending on the complexity of the system and the computational budget at disposal. In Section III, we will discuss the theoretical and algorithmic development that defines molecular fragmentation and effective Hamiltonian construction, parallel QCSC solvers, computational bit assignment strategy and MTPS construction, and inter-fragment correlation recovery protocols toward accurate treatment of extended chemical systems.

Problem Decomposition and Distributed QCSC Workflow

Hamiltonian Partitioning and Molecular Fragmentation

Molecular fragmentation provides a systematic route to reducing the computational cost of electronic-structure calculations by decomposing an extended molecular system into smaller, chemically meaningful subsystems. Among the established approaches, the fragment molecular orbital (FMO) method partitions a molecular system into NN fragments and evaluates fragment, fragment-pair, ... up to nn-meric fragment contributions to approximate the properties of the entire system [48]. Similarly, energy-based fragmentation methods express the total energy in terms of contributions from subsystems, often using a many-body expansion or related energy decomposition schemes [49]. These approaches offer a natural framework for quantum-centric supercomputing (QCSC), wherein individual fragments can be assigned to quantum-classical solvers and processed independently or in parallel, followed by classical reconstruction of the properties of the full system. For systems in which strong electronic correlation and fragment environment entanglement are important, density matrix embedding theory (DMET) provides a complementary strategy by embedding a fragment in an effective bath designed to capture its coupling to the surrounding system [50, 51]. Adapting these fragmentation and embedding principles to QCSC could enable resource-intensive quantum calculations to focus on the most chemically relevant subsystems, while treating interfragment interactions through an appropriate classical or embedding-based correction scheme.

In this paper, we have used FMO as the fragmentation tool, however, one may choose other relevant fragmentation mechanisms as well. FMO with nn-body approximation (known as FMO-nn) involves the following steps: (i) partitioning the entire system into NN fragments (monomers), (ii) monomeric calculation, dimeric calculation ⋯\cdots up to nn-meric calculation (where n≤Nn \le N) with a given method, and (iii) total property evaluation. For example, under the FMO-2-RHF approximation, a large molecular system is divided into individual fragments. The molecular orbitals (MOs) of each fragment are self consistently optimized in the field of external potential created by the surrounding ((N−1N-1)) fragments. The Hamiltonian for the monomer is given by,

H^k=∑i∈k[−12∇i2−∑AZA∣rI−RA∣+∑l≠k∫dr′ρl(r′)∣rI−r′∣]+∑l∈k∑I<J∈k1∣rI−rJ∣(11)\hat{H}_k = \sum_{i \in k} \left[ -\frac{1}{2}\nabla_i^2 - \sum_A \frac{Z_A}{\lvert r_I-R_A\rvert} + \sum_{l \ne k} \int\mathrm{d}r' \frac{\rho_l(r')}{\lvert r_I-r'\rvert} \right] + \sum_{l \in k} \sum_{I<J \in k} \frac{1}{\lvert r_I-r_J\rvert} \tag*{(11)}

to solve the monomeric problems: H^kΨk=EkΨk\hat{H}_k\Psi_k=E_k\Psi_k. Here, kk, ll and rIr_I, RAR_A represent different monomeric fragments and electronic and nuclear coordinates respectively, ∇I2\nabla_I^2 is the kinetic energy operator, ZAZ_A is the atomic number and ρl(r′)\rho_l(r') denotes the electronic density at position r′r' in ll-th fragment. Similarly the MOs of a fragment pair (dimer) are obtained self-consistently in the field produced by the leftover (N−2N-2) fragments. The associated Hamiltonian is given by,

H^kl=∑I∈k,l[−12∇I2−∑AZA∣rI−RA∣+∑m≠k,l∫dr′ρm(r′)∣rI−r′∣]+∑I∈k,l∑I<J∈k,l1∣rI−rJ∣(12)\hat{H}_{kl} = \sum_{I \in k,l} \left[ -\frac{1}{2}\nabla_I^2 - \sum_A \frac{Z_A}{\lvert r_I-R_A\rvert} + \sum_{m \ne k,l} \int\mathrm{d}r' \frac{\rho_m(r')}{\lvert r_I-r'\rvert} \right] + \sum_{I \in k,l} \sum_{I<J \in k,l} \frac{1}{\lvert r_I-r_J\rvert} \tag*{(12)}

which is used to solve each dimeric problem: H^klΨkl=EklΨkl\hat{H}_{kl}\Psi_{kl}=E_{kl}\Psi_{kl}. With the monomer and dimer calculations, one can combine the solutions to obtain the properties of the full system through the many-body expansion (MBE),

EFMO2−RHF=∑kEk+∑k<l[Ekl−Ek−El],(13)E^{\mathrm{FMO2-RHF}} = \sum_k E_k + \sum_{k<l} \left[ E_{kl} - E_k - E_l \right], \tag*{(13)}

where, EkE_k and EklE_{kl} are electrostatically embedded HF energies of monomer and dimers respectively. Since FMO-RHF lacks the description of explicit electronic correlation, QSCI methods can be implemented along with FMO approximation for accurate electronic energy calculation with the corresponding MBE of the form

EFMO2−QSCI=∑kEkQSCI+∑k<l[EklQSCI−EkQSCI−ElQSCI](14)E^{\mathrm{FMO2-QSCI}} = \sum_k E_k^{\mathrm{QSCI}} + \sum_{k<l} \left[ E_{kl}^{\mathrm{QSCI}} - E_k^{\mathrm{QSCI}} - E_l^{\mathrm{QSCI}} \right] \tag*{(14)}

So, within the FMO-2 approximation, monomer energies can be obtained with distributed QCSC solvers in parallel. Subsequently, to get the dimeric energy contributions (EklQSCIE_{kl}^{\mathrm{QSCI}}}), MTPS is constructed, followed by systematic retrieval of inter-fragment electronic correlation with the help of Gen-ML or S-CORE driven modules discussed in Section IV B. Note that, in this case, no quantum hardware samples for dimeric system is required as it is entirely initialized by MTPS. These steps can be generalized for arbitrary nn-meric FMO approximations as well, albeit at an increased computational cost.

PIGen-SQD Framework as Distributed Fragment QCSC Solver

PIGen-SQD is a class of efficient QCSC algorithms that uses quantum hardware samples, perturbative measures and generative machine learning (Gen-ML) to accurately identify the dominant sector of the Hilbert space that has substantial support to the target state. The initial configuration space is constructed by combining configurations obtained from quantum hardware measurements with additional determinants selected using perturbative estimates. Since hardware noise can produce configurations that violate the required particle-number and spin-projection symmetries, the measured configurations are filtered to retain only those belonging to the desired symmetry sector.

To complement the hardware samples, many-body perturbation theory (MBPT) is employed to identify additional configurations with potentially significant contributions to the target state. Starting from the HF reference, the second-order Møller-Plesset (MP2) amplitudes are given by

tijab=⟨ab∥ij⟩ϵi+ϵj−ϵa−ϵb,t_{ij}^{ab} = \frac{\langle ab \Vert ij \rangle}{\epsilon_i+\epsilon_j-\epsilon_a-\epsilon_b},

where ii, jj and aa, bb denote occupied and virtual spin-orbitals, respectively, and ϵp\epsilon_p is the energy of spin-orbital pp. These amplitudes provide perturbative estimates of the importance of doubly excited determinants. Higher-rank configurations can subsequently be identified through hierarchical perturbative constructions involving connected contractions of quasi two-body excitation operators (called scattering operators) and MP2 amplitudes. Explicit evaluation of all the tensor contractions required for higher-rank excitations can be computationally expensive. PIGen-SQD therefore employs sparse integral screening and index-based symbolic operations to identify candidate configurations without explicitly evaluating all possible tensor contractions. The resulting perturbatively selected configurations are combined with the symmetry-filtered hardware samples to form an enriched initial configuration space,

B(0)=Bhw∪BMBPT,(15)\mathcal{B}^{(0)} = \mathcal{B}_{\mathrm{hw}} \cup\mathcal{B}_{\mathrm{MBPT}}, \tag*{(15)}

where Bhw\mathcal{B}_{\mathrm{hw}} and BMBPT\mathcal{B}_{\mathrm{MBPT}} denote the hardware-derived and perturbatively selected determinant sets, respectively. The perturbative excitation rank and screening thresholds control the size of the initial space and the computational effort required for its construction. The molecular Hamiltonian of interest is projected onto the current determinant space B(m)\mathcal{B}^{(m)} and diagonalized classically,

H(m)c(m)=E(m)c(m),(16)\mathbf{H}^{(m)}\mathbf{c}^{(m)} = E^{(m)}\mathbf{c}^{(m)}, \tag*{(16)}

where H(m)\mathbf{H}^{(m)} is the Hamiltonian matrix in the current configuration space, E(m)E^{(m)} is the target eigenvalue, and c(m)\mathbf{c}^{(m)} is the corresponding CI coefficient vector. The coefficients are used to distinguish configurations with appreciable contributions to the target state from those with negligible contributions. For a prescribed coefficient threshold ϵcoeff\epsilon_{\mathrm{coeff}}, the retained dominant determinant set is defined as

Bdom(m)={∣Φμ⟩∈B(m):∣cμ(m)∣>ϵcoeff}.(17)\mathcal{B}_{\mathrm{dom}}^{(m)} = \left\{ \left|\Phi_\mu\right\rangle\in\mathcal{B}^{(m)} : \left|c_\mu^{(m)}\right| > \epsilon_{\mathrm{coeff}} \right\}. \tag*{(17)}

where μ\mu is the composite hole-particle index. Configurations with coefficients below the threshold are recorded in a blacklist, Bblk(m)\mathcal{B}_{\mathrm{blk}}^{(m)}, to prevent their repeated selection during subsequent iterations. The dominant configurations are then used to construct the training distribution for the generative model. Restricted Boltzmann Machine (RBM) (see Appendix A) is used as the generative model in this case to learn the distribution of configurations from the CI-weighted training data and to predict further dominant determinants from the unseen sectors of the Hilbert space. For a bias-free training, HF determinant is excluded from the RBM training data since the associated coefficient is usually several orders-of-magnitude higher than other configurations. RBM is trained on this distribution and generates new candidate configurations through Gibbs sampling. The newly generated configurations are screened against the blacklist and combined with the existing determinant space to produce an expanded configuration space,

B(m+1)=Bdom(m)∪(Bnew(m)∖Bblk(m)),(18)\mathcal{B}^{(m+1)} = \mathcal{B}_{\mathrm{dom}}^{(m)} \cup\left(\mathcal{B}_{\mathrm{new}}^{(m)} \setminus\mathcal{B}_{\mathrm{blk}}^{(m)}\right), \tag*{(18)}

where Bnew(m)\mathcal{B}_{\mathrm{new}}^{(m)} denotes the set of newly generated configurations after enforcing the required physical symmetries and removing blacklisted determinants. The Hamiltonian is subsequently re-diagonalized in the expanded space, and the resulting CI coefficients provide updated training data for the next iteration.

