CoChat

Limitations of quantum hardware for molecular energy estimation using VQE

Open access
Carreras, Abel, Orús, Román, Casanova, David · 2026 · Physical Chemistry Chemical Physics

Published version of a preprint

The DOI matches its Crossref record; 38 of the 38 references it cites were found in a registry or an index; 10 of the 16 sources its claims cite were checked for support.

Abstract

Variational quantum eigensolvers (VQEs) are among the most promising quantum algorithms for solving electronic structure problems in quantum chemistry, particularly during the noisy intermediate-scale quantum (NISQ) era. In this study, we investigate the capabilities and limitations of VQE algorithms implemented on current quantum hardware for determining molecular ground-state energies, focusing on the adaptive derivative-assembled pseudo-Trotter ansatz VQE (ADAPT-VQE). To address the significant computational challenges posed by molecular Hamiltonians, we explore various well known strategies to simplify the Hamiltonian, optimize the ansatz, and improve classical parameter optimization through modifications of the COBYLA optimizer. These enhancements are integrated into a tailored quantum computing implementation designed to minimize the circuit depth and computational cost. Using benzene as a benchmark system, we demonstrate the application of these optimizations on an IBM quantum computer. Despite these improvements, our results highlight the limitations imposed by current quantum hardware, particularly the impact of quantum noise on state preparation and energy measurement. The noise levels in today's devices prevent meaningful evaluations of molecular Hamiltonians with sufficient accuracy to produce reliable quantum chemical insights. Finally, we extrapolate the requirements for future quantum hardware to enable practical and scalable quantum chemistry calculations using VQE algorithms. This work provides an assessment of current quantum algorithms for molecular modeling on real quantum hardware, highlighting the impact of noise and hardware limitations on the achievable accuracy.

Overview

CoChat Summary

TL;DR

The paper studies how far VQE/ADAPT-VQE can be pushed on current IBM hardware for benzene, after simplifying the Hamiltonian, ansatz, optimizer, and circuit layout. Even with these optimizations, the authors find the hardware noise and coherence limits are too severe to reach reliable chemical accuracy, and they estimate coherence times would need to improve by about 50–100× or even two orders of magnitude.

Key findings

  • For benzene, using a four-electron, four-orbital active space reduces the problem to 8 qubits; the effective Hamiltonian has 1057 terms before compression and 233 terms after compression.
  • Hamiltonian compression with a threshold of 10^-2 Ha gives a 78% reduction in Hamiltonian terms while keeping the ground-state energy within about 10^-3 Ha of the full calculation.
  • Symmetry restriction of the ADAPT operator pool cuts the single- and double-excitation pool from 14 to 6 operators (57% reduction), and the paper reports that 6 ADAPT iterations/operators are enough to reach chemical accuracy in the ideal simulation, compared with 64 operators for UCCSD.
  • On IBM-Torino, even with 30 independent runs per iteration, 9000 shots per sample, and mitigation (ZNE + TREX), the measured energies still show large errors; the authors state that errors often exceed the 53.9 mHa correlation energy and remain well above chemical accuracy (1 kcal/mol).
  • Noise simulations indicate that achieving errors below 1 mHa for up to 7 operators would require coherence times of roughly α ≈ 100 relative to typical current hardware values, and full Qubit-ADAPT simulations break down for α < 5.

Methods

  • The study uses benzene as the benchmark system, with geometry from PubChem, molecular orbitals from PySCF, a STO-3G basis, and a Jordan–Wigner mapping.
  • The authors use an active space of 4 electrons in 4 orbitals and compare fermionic ADAPT-VQE with Qubit-ADAPT-VQE; the quantum hardware experiments are run on IBM-Torino.
  • Main analyses include Hamiltonian compression, natural vs canonical orbitals, symmetry-restricted operator pools, staircase circuit optimization, layout optimization, modified COBYLA optimization, hardware runs with ZNE/TREX, and thermal-relaxation noise simulations to estimate coherence requirements.

Limitations

  • The paper explicitly says that current quantum hardware noise prevents meaningful and reliable molecular energy estimation for medium-sized molecular Hamiltonians, and that full VQE optimization is unfeasible on present devices.
  • The authors state that their hardware experiment is limited to single-point energy evaluations; gradient-based operator selection and amplitude optimization are done classically, so the quantum device is not used for a full end-to-end VQE loop.
  • Evident from the text: results are shown for benzene and a few supplementary examples, so the study’s empirical validation is system-specific