The procedure alternates between RBM training and configuration generation, referred to as the micro-cycle, and Hamiltonian diagonalization and configuration-space updating, referred to as the macro-cycle. The iterative recovery is terminated when no new configurations are generated or when the energy change between successive macro-cycles satisfies

∣E(m)−E(m−1)∣<ϵE,(19)\left|E^{(m)} - E^{(m-1)}\right| < \epsilon_E, \tag*{(19)}

where ϵE\epsilon_E is the prescribed energy convergence threshold. Algorithmic procedure for PIGen-SQD is summarized in Algorithm 1. The converged eigenvalue and eigenvector are obtained by diagonalizing the Hamiltonian in the final recovered configuration space. The principal advantage of PIGen-SQD is the complementary role of its three components: quantum hardware provides samples from the prepared electronic state that are stochastically drawn irrespective of the excitation rank, perturbative screening introduces physically motivated configurations that may be absent from the measured distribution and provides an anchoring effect for the subsequent Gen-ML training, and generative learning facilitates iterative exploration of the dominant region of the symmetry-preserving Hilbert space. This combination aims to improve the accuracy of the recovered electronic

Algorithm 1 PIGen-SQD

  1. Input: Hamiltonian H^\hat{H}, circuit U^(θ)\hat{U}(\theta), shots NshotsN_{\mathrm{shots}}, perturbative rank nn, thresholds ϵint\epsilon_{\mathrm{int}}, ϵcoeff\epsilon_{\mathrm{coeff}}, ϵE\epsilon_E, sampling fraction x%x\%, particle-number and spin-projection targets

  2. Output: Eigenvalue EE, CI vector c\mathbf{c}

  3. Prepare ∣Ψ(θ)⟩=U^(θ)∣Φ0⟩|\Psi(\theta)\rangle= \hat{U}(\theta)|\Phi_0\rangle and measure NshotsN_{\mathrm{shots}} times in the computational basis

  4. Bhw←\mathcal{B}_{\mathrm{hw}} \leftarrow sampled configurations with correct particle number and spin projection

  5. BMBPT←\mathcal{B}_{\mathrm{MBPT}} \leftarrow configurations up to rank nn from MP2 amplitudes tijt_{ij} and screened scattering contractions (threshold ϵint\epsilon_{\mathrm{int}})

  6. B(0)←Bhw∪BMBPT\mathcal{B}^{(0)} \leftarrow\mathcal{B}_{\mathrm{hw}} \cup\mathcal{B}_{\mathrm{MBPT}}; Bblk(−1)←∅\mathcal{B}_{\mathrm{blk}}^{(-1)} \leftarrow\varnothing

  7. for m=0,1,2,…m=0,1,2,\ldots do

  8. Macro-cycle: solve H(m)c(m)=E(m)c(m)\mathbf{H}^{(m)}\mathbf{c}^{(m)} = E^{(m)}\mathbf{c}^{(m)} in B(m)\mathcal{B}^{(m)}

  9. if m>0m>0 and ∣E(m)−E(m−1)∣<ϵE\lvert E^{(m)}-E^{(m-1)}\rvert< \epsilon_E then break

  10. Bdom(m)←{∣Φμ⟩∈B(m):∣cμ(m)∣>ϵcoeff}\mathcal{B}_{\mathrm{dom}}^{(m)} \leftarrow\{|\Phi_\mu\rangle\in\mathcal{B}^{(m)}:\lvert c_\mu^{(m)}\rvert> \epsilon_{\mathrm{coeff}}\}

  11. Bblk(m)←Bblk(m−1)∪{∣Φμ⟩∈B(m):∣cμ(m)∣≤ϵcoeff}\mathcal{B}_{\mathrm{blk}}^{(m)} \leftarrow\mathcal{B}_{\mathrm{blk}}^{(m-1)} \cup\{|\Phi_\mu\rangle\in\mathcal{B}^{(m)}:\lvert c_\mu^{(m)}\rvert\leq\epsilon_{\mathrm{coeff}}\}

  12. Micro-cycle: train RBM on Bdom(m)∖{∣Φ0⟩}\mathcal{B}_{\mathrm{dom}}^{(m)}\setminus\{|\Phi_0\rangle\} with frequency of μ∝∣cμ(m)∣\mu\propto\lvert c_\mu^{(m)}\rvert

  13. Bnew(m)←\mathcal{B}_{\mathrm{new}}^{(m)} \leftarrow symmetry-preserving configurations generated by Gibbs sampling of x%x\% of the symmetry space

  14. B(m+1)←Bdom(m)∪(Bnew(m)∖Bblk(m))\mathcal{B}^{(m+1)} \leftarrow\mathcal{B}_{\mathrm{dom}}^{(m)} \cup\left(\mathcal{B}_{\mathrm{new}}^{(m)} \setminus\mathcal{B}_{\mathrm{blk}}^{(m)}\right)

  15. if B(m+1)=B(m)\mathcal{B}^{(m+1)}=\mathcal{B}^{(m)} then break

  16. return E(m)E^{(m)}, c(m)\mathbf{c}^{(m)}

state while controlling the dimension of the classical diagonalization problem. Due to the accuracy and efficiency over conventional SQD, in this work, PIGen-SQD is used as a subspace QCSC solver over monomeric fragments. This can be executed in parallel since the fragment eigenstates are independent of each other in the zeroth order approximation of the theory. PIGen-SQD would provide ∣Ψk⟩|\Psi_k\rangle in terms of a set of associated coefficients {cαkk}\{c_{\alpha_k}^k\} and the (local) computational basis {ϕαkk}\{\phi_{\alpha_k}^k\} for all fragments kk under consideration. These set of coefficients and configurations would be used further to construct the MTPS upon an optimal assignment of computational bit indices, followed by inter-fragment correlation recovery as discussed in the subsequent sections.

Mapped Aggregation of Partitioned Subspace

Computational Bit Assignment and MTPS Construction

Our assumption in this work is that the systems of interest have strong electronic correlations that are local in nature and the associated inter-fragment correlation is relatively weaker. While the local correlations are taken care of by distributed PIGen-SQD, an appropriate assignment of the local bits to the global indices is a necessary step to quantitatively reconstruct the nn-meric configurations from monomeric information. This assignment provides a direct correspondence between the occupied (1)/unoccupied (0) computational bits of the physically relevant sub-system configurations and the super-system bits and keeps track of the relative positions during the reconstruction. Conceptually, for a fragmentation scheme, an nn-meric supersystem is partitioned into nn disjoint subsystems, {Sk}k=1n\{S_k\}_{k=1}^{n}, each consisting of MkM_k orbitals and PkP_k particles. For each subsystem SkS_k, let Ik={1,2,…,Mk}\mathcal{I}_k=\{1,2,\ldots,M_k\} denote the index set of its local computational bits, where each configuration is represented by a bitstring of length MkM_k. The global index set of the supersystem is defined as Ig={1,2,…,M}\mathcal{I}_g=\{1,2,\ldots,M\}, where M=∑k=1nMkM=\sum_{k=1}^{n}M_k. The local-to-global index assignment is formalized by a family of injective mappings:

πk:Ik↪Ig,∀k∈{1,…,n}.(20)\pi_k:\mathcal{I}_k\hookrightarrow\mathcal{I}_g,\qquad\forall k\in\{1,\ldots,n\}. \tag*{(20)}

The proposed assignment consequently maps the local index of each bit to its corresponding nn-meric index, thereby enabling the systematic construction of the composite multimeric computational basis from the constituent subsystem basis.

For a quantitative description of this assignment problem, we start with the molecular orbitals (MOs) for fragments χβk\chi_{\beta_k} and super-system χ~g\widetilde{\chi}_g, which are linear combinations of atomic orbitals (AOs) assigned to the kk-th fragment μ~k\widetilde{\mu}_k and full system ξ~ν\widetilde{\xi}_\nu, respectively:

χβk=∑μkCμkβkξ~μk,χ~g=∑νC~νgξ~ν.(21)\chi_{\beta_k}=\sum_{\mu_k}C_{\mu_k\beta_k}\widetilde{\xi}_{\mu_k},\qquad\widetilde{\chi}_g=\sum_\nu\widetilde{C}_{\nu g}\widetilde{\xi}_\nu. \tag*{(21)}

Here, the notation χβk\chi_{\beta_k} denotes the β\beta-th MO of the kk-th molecular fragment, and CC and C~\widetilde{C} are the associated expansion coefficients for fragments and the super-system. Since the MOs form the occupation basis in the fermionic Hilbert space, the computational bit assignment in Eq. (20) translates to

πk:χβk↪χ~g.(22)\pi_k:\chi_{\beta_k}\hookrightarrow\widetilde{\chi}_g. \tag*{(22)}

This mapping operation πk\pi_k computationally translates to a linear assignment problem (LAP) and requires the minimization of a cost matrix JgβkJ_{g\beta_k} for each gg and βk\beta_k defined by (see Appendix D for details)

Jgβk=1−[∣Sgβk∣2Δϵgβk].(23)J_{g\beta_k}=1-\left[\lvert S_{g\beta_k}\rvert^2\Delta\epsilon_{g\beta_k}\right]. \tag*{(23)}

Here, the cost function JgβkJ_{g\beta_k} consists of two similarity scores: Sgβk=∑ν∑μkC~νgξ~ν∣ξμk⟩/⟨μkβkS_{g\beta_k}=\sum_\nu\sum_{\mu_k}\widetilde{C}_{\nu g}\widetilde{\xi}_\nu\lvert\xi_{\mu_k}\rangle/\langle\mu_k\beta_k, which is the MO overlap matrix element serving as a metric of spatial similarity, and a Gaussian energy similarity factor Δϵgβk=exp⁡(−(ϵg−ϵβk−δϵβk)22σ2)\Delta\epsilon_{g\beta_k}=\exp\left(-\frac{(\epsilon_g-\epsilon_{\beta_k}-\delta\epsilon_{\beta_k})^2}{2\sigma^2}\right), where ϵg\epsilon_g and ϵβk\epsilon_{\beta_k} are the orbital energies of the gg-th global orbital and β\beta-th local orbital of the kk-th fragment, σ\sigma is the energy error tolerance, and δϵβk\delta\epsilon_{\beta_k} is the corresponding mean of the Gaussian distribution (see Appendix D). The elements of the cost matrix provide a score for each gg and βk\beta_k, and our goal is to minimize this cost function such that the local index βk\beta_k can be optimally assigned a global position gg. This can be solved by the Hungarian algorithm [52], which is widely used to solve LAPs, particularly in the fields of multi-robot task allocations [53]. A detailed discussion regarding the LAP and the functional form of the cost function is provided in Appendix § 1.5.3 and Appendix § 1.14, respectively.

MTPS Construction: The fragment QCSC solvers (in this case PIGen-SQD) generate a set of dominant configurations for each fragment,

Bk={∣ϕ1k⟩,∣ϕ2k⟩,…,∣ϕαkk⟩,…,∣ϕdkk⟩},(24)\mathcal{B}_k = \left\{ \left|\phi_1^k\right\rangle, \left|\phi_2^k\right\rangle, \ldots, \left|\phi_{\alpha_k}^k\right\rangle, \ldots, \left|\phi_{d_k}^k\right\rangle\right\}, \tag*{(24)}

where ∣ϕαkk⟩=∣xMk…x2x1⟩∈{0,1}Mk\left|\phi_{\alpha_k}^k\right\rangle= \left|x_{M_k} \ldots x_2x_1\right\rangle\in\{0,1\}^{M_k} represents the computational-basis bitstring for the αk\alpha_k-th configuration of the kk-th fragment (with individual bits denoted by xx), and dkd_k denotes the number of configurations retained for the kk-th fragment by the fragment PIGen-SQD solver. In practice, configurations with an associated CI coefficient satisfying

∣cαkk∣>ϵ,(25)\left|c_{\alpha_k}^k\right| > \epsilon, \tag*{(25)}

where ϵ\epsilon is a pre-defined threshold, are retained to construct the dominant local configuration space.

Using the computational bit assignment, each local spin-orbital is assigned to its corresponding position in the computational basis of the composite system. The dominant configurations obtained from the individual fragment CI calculations can therefore be embedded into the composite computational basis and combined to construct a compact global subspace. We define this mapped tensor product state (MTPS) as

BMTPS={∣Φ~g⟩}=M(⨂k=1nBk),(26)\mathcal{B}_{\mathrm{MTPS}} = \left\{ \left|\widetilde{\Phi}_g\right\rangle\right\} = \mathcal{M}\left(\bigotimes_{k=1}^{n} \mathcal{B}_k\right), \tag*{(26)}

where ∣Φ~g⟩=∣x~M…x~2x~1⟩∈{0,1}M\left|\widetilde{\Phi}_g\right\rangle= \left|\widetilde{x}_M \ldots\widetilde{x}_2\widetilde{x}_1\right\rangle\in\{0,1\}^{M} are the composite super-system configurations, and M(⋅)\mathcal{M}(\cdot) denotes the mapped tensor-product operation according to the computational bit assignment formalism. In contrast to a conventional tensor product, MTPS embeds the local configurations into the computational basis of the composite system according to the orbital assignment. This distinction is essential because the local configurations ∣ϕαkk⟩\left|\phi_{\alpha_k}^k\right\rangle are defined in the Hilbert space of the individual fragments, whereas the resulting configurations ∣Φ~g⟩\left|\widetilde{\Phi}_g\right\rangle belong to the Hilbert space of the composite system. The resulting mapped product configurations form a subset of the computational basis of the composite system {∣Φg⟩}\{\left|\Phi_g\right\rangle\},

{∣Φ~g⟩}⊆{∣Φg⟩},(27)\left\{ \left|\widetilde{\Phi}_g\right\rangle\right\} \subseteq\{\left|\Phi_g\right\rangle\}, \tag*{(27)}

and define a compact subspace of the exponentially large Hilbert space. Since this subspace is constructed from the dominant configurations of the individual fragment CI solutions, it is expected to retain substantial support to the target eigenstate of the full Hamiltonian. The resulting compact subspace can subsequently be used for projection and diagonalization of the full Hamiltonian, thereby providing an efficient route to incorporate inter-fragment correlation beyond the zeroth-order product-state description.

Cross-Fragment Configurations from Local CI Solutions: The MTPS constructed from the individual monomer determinants intrinsically accounts for intra-fragment excitations and embeds them appropriately in the Fock space of the composite system. However, in general, it does not contain inter-fragment correlation effects, which include configurations corresponding to electronic excitations between the constituent fragments. Such inter-fragment excitations can be particularly important for weakly bound donor-acceptor complexes, where the frontier orbitals of the composite system may be localized predominantly on different monomeric subunits. The omission of these configurations can therefore restrict the description of inter-fragment charge redistribution and charge-transfer characters. To account for these effects, we augment the mapped product space with a set of cross-fragment excited configurations.

Since the origin of each orbital (or bit in the computational basis) is identified and can be associated with a particular fragment using the computational bit assignment, cross-fragment terms can be generated by exciting electrons from each determinant of a particular fragment to the unoccupied spinorbitals of neighboring fragments. Although the number of all cross-fragment excited configurations grows combinatorially with the number of electrons in one fragment and the number of unoccupied orbitals in the other fragments, we restrict this expansion by applying a one-body cross-fragment excitation (within the same spin sector) from ∣ϕαkk⟩\left|\phi_{\alpha_k}^k\right\rangle to ∣ϕαll⟩\left|\phi_{\alpha_l}^l\right\rangle (and vice versa) only if the numerical product of the corresponding monomeric CI coefficients satisfies

∣cαkkcαll∣>ϵcf,(28)\left|c_{\alpha_k}^k c_{\alpha_l}^l\right| > \epsilon_{\mathrm{cf}}, \tag*{(28)}

where cαkkc_{\alpha_k}^k and cαllc_{\alpha_l}^l denote the corresponding coefficients of the associated determinants, and ϵcf\epsilon_{\mathrm{cf}} is the cross-fragment threshold. This set of cross-fragment configurations supplements the mapped product space and explicitly introduces low-rank inter-fragment correlation effects. From here onwards, the definition of MTPS includes collectively Eq. (26) as well as the cross fragment single excitations discussed here.

Systematic Recovery of Inter-Fragment Correlation Effects

The MTPS construction described in Section IV A exploits the chemically motivated fragmentation involving strong local electronic correlations and partially recovers the inter-fragment correlations effects with low-rank cross-fragment excitations. While MTPS provides a substantial overlap with the target state, a systematic recovery of inter-fragment correlation is necessary for an accurate description of the composite system. Toward this, MTPS provides an extremely good initialization for subsequent self-consistent inter-fragment correlation recovery. Here we consider two possible inter-fragment correlation recovery methods as follows:

S-CORE Initialized by MTPS: The self-consistent configuration recovery or S-CORE module[36] in the SQD protocol provides a systematic recovery of dominant determinants for accurate electronic correlation energy estimation with Hamiltonian projection and diagonalization. However, the accuracy and efficiency of S-CORE depends on the initial point of the iterative procedure in a way that a better initial set of dominant configurations leads to more accurate and more compact diagonalization subspace. Thus for extended systems with multiple chemically motivated fragments, MTPS qualifies as a natural choice for such an initialization. In this paper we have used S-CORE with carryover mechanism[39], which uses an additional retention step to save and carry important set of bitstrings from one iteration to the next iteration to ensure that energetically relevant configurations are incorporated in the recovery process. This inter-fragment recovery with S-CORE initialized by MTPS will be referred to as S-CORE(MTPS).

Generative Machine-Learning-Driven Recovery: A more sophisticated and systematic approach would be to self-consistently train a secondary layer of Gen-ML model initialized with MTPS to search and select unseen dominant sectors of the Hilbert space. So in this entire workflow, we use two layer of Gen-ML: one in PIGen-SQD and one in the cross-fragment recovery stage and both the stages use the same RBM module discussed in details in Appendix A. This pipeline consisting of PIGen-SQD, MTPS and Gen-ML driven cross-fragment recovery will be referred to as PIGen-SQD with mapped aggregation of partitioned subspace or PIGen-SQD(MAPS).

To summarize, the PIGen-SQD(MAPS) algorithm involves the following steps:

  1. Parallel Fragment QSCI Solvers: The molecular system is divided into chemically motivated fragments with dominant localized electronic correlations. Once the optimal fragments are obtained, PIGen-SQD (as discussed in section III B) modules are used for each fragment as distributed QCSC solver.

2. Computational Bit Assignment and MTPS Construction: After the convergence of PIGen-SQD cycles, all the dominant configurations based on their associated coefficients are collected. The computational bit positions are appropriately assigned according to the approach described in Section IV A to map each local bit to a global bit. This enables one to associate local orbitals to global orbitals. With the assigned bits, MTPS is constructed from dominant fragment configurations obtained from PIGen-SQD.

3. Systematic Recovery of Inter-fragment Correlation: Inter-fragment correlation, which is only partially recovered MTPS construction, is self-consistently recovered using standard S-CORE initialized with MTPS or a secondary layer of Gen-ML module, as discussed in Section IV B.

The associated schematics of the workflow is shown in Figure 1.

PIGen-SQD workflow, computational bit mapping and MTPS construction, decomposition, and aggregation

Figure 1. (a) PIGen-SQD workflow: A molecular system is encoded into a PQC and measurement is done to obtain the samples. These samples are then classically post-processed with Gen-ML driven modules for fermionic configuration reconstruction. (b) Computational bit mapping and MTPS construction: The orbitals of two sub-systems, A and B are appropriately assigned to the super-system bits. The dominant configurations obtained by distributed PIGen-SQD is then translated to the global bits from local bits accordingly. PIGen-SQD(MAPS) workflow is shown in the lower panel with (c) showing the problem decomposition and distributed PIGen-SQD solver, followed by (d) aggregation with MTPS and Gen-ML driven inter-fragment correlation recovery.

Most importantly, in this workflow, we have entirely circumvented the circuit preparation for the composite system, which would otherwise require substantially deeper circuits and extremely high number of shots for meaningful sampling. The mapped product state is directly fed into the S-CORE or Gen-ML based configuration recovery module, which self-consistently recovers additional dominant configurations until convergence for an accurate description of the composite system while only exploring a restricted space at a time.

Results and Discussion

Computational details

Molecular Calculations and Fragmentation: In this paper, biochemically relevant (H2O)2(\mathrm{H_2O})_2 and formamide (FMA) dimer (geometry given in Supplementary Information) are studied at different intermolecular separations. In particular, accurate electronic structure calculations of FMA dimers are of immense biochemical importance as it is the smallest system containing the peptide bond. For the dimeric systems (which is the full system in this particular example), restricted Hartree-Fock (RHF) calculations are performed with PySCF using 6-31G basis set. The dimer active spaces for (H2O)2(\mathrm{H_2O})_2 and FMA homo-dimer are chosen to be (16e, 20o) and (36e, 24o) which forms 1.58×10101.58 \times10^{10} and 1.81×10101.81 \times10^{10} dimensional FCI space respectively. Effective one-electron and two-electron active space integrals with the constant core energy contributions are obtained from PySCF CASCI integral transformations. For distributed QCSC calculations, the dimeric systems are fragmented into two monomeric sub-systems such that each fragment of H2O\mathrm{H_2O} and FMA consists of (8e,10o) and (18e,12o) active spaces with (104)2=44100\binom{10}{4}^{2} = 44100 and (129)2=48400\binom{12}{9}^{2} = 48400 dimensional FCI space respectively.

Calculations with GAMESS-US[54] are performed to obtain fragment-resolved molecular orbital coefficients, one-electron integrals, two electron integrals, fock matrices, density matrices and nuclear-repulsion energies for each monomer, embedded in the mean-field of the other monomer. Also, to test the capabilities of the PIGen-SQD(MAPS) workflow for studying extended systems, we have studied (H2O)6(\mathrm{H_2O})_6 (geometry given in Supplementary Information) system under the FMO-2 approximation. Different classical SCI methods have been studied such as Heat-bath CI (HCI) and standard SCI implementation in PySCF (which we will refer to as SCI-PySCF) for comparing the performance of PIGen-SQD(MAPS). HCI results are obtained with qiskit_addon_dice[55, 56] package with the variational selection threshold kept at 10−510^{-5}, while SCI-PySCF results are obtained with pyscf.fci.selected_ci routine in PySCF[57].

Quantum Circuit Details: In the PIGen-SQD(MAPS) workflow, PIGen-SQD routines are used in parallel for each monomer which requires monomeric quantum hardware samples followed by Gen-ML driven fermionic state reconstruction[43]. The perturbative configurations in the PIGen-SQD workflow are obtained with sparse-tensor operations and the corresponding algebraic terms are given in Appendix F. Conventional SQD is also studied over the dimeric systems to compare the results that uses the samples obtained for dimer system followed by S-CORE modules with carryover mechanism[39] as implemented in qiskit_addon_sqd[58] (version 0.12.1) with carryover threshold set at 10−610^{-6}. LUCJ ansatz with optimized CCSD parameters are implemented with ffsim[59] package for both monomeric and dimeric systems. The corresponding spinorbitals are mapped onto ibm_boston hardware that runs on IBM Heron R3 processor with heavy-hex coupling map. The qiskit_zigzag_layout function is used for the qubit assignments on quantum processors. S-CORE(MTPS) results are obtained with standard carryover S-CORE module. The hyperparameters used for SQD and S-CORE(MTPS) are given in Table IV in Appendix B. For PIGen-SQD(MAPS) workflow, only monomeric hardware samples are used along with multi-layered Gen-ML driven recovery modules to get the solutions for each fragment and for subsequent inter-fragment recovery. No dimeric hardware samples are used for PIGen-SQD(MAPS) framework to demonstrate it as a method with low shot requirement. However, it is straightforward to add dimeric samples as well along with MTPS. The coupling map of the QPU and associated qubit assignments are shown in Figure 4 (a, Figure 4c) for H2O\mathrm{H_2O} and FMA monomer. The details regarding the system active space, number of spinorbitals, number of qubits (including the auxiliary qubits), NshotsN_{\mathrm{shots}}, FCI dimension, two-qubit gate count and two-qubit gate depth are shown in Table I.

Coupling maps and qubit assignments for water and formamide dimers, with energy errors and diagonalization dimensions

Figure 4. Coupling map and qubit assignment on IBM Heron R3 processor for monomeric fragments of water and FMA dimer are shown in (a) and (c) respectively. The intermolecular Oxygen-Oxygen distance between two H2OH_2O monomers is 2.2 Å and the perpendicular distance between the parallel C–N bonds of the two FMA monomers is 3.23 Å. Here α\alpha (β\beta) orbitals are shown in green (red) along with the auxiliary qubits (in yellow) that encode density-density interactions in the LUCJ ansatz. Energy error (in EhE_h) with respect to HCI and the associated diagonalization dimension are shown in (b) and (d) for SQD, S-CORE(MTPS) and PIGen-SQD(MAPS). Grey dotted line shows the 1 mEhE_h error with respect to HCI.

HyperparametersSQDS-CORE(MTPS)
Iterations1010
Number of Batches22
Samples Per Batch100010000
Energy Tolerance10−510^{-5}10−510^{-5}
Occupancies Tolerance10−510^{-5}10−510^{-5}
Carryover Threshold10−610^{-6}10−610^{-6}

Table IV. Hyperparameters used in SQD and S-CORE(MTPS).

Computational Bit Assignment, MTPS and Inter-fragment Correlation Recovery: The LAP associated with computational bit asassignment (Section IV A) was solved via the scipy.optimize.linear_sum_assignment routine in SciPy (v1.17.0) [60]. The routine is based on Crouse’s formulation [61], representing one of several highly efficient adaptations of the original Hungarian algorithm [62, 63].

System (Active Space)SpinorbitalsQubitsNshotsN_{\mathrm{shots}}dim⁡(FCI)\operatorname{dim}(\mathrm{FCI})Gate CountGate Depth
H2_2O (8e, 10o)20232×1052 \times10^{5}4410052486
(H2_2O)2_2 (16e, 20o)40455×1065 \times10^{6}1.58×10101.58 \times10^{10}2072154
FMA (18e, 12o)24272×1052 \times10^{5}4840064482
(FMA)2_2 (36e, 24o)48543×1063 \times10^{6}1.81×10101.81 \times10^{10}2672170

Table 1. Details of number of spinorbitals, number of qubits (including auxiliary qubits), number of shots, FCI dimension, two-qubit quantum gate count, and two qubit circuit depth for the molecular systems studied with the LUCJ ansatz.

The associated MTPS construction is done in two phases which incorporate both intra and inter-fragment correlation effects as described in Section IV A. For generating the intra-fragment contributions, the dominant configurations from each monomer are assembled based on a pre-defined cut-off threshold ∣cαkk∣>ϵ\lvert c_{\alpha_k}^{k} \rvert> \epsilon (in this case ϵ=10−6\epsilon= 10^{-6}. With the assumption that strong correlation effects have intra-fragment characteristics, and to restrict the dimensionality of the product space, the monomer configurations were ordered by decreasing coefficient magnitude and MTPS is formed only with those cαkkc_{\alpha_k}^{k} and cαllc_{\alpha_l}^{l} that satisfy ∣cαkkcαll∣>10−7\lvert c_{\alpha_k}^{k}c_{\alpha_l}^{l}\rvert> 10^{-7}. The corresponding occupation vectors were then permuted into the canonical dimer orbital ordering using the LAP. Product-space generation was parallelized over all available CPU cores. For generating the inter-fragment sector of the MTPS we have used ϵcf=10−5\epsilon_{cf} = 10^{-5} in Eq. (28). Inter-fragment excitations are generated in a way such that the S^z\hat{S}_z spin symmetry of the super-system is maintained. This cross-fragment excitation generation is parallelized similar to the mapped product space generation and duplicate determinants were removed. HF, CISD and perturbatively obtained triples (see Appendix F) for dimeric system are augmented with MTPS to provide a better initialization for the subsequent configuration recovery modules. Spin-mirror partners, generated by interchanging the up and down spin occupation sectors, were subsequently included. This is followed by a final removal of any additional duplicate determinants in the entire set of configurations and this de-duplicated set of configurations are passed to the diagonalizer. Note that from here onwards, MTPS refers to this combined subspace which contains manual augmentation of HF, CISD perturbative triples followed by de-duplication to ensure that no obvious lower order dominant configurations are left out. In the next section (§ 1.5 B), the numerical demonstrations different inter-fragment recovery mechanisms initialized with MTPS are discussed in details.

Water and Formamide Dimer

Computational Bit Assignment: One of the most important steps of the PIGen-SQD(MAPS) workflow is the computational bit assignment from local monomeric system to nn-meric system since an optimal local to global mapping is essential for accurate reconstruction of super-system orbitals from monomeric PIGen-SQD solutions. Figure 2 shows a subset of dimer orbitals for (H2O)2(\mathrm{H}_2\mathrm{O})_2 adjacent to highest occupied MO (HOMO) and lowest unoccupied MO (LUMO) region (center) along with the assigned monomeric orbitals, denoted as M1 (right) and M2 (left) for monomer 1 and 2, respectively, obtained with computational bit mapping. The assignment cost function combines the squared AO-metric orbital overlap with a Gaussian orbital-energy compatibility factor as discussed in section IV A. The orbitals are shown with the same isosurface value and molecular orientations. Red and blue colour denote the global phase of the orbitals, however, the overall phase is arbitrary and does not affect the mapping because square of the orbital overlap (∣Sgβk∣2\lvert S_{g_{\beta_k}}\rvert^2) is used in the assignment. Figure 2 pictorially represents that the LAP used in this study is fairly optimal as the local molecular orbitals retain their shape in the dimeric super-system (up to a constant phase factor). The orbital assignment for FMA dimer is shown in Figure E1 in Appendix E.

Orbital correspondence in (H₂O)₂

Figure 2. Orbital correspondence in (H2_2O)2_2. (Center) Subset of dimer molecular orbitals situated near the HOMO-LUMO region. (Right, Left) Assigned monomeric orbitals for monomer 1 (M1) and monomer 2 (M2), respectively, established through computational bit assignment.

Orbital correspondence in FMA dimer

Figure E1. Orbital correspondence in FMA dimer. (Center) Subset of dimer molecular orbitals situated near the HOMO-LUMO region. (Left, Right) Mapped monomeric orbitals for monomer 1 (M1) and monomer 2 (M2), respectively, established through computational bit assignment.

MTPS: Figure 3 shows the contribution of MTPS toward the recovery of electronic correlation directly from monomeric solutions without any explicit iterative correlation recovery performed over dimeric system. The y-axes show the energy error with respect to HCI (in mEh_{\mathrm{h}}) and on top of each bar plot, the associated diagonalization dimension is mentioned. It shows the systematic improvements upon incorporating the perturbative triples and MTPS over CISD determinants. For equilibrium geometries, MTPS shows around 7 mEh_{\mathrm{h}} and 3 mEh_{\mathrm{h}} improvements over CISD for (H2O)2(\mathrm{H}_2\mathrm{O})_2 and (FMA)2(\mathrm{FMA})_2 respectively. For stretched inter-molecular separation of water dimer, MTPS is extremely accurate (<1< 1 mEh_{\mathrm{h}} error) since electronic correlations effects are highly localized and thus MTPS provides an accurate description of the weak inter-molecular interactions in such cases.

Contribution of CISD, CISD + perturbative triples and MTPS configurations toward recovery of correlation energy

Figure 3. Contribution of CISD, CISD + perturbative triples and MTPS configurations toward recovery of correlation energy for water and FMA dimer for two different separation distances each. In y-axes energy error (in mEh_{\mathrm{h}}) with respect to HCI is shown. On top of each bar plot, the associated diagonalization dimension is also shown. The dotted line shows 1 mEh_{\mathrm{h}} error from HCI.

Inter-Fragment Correlation Recovery: Since the performance of any inter-fragment correlation recovery module, be it Gen-ML or S-CORE driven, is sensitive to the quality of the initial seed, MTPS provides extremely good initialization for the same. To demonstrate this, in Figure 4 various configuration recovery strategies are compared. For a fair and head-to-head comparison, a study is shown between standard SQD with quantum hardware samples seed (obtained from circuits constructed for dimeric systems) and S-CORE(MTPS), which refers to the standard S-CORE module of SQD with MTPS seed. S-CORE(MTPS) differs from SQD only in terms of the initialization of the configuration recovery module: SQD uses quantum hardware samples for dimeric systems while S-CORE(MTPS) uses MTPS obtained from monomeric PIGen-SQD solutions. In this case, one can clearly see the importance of initial seed for S-CORE module. While SQD suffers from slow convergence, inaccurate energy estimation and huge diagonalization dimension, the corresponding S-CORE(MTPS) seed shows faster convergence, provides significantly more accurate results while using substantially low dimensional diagonalization space for both (H2O)2(\mathrm{H}_2\mathrm{O})_2 and FMA dimer (Figure 4 (b) and (d)). This comparison shows that MTPS, which is obtained from shallower monomeric circuits, low shot budget and consequently better span of the monomeric Hilbert space, shows substantial improvements in terms of both accuracy and efficiency compared to standard SQD even with the same S-CORE mechanism.

The corresponding PIGen-SQD(MAPS) in both the cases provides the perfect balance of accuracy and efficiency. The systematic iterative identification of dominant configurations and blacklisting mechanism of PIGen-SQD(MAPS) provides 95% reduction in diagonalization space dimension while having an order-of-magnitude better energy accuracy for (H2O)2(\mathrm{H}_2\mathrm{O})_2. Somewhat similar trend is observed in (FMA)2(\mathrm{FMA})_2 as well.

Potential Energy Surface (PES): Figure 5(a) shows the PES for water dimer obtained with SQD, PIGen-SQD(MAPS) and S-CORE(MTPS) with varying oxygen-oxygen distance (in Å) of two different monomers. For comparison, corresponding HCI, SCI-PySCF and CCSD results are also plotted. Figure 5(b) and (c) shows the corresponding error with respect to HCI and diagonalization dimension respectively. The results show that PIGen-SQD(MAPS), SCI-PySCF and S-CORE(MTPS) are all within 1 mEhE_h accuracy with HCI for all the intermolecular Oxygen-Oxygen distances under consideration. CCSD has around 2 mEhE_h error while SQD shows around 8-10 mEhE_h error throughout all the distances. The interesting observation is that even though SQD and S-CORE(MTPS) uses the same configuration recovery module, the MTPS initialization steers S-CORE toward exploring the relevant sectors of the Hilbert space and shows remarkable accuracy while restricting the diagonalization to substantially compact subspaces in comparison with SQD.

Potential energy surfaces and corresponding errors and diagonalization dimensions for (H₂O)₂

Figure 5. Potential energy surface (PES) for (H₂O)₂. Along the y-axes (a) shows the energy (in Eₕ) with different methods: CCSD, SCI-PySCF, SQD, S-CORE(MTPS), PIGen-SQD(MAPS) and HCI, (b) shows error with respect to HCI and (c) shows the corresponding diagonalization dimension. Different Oxygen-Oxygen distance (in Å) between two monomers are represented along x-axes.

ion with SQD. However, SQD accuracy can be further increased by tuning the associated hyperparameters albeit with increased diagonalization dimension. For all the distances, PIGen-SQD(MAPS) provides the most compact diagonalization subspace in this study while attaining consistently around 0.4 mEhE_h accuracy with HCI. Figure 6(a) shows the potential energy curve of the FMA dimer as a function of the perpendicular distance between the parallel C–N bonds of the two monomers. Figure 6(b) and (c) demonstrates S-CORE(MTPS) is the most accurate, providing an order-of-magnitude more accurate energetics compared to conventional SQD while utilizing only one third of the diagonalization cost. In this case also, PIGen-SQD(MAPS) shows the perfect balance between accuracy and compact diagonalization space as it consistently has 0.4 mEhE_h error over the surface and requires almost an order-of-magnitude reduced number of configurations in the diagonalization space. SCI-PySCF lies somewhere in between in terms of accuracy and efficiency.

Potential energy surfaces and corresponding quantities for (FMA)₂

Figure 6. PES and other quantities for (FMA)₂. Axes information are the same as that of Fig. 5.

Water Hexamer

To test the scalability and accuracy of our distributed workflow for extended systems, we investigated the cage conformer of the water hexamer, (H2O)6(\mathrm{H_2O})_6, under the two-body fragment molecular orbital (FMO-2) approximation. Within this framework, the total system is partitioned into N=6N = 6 monomeric units, yielding (62)=15\binom{6}{2} = 15 unique dimeric pairs (fragment pairs). Each dimer is treated independently in parallel using the PIGen-SQD(MAPS) algorithm, and the individual subsystem energetics are aggregated via the many-body expansion according to Eq. (14).

The resulting FMO-2 energies obtained across restricted Hartree–Fock (RHF), heat-bath configuration interaction (HCI), PIGen-SQD(MAPS), and coupled-cluster with singles and doubles (CCSD) are summarized in Table II. Across the 120-qubit active space spanned by the hexamer, direct full-system state preparation and measurement are entirely bypassed; instead, quantum sampling is restricted to 20-qubit monomeric circuits, followed by inter-fragment correlation recovery across 40-qubit dimeric subsystems.

Remarkably, PIGen-SQD(MAPS) achieves an FMO-2 ground-state energy of −456.544231 Eh-456.544231\ E_{\mathrm{h}}, recovering the electronic correlation within sub-millihartree accuracy (0.915 mEhmE_{\mathrm{h}}) relative to the classical reference EHCIFMO−2=−456.545146 EhE_{\mathrm{HCI}}^{\mathrm{FMO-2}} = -456.545146\ E_{\mathrm{h}}. Furthermore, PIGen-SQD(MAPS) outperforms ECCSDFMO−2=−456.542174 EhE_{\mathrm{CCSD}}^{\mathrm{FMO-2}} = -456.542174\ E_{\mathrm{h}} by nearly 2.06 mEhmE_{\mathrm{h}}, demonstrating the superior capability of the physics-informed generative recovery scheme to capture higher-order inter-fragment correlation effects beyond standard single- and double-excitation manifolds. These results validate PIGen-SQD(MAPS) as an accurate and resource-efficient solver for utility-scale chemical simulations on near-term quantum-HPC architectures.

ERHFFMO−2E_{\mathrm{RHF}}^{\mathrm{FMO}-2}ECCSDFMO−2E_{\mathrm{CCSD}}^{\mathrm{FMO}-2}EPIGen−SQD(MAPS)FMO−2E_{\mathrm{PIGen-SQD(MAPS)}}^{\mathrm{FMO}-2}EHCIFMO−2E_{\mathrm{HCI}}^{\mathrm{FMO}-2}
-456.00503267-456.54217401-456.54423085-456.54514579

Table 2. Comparison between RHF, HCI, PIGen-SQD(MAPS) and CCSD under the FMO-2 approximation.

Conclusion and Future Outlook

In this work, we propose a distributed quantum-HPC hybrid architecture to treat inter-molecular electronic correlations, larger active spaces and extended systems. The developed algorithm involves distributed QCSC solvers and multi-layered Gen-ML modules to decompose a large-scale problem into parallel tractable computational tasks, followed by a two-stage aggregation process to integrate the individual sub-systems and reconstruct the solution of the super-system. Our algorithm, which we refer to as PIGen-SQD(MAPS), employs PIGen-SQD as the distributed QCSC solver which provides a set of dominant configurations and the associated coefficients for each sub-system. These individual solutions are collected and fed into a two-stage aggregation process to get the target solution at a much reduced computational cost compared to the parent QCSC method. The aggregation step starts with a computational bit assignment, which provides an one-to-one mapping between the local and global bits. Subsequently, a mapped tensor product state (MTPS) is constructed to recover a dominant subspace of the global Hilbert space from the dominant configurations of the individual fragments. Finally MTPS is fed into an iterative inter-fragment correlation recovery module driven by Gen-ML or carryover S-CORE mechanism. The multi-layered Gen-ML engine, one at the parallel QCSC level and one at inter-fragment recovery stage drives the algorithm along with quantum hardware samples (obtained for each fragmented systems) and physically motivated MTPS construction. Due to the distributed nature, PIGen-SQD(MAPS) requires fewer qubits, shallower circuits, and drastically reduced shot requirements per QCSC task that improves the quality of the samples obtained for each sub-problem compared to conventional SQD.

Our numerical demonstrations on non-covalent dimeric systems, including the hydrogen-bonded water dimer (H2O)2(\mathrm{H}_2\mathrm{O})_2 and the peptide-bond prototype formamide dimer (FMA)2(\mathrm{FMA})_2, reveal that PIGen-SQD(MAPS) yields an order-of-magnitude improvement in energy accuracy compared to standard SQD performed on the entire composite system, while simultaneously reducing the classical Hamiltonian diagonalization dimension by up to 95%. Furthermore, the application to the cage conformer of the water hexamer (H2O)6(\mathrm{H}_2\mathrm{O})_6 under the FMO-2 approximation illustrates the scalability of our approach, effectively decomposing a demanding 120-qubit problem into six parallel 20-qubit monomeric QCSC tasks, followed by aggregation stages. The resulting aggregated subspace shows strong agreement with the reference calculations with HCI and provides better accuracy than reference FMO-2-CCSD calculations.

By restricting expensive quantum measurements to localized active spaces and Gen-ML driven cross-fragment entanglement, PIGen-SQD(MAPS) provides a more scalable and systematically improvable avenue for electronic structure calculations on utility-scale quantum-HPC architectures. Furthermore, the framework provides us with great flexibility in selectively capturing the many-body correlation among the weakly entangled fragments through MTPS with further cross-fragment entanglement build-up by Gen-ML or S-CORE(MTPS). This hybrid mechanism allows us to treat different parts of a large system with various degrees of sophistication as warranted by its electronic complexity. Future extensions will focus on generalizing the bit-mapping protocol to heterogeneous environments and integrating advanced generative architectures such as transformers[64] or autoencoders[65] for multi-fragment correlation recovery.

Author contributions

Conceptualization: Chayan Patra, Prayas Ghosh, Rahul Maitra

Methodology: Chayan Patra, Prayas Ghosh, Rahul Maitra

Investigation: Chayan Patra, Prayas Ghosh, Rahul Maitra

Formal analysis: Chayan Patra, Prayas Ghosh, Rahul Maitra

Data curation: Chayan Patra, Prayas Ghosh

Writing - original draft: Chayan Patra, Rahul Maitra

Writing - review and editing: Chayan Patra, Prayas Ghosh, Rahul Maitra

Software: Chayan Patra, Prayas Ghosh, Adwait Pathadekar

Visualization: Chayan Patra, Prayas Ghosh

Supervision: Rahul Maitra

Funding acquisition: Rahul Maitra

CONFLICT OF INTEREST

The authors declare no conflict of interest.

Data and Code Availability

The data and code are available upon reasonable request to the corresponding author.

Acknowledgments

We acknowledge the use of IBM Quantum Credits for this work. The views expressed are those of the authors, and do not reflect the official policy or position of IBM or the IBM Quantum team. RM and CP acknowledges the financial support from Quantum Algorithms project with sanction number DST/QTC/NQM/QA/2026/01 under the National Quantum Mission (NQM) scheme, Department of Science and Technology (DST), Government of India. AP and RM acknowledge financial support from Industrial Research and Consultancy Centre (IRCC), IIT Bombay. The authors acknowledge the discussions with Sonaldeep Halder on some technical aspects of the code during the initial phase of the project.

Restricted Boltzmann Machines for Configuration Recovery

Restricted Boltzmann Machines (RBMs) constitute a class of unsupervised, generative stochastic neural networks capable of modeling complex, high-dimensional nonlinear probability distributions. Structurally, an RBM is organized as a bipartite graph comprising a visible layer g=({gi}i=1D,gi∈{0,1})g = (\{g_i\}_{i=1}^{D}, g_i \in\{0,1\}) that encodes system configurations, and a hidden layer h=({hj}j=1J,hj∈{0,1})h = (\{h_j\}_{j=1}^{J}, h_j \in\{0,1\}) that introduces auxiliary degrees of freedom to enhance expressivity. The network parameters θ={u,y,W}\theta= \{\mathbf{u},\mathbf{y},\mathbf{W}\} consisting of the visible bias vector u\mathbf{u}, the hidden bias vector y\mathbf{y}, and the inter-layer weight matrix W\mathbf{W}, are optimized to accurately approximate the underlying data distribution.

Mathematically, the joint probability distribution over the visible and hidden units is governed by an energy function:

E(g,h,θ)=−∑i=1D∑j=1JgiWijhj−∑j=1Jyjhj−∑i=1Duigi,(29)E(\mathbf{g},\mathbf{h},\theta)=-\sum_{i=1}^{D}\sum_{j=1}^{J}g_iW_{ij}h_j-\sum_{j=1}^{J}y_jh_j-\sum_{i=1}^{D}u_ig_i, \tag*{(29)}

from which the marginal probability of a visible configuration g\mathbf{g} is derived via the canonical Gibbs distribution:

P(g;θ)=∑hP(g,h;θ)=∑h1Z(θ)exp⁡(−E(g,h,θ)),(30)P(\mathbf{g};\theta)=\sum_{\mathbf{h}}P(\mathbf{g},\mathbf{h};\theta)=\sum_{\mathbf{h}}\frac{1}{Z(\theta)}\exp\left(-E(\mathbf{g},\mathbf{h},\theta)\right), \tag*{(30)}

where the partition function Z(θ)Z(\theta) normalizes the distribution over all possible states:

Z(θ)=∑g∑hexp⁡(−E(g,h,θ)).(31)Z(\theta)=\sum_{\mathbf{g}}\sum_{\mathbf{h}}\exp\left(-E(\mathbf{g},\mathbf{h},\theta)\right). \tag*{(31)}

Because evaluating the partition function exactly is computationally intractable, RBM training typically relies on the contrastive divergence (CD) algorithm[66]. CD performs stochastic approximations of the log-likelihood gradient, effectively minimizing the Kullback–Leibler divergence between the empirical data distribution and the model’s learned distribution. Owing to the absence of intra-layer connections, the conditional activation probabilities update alternately via the logistic sigmoid function σ(x)=(1+e−x)−1\sigma(x) = (1 + e^{-x})^{-1}:

P(hj=1∣g)=σ(∑i=1DgiWij+yj),(32)P(h_j=1\mid\mathbf{g})=\sigma\left(\sum_{i=1}^{D}g_iW_{ij}+y_j\right), \tag*{(32)}
P(gi=1∣h)=σ(∑j=1JWijhj+ui).(33)P(g_i=1\mid\mathbf{h})=\sigma\left(\sum_{j=1}^{J}W_{ij}h_j+u_i\right). \tag*{(33)}

Following training, novel samples are drawn according to the learned probability distribution P(g)P(\mathbf{g}) using the block Gibbs sampling technique[67]. Detailed descriptions of RBM can be found in Refs.[66, 68, 69, 70, 71].

In the domain of electronic structure theory, RBMs offer exceptional representational capacity for capturing intricate many-body electron correlations[72, 73, 74, 75, 76, 77]. When properly trained, an RBM directs configuration sampling toward physically significant sectors of the Hilbert space consisting of determinants possessing the largest configuration interaction (CI) amplitudes. Consequently, the diagonalization subspace drastically shrinks and computational overhead is mitigated. Nevertheless, as unsupervised architectures, RBMs are inherently sensitive to the quality of the initial visible layer training data. In this paper, for PIGen-SQD, we have initialized RBM training with hardware samples augmented by CISD configurations and perturbatively obtained triples (see Appendix F). The second layer of RBM for PIGen-SQD(MAPS) workflow uses MTPS as the initial training data set. The details of the associated hyperparameters used in this work is given in Table III which is same as the original PIGen-SQD work in Ref. [43].

Thresholds and HyperparametersValue
Two-Body Integral Threshold (ϵint\epsilon_{\mathrm{int}})10−1010^{-10}
Blacklist Screening Coeff. (ϵcoeff\epsilon_{\mathrm{coeff}})10−610^{-6}
PIGen-SQD Convergence Threshold (ϵE\epsilon_E)10−510^{-5}
RBM Learning Rate (η\eta)10−310^{-3}
RBM Intrinsic Gibbs Sampling (nGn_G)20
RBM Number of Training Epochs (ntrainn_{\mathrm{train}})3
RBM Generation Gibbs Sampling (ngn_g)1

Table III. Summary of thresholds and hyperparameters used in PIGen-SQD workflow.

Hyperparameters for SQD and S-CORE(MTPS)

The hyperparameters for iterative carryover mechanism of SQD and S-CORE(MTPS) are given in Table IV. The two methods use the same set of hyperparameters, except for the number of samples per batch.

THE LINEAR ASSIGNMENT PROBLEM

To determine the optimal pairing between two distinct sets of equal size, such as mapping initial states to final states, assigning tasks to workers of a company for optimal output or assigning discrete particles to lattice sites, we formulate the system as a Linear Assignment Problem (LAP). The LAP is resolved exactly and deterministically in polynomial time using the Kuhn-Munkres optimization procedure, commonly known as the Hungarian algorithm.

Mathematical Formulation of LAP

Consider a bipartite system comprising two disjoint sets of size nn, denoted U\mathcal{U} and V\mathcal{V}. Let C∈Rn×nC \in\mathbb{R}^{n \times n} denote the cost matrix, where the matrix element cijc_{ij} represents the physical cost, energy, or distance associated with assigning element i∈Ui \in\mathcal{U} to element j∈Vj \in\mathcal{V}.

The objective is to find a bijective mapping that minimizes the global cost function. We introduce a boolean decision matrix X∈{0,1}n×nX \in\{0,1\}^{n \times n}, where xij=1x_{ij}=1 if ii is assigned to jj, and xij=0x_{ij}=0 otherwise. The optimization problem is mathematically defined as:

min⁡X∑i=1n∑j=1ncijxij(34)\min_X \sum_{i=1}^{n}\sum_{j=1}^{n} c_{ij}x_{ij} \tag*{(34)}

subject to the constraints of a perfect bipartite matching (each element in U\mathcal{U} is mapped to exactly one element in V\mathcal{V}, and vice versa):

∑j=1nxij=1∀i∈{1,2,…,n}(35)\sum_{j=1}^{n} x_{ij}=1 \quad\forall i \in\{1,2,\ldots,n\} \tag*{(35)}
∑i=1nxij=1∀j∈{1,2,…,n}(36)\sum_{i=1}^{n} x_{ij}=1 \quad\forall j \in\{1,2,\ldots,n\} \tag*{(36)}

Algorithmic Implementation of Hungarian Method

The Hungarian algorithm exploits the property that adding or subtracting a constant from any row or column of the cost matrix shifts the total objective function by a constant, but preserves the configuration of the optimal assignment. The algorithm systematically reduces the matrix CC to induce a zero-cost perfect matching. The procedure operates in O(n3)O(n^3) time and executes as follows:

a. Matrix Reduction First, we normalize the cost matrix by establishing local minima in every row and column. For each row ii, we subtract the minimum element from all entries in that row:

cij′=cij−min⁡kcik∀j(37)c'_{ij}=c_{ij}-\min_k c_{ik} \quad\forall j \tag*{(37)}

Subsequently, for each column jj in the intermediate matrix C′C', we subtract the column minimum:

cij′′=cij′−min⁡kckj′∀i(38)c''_{ij}=c'_{ij}-\min_k c'_{kj} \quad\forall i \tag*{(38)}

The resulting reduced matrix C′′C'' contains at least one zero in every row and column.

b. Graph Construction and Vertex Cover We define an unweighted bipartite graph G=(U,V,E)\mathcal{G}=(\mathcal{U},\mathcal{V},\mathcal{E}) where an edge (i,j)∈E(i,j)\in\mathcal{E} exists if and only if cij′′=0c''_{ij}=0. By König’s theorem, the size of the maximum bipartite matching in G\mathcal{G} is exactly equal to the minimum vertex cover.

We determine the minimum number of lines—corresponding to a set of rows R\mathcal{R} and a set of columns C\mathcal{C}—required to cover all zero elements in C′′C''. Let k=∣R∣+∣C∣k=|\mathcal{R}|+|\mathcal{C}| be the total number of covering lines.

c. Optimality Test

  • If k=nk=n, the zero elements covered by these lines constitute a perfect matching. The algorithm terminates, and the optimal assignment matrix XX is constructed from a set of nn independent zeros.

  • If k<nk<n, a perfect matching cannot yet be formed using purely zero-cost edges. The matrix must be iteratively updated to introduce new zeros.

d. Matrix Update To iteratively approach optimality without discarding previously discovered partial matchings, we define Δ\Delta as the minimum element in C′′C'' that is not covered by any line in R\mathcal{R} or C\mathcal{C}:

Δ=min⁡i∉R, j∉Ccij′′(39)\Delta=\min_{i\notin\mathcal{R},\,j\notin\mathcal{C}} c''_{ij} \tag*{(39)}

The matrix elements are updated according to the following piecewise transformation:

cij(new)={cij′′−Δ,if i∉R and j∉C (uncovered)cij′′+Δ,if i∈R and j∈C (doubly covered)cij′′,otherwise (singly covered)(40)c_{ij}^{(new)} = \begin{cases} c_{ij}^{\prime\prime} - \Delta, & \text{if } i \notin R \text{ and } j \notin C \text{ (uncovered)} \\ c_{ij}^{\prime\prime} + \Delta, & \text{if } i \in R \text{ and } j \in C \text{ (doubly covered)} \\ c_{ij}^{\prime\prime}, & \text{otherwise (singly covered)} \end{cases} \tag*{(40)}

We substitute C′′C^{\prime\prime} with C(new)C^{(new)} and repeat the procedure from Step 2 until condition k=nk = n is satisfied.

Fragment-to-Supersystem Computational Bit Mapping and Cost Function Construction

The construction of MTPS from configurations obtained from fragment QCSC solvers requires a consistent correspondence between the fragment molecular orbitals and the supersystem molecular orbitals. The correspondence is non-trivial because the supersystem orbitals are not required to retain the same ordering or strict fragment localization as the isolated-fragment orbitals. Orbital mixing, energetic reordering and near degeneracies may therefore render a direct index-by-index correspondence. To obtain a one to one correspondence, the fragment to supersystem mapping is constructed as linear assignment problem (LAP). LAP with Hungarian method is described in the Appendix C. In this section, we are describing the construction of physically motivated cost function supplied to the assignment solver for the Fragment-to-Supersystem Orbital Mapping.

Orbital Overlap Similarity

The spatial correspondence between supersystem MO χ~g\widetilde{\chi}_{g} and fragment MO χβk\chi_{\beta_{k}} is

Sgβk=⟨χ~g∣χβk⟩(41)S_{g\beta_{k}} = \langle\widetilde{\chi}_{g} \mid\chi_{\beta_{k}} \rangle \tag*{(41)}

Plugging in the expansion in Eq. (21) we have

Sgβk=∑ν∑μC~gν⟨ξ~ν∣ξμk⟩Cμkβk(42)S_{g\beta_{k}} = \sum_{\nu} \sum_{\mu} \widetilde{C}_{g\nu} \langle\widetilde{\xi}_{\nu} \mid\xi_{\mu_{k}} \rangle C_{\mu_{k}\beta_{k}} \tag*{(42)}

Here, it is important to note that computationally there is a mismatch between the dimension of C~gν\widetilde{C}_{g\nu} and CμβkC_{\mu\beta_{k}} which is taken care of by padding extra zero coefficients before evaluating the overlap for each block involving a particular kk of the entire SS matrix.

As the global phase of a molecular orbital is arbitrary and consequently any two orbitals differing only by a global phase have identical mapping similarity, we define the spatial similarity factor as the square of the orbital similarity,

Agβk=∣Sgβk∣2(43)A_{g\beta_{k}} = \lvert S_{g\beta_{k}} \rvert^{2} \tag*{(43)}

For normalized orbitals, 0≤Agβk≤10 \leq A_{g\beta_{k}} \leq1 where values approaching unity indicate strong spatial overlap, whereas values close to zero reflect negligible spatial correspondence between the fragment and supersystem molecular orbitals.

Orbital Energy Compatibility

Spatial similarity alone may not always uniquely distinguish the physically appropriate correspondence. This particularly becomes more relevant for groups of closely spaced orbitals having similar spatial character or for orbitals that mix upon formation of the supersystem. Therefore, we introduced orbital energy approximately as an additional compatibility criterion.

For gg-th global orbital with orbital energy ϵg\epsilon_{g} and β\beta-th orbital of the kk-th fragment is ϵβk\epsilon_{\beta_{k}} we defined energy compatibility factor as a Gaussian function

(Δϵ)gβk=e−(ϵg−ϵβk−δϵ‾βk)22σϵ2(44)(\Delta\epsilon)_{g\beta_{k}} = e^{-\frac{\left(\epsilon_{g}-\epsilon_{\beta_{k}}-\overline{\delta\epsilon}_{\beta_{k}}\right)^{2}}{2\sigma_{\epsilon}^{2}}} \tag*{(44)}

here σ\sigma controls the energy window over which fragment and supersystem orbitals are considered energetically compatible. The Gaussian factor satisfies 0≤(Δϵ)gβk≤10 \leq(\Delta\epsilon)_{g\beta_{k}} \leq1 and smoothly penalizes increasingly large energy separations without introducing a hard energy cutoff.

System dependent calibrated energy window

The energy compatibility factor introduced above needs an energy scale σ\sigma that determines how fast the compatibility between a fragment orbital and a supersystem orbital decreases with increase in orbital energy separation. A fixed value of σ\sigma would mean same absolute energy tolerance on chemically and spectrally different systems and thus introduce a system independent scale into the mapping. Instead, we determined σ\sigma directly through the overlap-resolved supersystem spectrum associated with the fragment orbitals. First we made the overlap distribution normalized according to

pgβk=Agβk∑fAfβk,∑gpgβk=1(45)p_{g\beta_{k}} = \frac{A_{g\beta_{k}}}{\sum_{f} A_{f\beta_{k}}}, \qquad\sum_{g} p_{g\beta_{k}} = 1 \tag*{(45)}

where AgβkA_{g\beta_{k}} is defined as equation D3. The weights pgβkp_{g\beta_{k}} therefore identify the regions of the supersystem spectrum that actually carry the character of the fragment orbital χβk\chi_{\beta_{k}}. Now we have the overlap-conditioned mean energy separation as

δϵ‾βk=∑gpgβkδϵgβk(46)\overline{\delta\epsilon}_{\beta_{k}} = \sum_{g} p_{g\beta_{k}} \delta\epsilon_{g\beta_{k}} \tag*{(46)}

where δϵgβk=ϵg−ϵβk\delta\epsilon_{g\beta_{k}} = \epsilon_{g} - \epsilon_{\beta_{k}}, and the corresponding spectral spread as

τβk=[∑gpgβk(δϵgβk−δϵ‾βk)2]1/2(47)\tau_{\beta_{k}} = \left[\sum_{g} p_{g\beta_{k}} \left(\delta\epsilon_{g\beta_{k}} - \overline{\delta\epsilon}_{\beta_{k}}\right)^{2}\right]^{1/2} \tag*{(47)}

The quantity in equation D6 measures a systematic displacement of the fragment orbital relative to the supersystem spectrum (i.e. it finds out location of the overlap-supported supersystem orbitals carrying the character of the fragment orbital), while the quantity of the equation D7 describes the energy range over which orbital characters are spread (the intrinsic width of this overlap-supported distribution).

The single characteristic energy scale for the overall fragment to supersystem mapping problem is obtained from the root mean square of the orbital specific scales,

τspec=[1Nfrag∑βκτβκ2]1/2,(48)\tau_{\mathrm{spec}} = \left[\frac{1}{N_{\mathrm{frag}}}\sum_{\beta\kappa}\tau_{\beta\kappa}^{2}\right]^{1/2}, \tag*{(48)}

where NfragN_{\mathrm{frag}} is the total number of active fragment orbitals included in the mapping. The width of the gaussian energy compatibility factor is then defined by

σE=τspec(49)\sigma_E = \tau_{\mathrm{spec}} \tag*{(49)}

and the energy compatibility factor becomes full defined as

(Δϵ)gβκ=e−(ϵg−ϵβκ−δϵβκE)22σE2(50)(\Delta\epsilon)_{g\beta\kappa} = e^{-\frac{(\epsilon_g-\epsilon_{\beta\kappa}-\delta\epsilon_{\beta\kappa}^{E})^{2}}{2\sigma_E^{2}}} \tag*{(50)}

Combined Orbital Mapping Cost Function

For the overall cost function we combined the spatial similarity factor and energy penalty factor multiplicatively such that the joint compatibility factor is

Qgβκ=Agβκ(Δϵ)gβκ(51)Q_{g\beta\kappa}=A_{g\beta\kappa}(\Delta\epsilon)_{g\beta\kappa} \tag*{(51)}

and the cost function becomes

Jgβκ=1−Qgβκ(52)J_{g\beta\kappa}=1-Q_{g\beta\kappa} \tag*{(52)}

The multiplicative construction is deliberately different from an additive combination of independently weighted spatial and energetic penalties. The two factors here are not competing objectives (whose relative weight are to be adjusted through empirical coefficients), rather they provide complementary conditions for identifying the same physical correspondence i.e. the orbitals should have compatible spatial character and the energy of the spatially related states should be spectrally reasonable. In this aspect, it is important to note that, because

0≤Agβκ≤1,0<(Δϵ)gβκ≤1,(53)0\leq A_{g\beta\kappa}\leq1,\qquad0<(\Delta\epsilon)_{g\beta\kappa}\leq1, \tag*{(53)}

the joint compatibility satisfies

0≤Qgβκ≤Agβκ.(54)0\leq Q_{g\beta\kappa}\leq A_{g\beta\kappa}. \tag*{(54)}

This ensures that energy proximity can never overrule the absence of the spatial compatibility. It can only reduce the compatibility already supported by the spatial similarity or at most keep it preserved. Thus, if

Agβκ→0,(55)A_{g\beta\kappa}\to0, \tag*{(55)}

then

Qgβκ→0andJgβκ→1(56)Q_{g\beta\kappa}\to0 \qquad\text{and} \qquad J_{g\beta\kappa}\to1 \tag*{(56)}

for any value of (Δϵ)gβκ(\Delta\epsilon)_{g\beta\kappa}. An accidental near degeneracy in orbital energy therefore cannot compensate for negligible spatial similarity. On the other hand, when two orbitals exhibit strong spatial similarity,

Agβκ→1⟹Jgβκ→1−(Δϵ)gβκ.(57)A_{g\beta\kappa}\to1 \quad\Longrightarrow\quad J_{g\beta\kappa}\to1-(\Delta\epsilon)_{g\beta\kappa}. \tag*{(57)}

and here the orbital energy difference serves as additional distinguishing factor between the competing high overlap candidates. The cooperative character of the two factors can also be expressed directly through

∂Jgβκ∂Agβκ∣(Δϵ)gβκ=−(Δϵ)gβκ≤0,(58)\left.\frac{\partial J_{g\beta\kappa}}{\partial A_{g\beta\kappa}}\right|_{(\Delta\epsilon)_{g\beta\kappa}}=-(\Delta\epsilon)_{g\beta\kappa}\leq0, \tag*{(58)}

and

∂Jgβκ∂(Δϵ)gβκ∣Agβκ=−Agβκ≤0.(59)\left.\frac{\partial J_{g\beta\kappa}}{\partial(\Delta\epsilon)_{g\beta\kappa}}\right|_{A_{g\beta\kappa}}=-A_{g\beta\kappa}\leq0. \tag*{(59)}

Thus, an increase in either spatial similarity or energy compatibility can only decrease the cost; improvement in one descriptor never produces a penalty in the other. In this precise sense, overlap and energy similarities act cooperatively rather than antagonistically.

Another very important point to note that, the cost can be written as

Jgβκ=(1−Agβκ)+Agβκ(1−(Δϵ)gβκ).(60)J_{g\beta\kappa}=(1-A_{g\beta\kappa})+A_{g\beta\kappa}\left(1-(\Delta\epsilon)_{g\beta\kappa}\right). \tag*{(60)}

which illustrates that the first contribution represents the spatial mismatch, whereas the second contribution represents the energy penalty. But the energy penalty is weighted by the spatial similarity itself. It shows that energetic disagreement is relevant only to the extent that the candidate pair already shares common orbital character.

Mapped Orbitals for FMA Dimer

Figure E1 shows a set of super-system orbitals near the HOMO-LUMO region along with the associated assigned fragment orbitals for monomer 1 and monomer 2 according to the computational bit mapping.

Algebraic Terms for Perturbative Calculations

The perturbative augmentation in the PIGen-SQD framework configurations up to rank three includes the following algebraic terms:

cijkabc←{vijamtmkbc,vieabtjkec}(61)c_{ijk}^{abc}\leftarrow\left\{v_{ij}^{am}t_{mk}^{bc},v_{ie}^{ab}t_{jk}^{ec}\right\} \tag*{(61)}

Here vv and tt represent the two-body integral and MP2 tensor. Note the signs of the algebraic terms are not explicitly shown in these expressions are assumed to be absorbed within each term.

Supplementary Information

Geometry of the Molecular Systems Studied

The Cartesian coordinates of the water dimer, (H2O)2(\mathrm{H}_2\mathrm{O})_2, with an oxygen–oxygen separation of 2.2 Å are reported in Table S1. All coordinates are given in Å.

Atomxx (Å)yy (Å)zz (Å)
O0.63877561-0.014165420.04685218
H1.04765561-0.58046842-0.58807482
H0.918799610.86313358-0.16095782
O-1.555212000.03613800-0.10770500
H-0.61262100-0.04017400-0.04964000
H-1.88509500-0.237998000.73159800

Table S1. Cartesian coordinates of the water dimer (H2O)2(\mathrm{H}_2\mathrm{O})_2 with an oxygen-oxygen separation of 2.2 Å.

The Cartesian coordinates of the formamide dimer at an intermolecular separation of 3.67 Å are reported in Table S2. All coordinates are given in Å.

Atomxx (Å)yy (Å)zz (Å)
C-2.024327540.053336770.00000000
N-1.41581113-1.143387670.00000000
O-1.457103641.147496430.00000000
H-0.38812535-1.20905356-0.00000000
H-1.96870376-1.986156690.00000000
H-3.12868649-0.009514980.00000000
O1.58210808-1.14749705-0.00000000
C2.14932887-0.05333586-0.00000000
N1.540809341.143386880.00000000
H0.513123261.20904999-0.00000000
H3.253687900.009518750.00000000
H2.093700451.986156990.00000000

Table S2. Cartesian coordinates of the formamide dimer at 3.67 Å.

The Cartesian coordinates of the cage-shaped water hexamer, (H2O)6(\mathrm{H}_2\mathrm{O})_6, used in this work are reported in Table S3. All coordinates are given in Å.

Atomxx (Å)yy (Å)zz (Å)
O0.679821.668470.22469
H1.544721.235570.26190
H0.801992.608830.28673
O-0.92115-0.001761.65020
H-0.448060.773311.33150
H-1.76238-0.051711.18135
O0.63241-1.649840.20904
H0.08923-1.222680.89198
H0.53282-2.592140.27953
O-0.424780.02956-1.62376
H-0.096160.81284-1.17004
H0.00023-0.73782-1.22438
O2.74478-0.02765-0.08537
H2.20281-0.788800.15326
H2.94467-0.09143-1.01363
O-2.76330-0.07803-0.30968
H-2.08455-0.14812-0.99436
H-3.307630.67610-0.50478

Table S3. Cartesian coordinates of the cage-shaped water hexamer (H2O)6(\mathrm{H}_2\mathrm{O})_6.

References

  1. [1]J. Čížek. On the correlation problem in atomic and molecular systems: calculation of wavefunction components in ursell-type expansion using quantum-field theoretical methods, J. Chem. Phys. 45, 4256 (1966).DOI
  2. [2]J. Čížek and J. Paldus, Correlation problems in atomic and molecular systems iii. rederivation of the coupled-pair many-electron theory using the traditional quantum chemical methods, Int. J. Quantum Chem. 5, 359 (1971).
  3. [3]T. D. Crawford and H. F. Schaefer, An introduction to coupled cluster theory for computational chemists, Reviews in computational chemistry 14, 33 (2000).DOI
  4. [4]C. D. Sherrill and H. F. Schaefer III, The configuration interaction method: Advances in highly correlated approaches, in Advances in quantum chemistry, Vol. 34 (Elsevier, 1999) pp. 143–269.DOI
  5. [5]S. R. White, Density matrix formulation for quantum renormalization groups, Physical review letters 69, 2863 (1992).pubmed.ncbi.nlm.nih.gov/10046608
  6. [6]A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, A variational eigenvalue solver on a photonic quantum processor, Nature communications 5, 4213 (2014).DOI
  7. [7]M. Cerezo, A. Arrasmith, B. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, et al., Variational quantum algorithms, Nature Reviews Physics 3, 625 (2021).
  8. [8]K. Bharti, A. Cervera-Lierta, T. H. Kyaw, T. Haug, S. Alperin-Lea, A. Anand, M. Degroote, H. Heimonen, J. S. Kottmann, T. Menke, et al., Noisy intermediate-scale quantum algorithms, Reviews of Modern Physics 94, 015004 (2022).DOI
  9. [9]N. H. Stair and F. A. Evangelista, Simulating many-body systems with a projective quantum eigensolver, PRX Quantum 2, 030301 (2021).DOI
  10. [10]C. Patra, S. Halder, and R. Maitra, Projective quantum eigensolver via adiabatically decoupled subsystem evolution: A resource efficient approach to molecular energetics in noisy quantum computers, The Journal of Chemical Physics 160 (2024).arxiv.org/abs/2403.08519
  11. [11]C. Patra, D. Mukherjee, S. Halder, D. Mondal, and R. Maitra, Toward a resource-optimized dynamic quantum algorithm via non-iterative auxiliary subspace corrections, The Journal of Chemical Physics 161 (2024).
  12. [12]K. Kanno, M. Kohda, R. Imai, S. Koh, K. Mitarai, W. Mizukami, and Y. O. Nakagawa, Quantum-selected configuration interaction: Classical diagonalization of hamiltonians in subspaces selected by quantum computers, arXiv preprint arXiv:2302.11320 (2023).DOI
  13. [13]E. R. Anschuetz and B. T. Kiani, Quantum variational algorithms are swapped with traps, Nature Communications 13, 7760 (2022).
  14. [14]H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, An adaptive variational algorithm for exact molecular simulations on a quantum computer, Nature communications 10, 3007 (2019).DOI
  15. [15]S. E. Smart and D. A. Mazziotti, Quantum solver of contracted eigenvalue equations for scalable molecular simulations on quantum computing devices, Physical Review Letters 126, 070504 (2021).DOI
  16. [16]H. R. Grimsley, G. S. Barron, E. Barnes, S. E. Economou, and N. J. Mayhall, Adaptive, problem-tailored variational quantum eigensolver mitigates rough parameter landscapes and barren plateaus, npj Quantum Information 9, 19 (2023).DOI
  17. [17]I. G. Ryabinkin, T.-C. Yen, S. N. Genin, and A. F. Izmaylov, Qubit coupled cluster method: a systematic approach to quantum chemistry on a quantum computer, Journal of chemical theory and computation 14, 6317 (2018).arxiv.org/abs/1809.03827
  18. [18]N. P. Bauman, E. J. Bylaska, S. Krishnamoorthy, G. H. Low, N. Wiebe, C. E. Granade, M. Roetteler, M. Troyer, and K. Kowalski, Downfolding of many-body hamiltonians using active-space models: Extension of the subsystem embedding sub-algebras approach to unitary coupled cluster formalisms, The Journal of Chemical Physics 151, 014107 (2019).DOI
  19. [19]Y. Matsuzawa and Y. Kurashige, Jastrow-type decomposition in quantum chemistry for low-depth quantum circuits, Journal of chemical theory and computation 16, 944 (2020).DOI
  20. [20]I. G. Ryabinkin, R. A. Lang, S. N. Genin, and A. F. Izmaylov, Iterative qubit coupled cluster approach with efficient screening of generators, Journal of chemical theory and computation 16, 1055 (2020).DOI
  21. [21]Y. S. Yordanov, D. R. Arvidsson-Shukur, and C. H. Barnes, Efficient quantum circuits for quantum computational chemistry, Physical Review A 102, 062612 (2020).DOI
  22. [22]K. Kowalski, Dimensionality reduction of the many-body problem using coupled-cluster subsystem flow equations: Classical and quantum computing perspective, Physical Review A 104, 032804 (2021).DOI
  23. [23]D. Mondal, D. Halder, S. Halder, and R. Maitra, Development of a compact Ansatz via operator commutativity screening: Digital quantum simulation of molecular systems, The Journal of Chemical Physics 159, 014105 (2023).DOI
  24. [24]D. Halder, V. S. Prasannaa, and R. Maitra, Dual exponential coupled cluster theory: Unitary adaption, implementation in the variational quantum eigensolver framework and pilot applications, The Journal of Chemical Physics 157, 174117 (2022).
  25. [25]D. Claudino, B. Peng, N. P. Bauman, K. Kowalski, and T. S. Humble, Improving the accuracy and efficiency of quantum connected moments expansions, Quantum Science and Technology 6, 034012 (2021).DOI
  26. [26]D. Halder, S. Halder, D. Mondal, C. Patra, A. Chakraborty, and R. Maitra, Corrections beyond coupled cluster singles and doubles through selected generalized rank-two operators: digital quantum simulation of strongly correlated systems, Journal of Chemical Sciences 135, 41 (2023).DOI
  27. [27]D. Halder, D. Mondal, and R. Maitra, Measurement free approach towards dynamic construction of a shallow depth ansatz for molecular energetics in noisy quantum computers, arXiv preprint arXiv:2311.09895 (2023).
  28. [28]K. Kowalski and N. P. Bauman, Quantum flow algorithms for simulating many-body systems on quantum computers, Physical Review Letters 131, 200601 (2023).arxiv.org/abs/2305.05168
  29. [29]H. G. Burton, D. Marti-Dafcik, D. P. Tew, and D. J. Wales, Exact electronic states with shallow quantum circuits from global optimisation, npj Quantum Information 9, 75 (2023).DOI
  30. [30]H. G. Burton, Accurate and gate-efficient quantum ansätze for electronic states without adaptive optimization, Physical Review Research 6, 023300 (2024).DOI
  31. [31]D. Halder, D. Mondal, and R. Maitra, Noise-independent route toward the genesis of a compact ansatz for molecular energetics: A dynamic approach, The Journal of Chemical Physics 160 (2024).
  32. [32]S. Halder, C. Patra, D. Mondal, and R. Maitra, Machine learning aided dimensionality reduction toward a resource efficient projective quantum eigensolver: Formal development and pilot applications, The Journal of Chemical Physics 158 (2023).DOI
  33. [33]C. Patra and R. Maitra, Energy landscape plummeting in variational quantum eigensolver: Subspace optimization, non-iterative corrections, and generator-informed initialization for improved quantum efficiency, The Journal of Chemical Physics 163, 024112 (2025).
  34. [34]M. Haidar, O. Adjoua, S. Badreddine, A. Peruzzo, and J.-P. Piquemal, Non-iterative disentangled unitary coupled-cluster based on lie-algebraic structure, Quantum Science and Technology 10, 025031 (2025).DOI
  35. [36]J. Robledo-Moreno, M. Motta, H. Haas, A. Javadi-Abhari, P. Jurcevic, W. Kirby, S. Martiel, K. Sharma, S. Sharma, T. Shirakawa, et al., Chemistry beyond the scale of exact diagonalization on a quantum-centric supercomputer, Science Advances 11, eadu9991 (2025).
  36. [37]A. Shajan, D. Kaliakin, A. Mitra, J. Robledo Moreno, Z. Li, M. Motta, C. Johnson, A. A. Saki, S. Das, I. Sitdikov, et al., Toward quantum-centric simulations of extended molecules: Sample-based quantum diagonalization enhanced with density matrix embedding theory, Journal of Chemical Theory and Computation 21, 6801 (2025).DOI
  37. [38]A. K. Patra, A. KSV, R. Bhat, R. Maitra, et al., Quantum simulation of ligand-like molecules through sample-based quantum diagonalization in density matrix embedding framework, The Journal of Chemical Physics 165 (2026).
  38. [39]Q. Wang, K. J. Sung, R. DaÃŽCunha, M. R. Hermes, T. Gujarati, Y. Kawashima, Y.-y. Ohnishi, G. O. Jones, M. Motta, and L. Gagliardi, Localized sample-based quantum diagonalization for strongly correlated chemistry, Proceedings of the National Academy of Sciences 123, e2603914123 (2026).DOI
  39. [40]A. Shajan, D. Kaliakin, F. Liang, T. Pellegrini, H. Doga, S. Bhowmik, S. Das, A. Mezzacapo, M. Motta, and K. M. Merz Jr, Molecular quantum computations on a protein, Journal of Chemical Theory and Computation 22, 6041 (2026).
  40. [41]K. M. Merz Jr, A. Shajan, D. Kaliakin, F. Liang, Y. Otsuka, T. Shirakawa, L. Broers, H. Xu, M. Tsuji, M. Sato, et al., Crossing the 12,000-atom barrier with heterogeneous quantum-classical supercomputing: quantum chemistry of protein-ligand complexes, arXiv preprint arXiv:2605.01138 (2026).arxiv.org/abs/2605.01138
  41. [42]K. Yamamoto, R. Masui, T. Nakajima, M. Tsuji, M. Sato, P. Schow, E. Heidemann, M. Burke, P. Seitz, O. J. Backhouse, et al., Quantum-hpc hybrid computation of biomolecular excited-state energies, arXiv preprint arXiv:2601.15677 (2026).arxiv.org/abs/2601.15677
  42. [43]C. Patra, D. Mondal, S. Halder, D. Halder, M. R. Laskar, R. Goel, and R. Maitra, Accelerated quantum-centric supercomputing through perturbation-theoretic measures and generative machine learning, Quantum Science and Technology 11, 035075 (2026).DOI
  43. [44]P. Jordan and E. P. Wigner, Über das paulsche äquivalenzverbot (Springer, 1993).DOI
  44. [45]M. Motta, K. J. Sung, K. B. Whaley, M. Head-Gordon, and J. Shee, Bridging physical intuition and hardware efficiency for correlated electronic states: the local unitary cluster jastrow ansatz for electronic structure, Chemical Science 14, 11213 (2023).DOI
  45. [46]H. Belgallali, T. V. Camp, R. Pradeep, S. Das, N. Anand, and R. LaRose, Efficient classical simulation of large-scale unitary cluster jastrow circuits (2026), arXiv:2607.21337 [quant-ph].
  46. [47]E. Koridze, S. Sen, L. Visscher, and S. Polla, Fragpt2: Multifragment wave function embedding with perturbative interactions, Journal of Chemical Theory and Computation 21, 655 (2025).
  47. [48]K. Kitaura, E. Ikeo, T. Asada, T. Nakano, and M. Uebayasi, Fragment molecular orbital method: an approximate computational method for large molecules, Chemical Physics Letters 313, 701 (1999).DOI
  48. [49]M. A. Collins and R. P. A. Bettens, Energy-based molecular fragmentation methods, Chemical Reviews 115, 5607 (2015).DOI
  49. [50]G. Knizia and G. K.-L. Chan, Density matrix embedding: a simple alternative to dynamical mean-field theory, Physical Review Letters 109, 186404 (2012).pubmed.ncbi.nlm.nih.gov/23215304
  50. [51]G. Knizia and G. K.-L. Chan, Density matrix embedding: A strong-coupling quantum embedding theory, Journal of Chemical Theory and Computation 9, 1428 (2013).DOI
  51. [52]H. W. Kuhn, The hungarian method for the assignment problem, Naval research logistics quarterly 2, 83 (1955).
  52. [53]H. Chakraa, F. Guérin, E. Leclercq, and D. Lefebvre, Optimization techniques for multi-robot task allocation problems: Review on the state-of-the-art, Robotics and Autonomous Systems 168, 104492 (2023).DOI
  53. [54]M. W. Schmidt, K. K. Baldridge, J. A. Boatz, S. T. Elbert, M. S. Gordon, J. H. Jensen, S. Koseki, N. Mat­sunaga, K. A. Nguyen, S. Su, et al., General atomic and molecular electronic structure system, Journal of computational chemistry 14, 1347 (1993).DOI
  54. [55]S. Sharma, A. A. Holmes, G. Jeanmairet, A. Alavi, and C. J. Umrigar, Semistochastic heat-bath configuration interaction method: Selected configuration interaction with semistochastic perturbation theory, Journal of chemical theory and computation 13, 1595 (2017).DOI
  55. [56]A. A. Holmes, N. M. Tubman, and C. Umrigar, Heat-bath configuration interaction: An efficient selected configuration interaction algorithm inspired by heat-bath sampling, Journal of chemical theory and computation 12, 3674 (2016).DOI
  56. [57]Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo, Z. Li, J. Liu, J. D. McClain, E. R. Sayfutyaro­va, S. Sharma, et al., PySCF: the python-based simulations of chemistry framework, Wiley Interdisciplinary Reviews: Computational Molecular Science 8, e1340 (2018).arxiv.org/abs/1701.08223
  57. [58]A. A. Saki, S. Barison, B. Fuller, J. R. Garrison, J. R. Glick, C. Johnson, A. Mezzacapo, J. Robledo-Moreno, M. Rossmannek, P. Schweigert, I. Sitdikov, and K. J. Sung, Qiskit addon: sample-based quantum diagonalization, https://github.com/Qiskit/qiskit-addon-sqd (2024).
  58. [59]The ffsim developers, ffsim: Faster simulations of fermionic quantum circuits, https://github.com/qiskit-community/ffsim.
  59. [60]P. Virtanen et al., SciPy 1.0: Fundamental algorithms for scientific computing in Python, Nature Methods 17, 261 (2020).
  60. [61]D. F. Crouse, On implementing 2d rectangular assignment algorithms, IEEE Transactions on Aerospace and Electronic Systems 52, 1679 (2016).DOI
  61. [62]R. Jonker and A. Volgenant, Improving the hungarian assignment algorithm, Operations Research Letters 5, 171 (1986).DOI
  62. [63]R. Jonker and A. Volgenant, A shortest augmenting path algorithm for dense and sparse linear assignment problems, Computing 38, 325 (1987).
  63. [64]X. Zeng, M. Gong, B. Kan, Y. Fan, H. Ma, J. Cai, Y. Liu, N. Zhou, T. Jiang, S. Guo, et al., Transformer refined quantum sampling for strongly correlated electronic structure, arXiv preprint arXiv:2605.24617 (2026).arxiv.org/abs/2605.24617
  64. [65]S. Halder, C. Patra, and R. Maitra, Quantum wavefunction augmentation via variational autoencoders, arXiv preprint arXiv:2609.07972 (2026).arxiv.org/abs/2609.07972
  65. [66]G. E. Hinton, A practical guide to training restricted boltzmann machines, in Neural Networks: Tricks of the Trade: Second Edition (Springer, 2012) pp. 599–619.DOI
  66. [67]D. J. MacKay, Information theory, inference and learning algorithms (Cambridge university press, 2003).
  67. [68]N. Zhang, S. Ding, J. Zhang, and Y. Xue, An overview on restricted boltzmann machines, Neurocomputing 275, 1186 (2018).DOI
  68. [69]G. E. Hinton, Training products of experts by minimizing contrastive divergence, Neural computation 14, 1771 (2002).pubmed.ncbi.nlm.nih.gov/12180402
  69. [70]G. E. Hinton and R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks, science 313, 504 (2006).pubmed.ncbi.nlm.nih.gov/16873662
  70. [71]R. G. Melko, G. Carleo, J. Carrasquilla, and J. I. Cirac, Restricted boltzmann machines in quantum physics, Nature Physics 15, 887 (2019).DOI
  71. [72]B. Herzog, B. Casier, S. Lébègue, and D. Rocca, Solving the schrÃudinger equation in the configuration space with generative machine learning, Journal of Chemical Theory and Computation 19, 2484 (2023).
  72. [73]R. Xia and S. Kais, Quantum machine learning for electronic structure calculations, Nature communications 9, 4195 (2018).
  73. [74]V. Singh, R. Gupta, M. Sajjan, F. Remacle, R. D. Levine, and S. Kais, Maximal entropy formalism and the restricted boltzmann machine, The Journal of Physical Chemistry A.
  74. [75]S. Halder, A. Dey, C. Shrikhande, and R. Maitra, Machine learning assisted construction of a shallow depth dynamic ansatz for noisy quantum hardware, Chemical Science 15, 3279 (2024).
  75. [76]S. Halder, K. Anand, and R. Maitra, Construction of chemistry-inspired dynamic ansatz utilizing generative machine learning, The Journal of Physical Chemistry A.
  76. [77]J. I. Hernandez-Martinez, A. Mendez-Vazquez, G. Rodriguez-Hernandez, and S. L. Juarez-Osorio, Configuration interaction guided sampling with interpretable restricted boltzmann machine, Machine Learning: Science and Technology (2025).

Paper details

vlnc:W_kr93hce2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

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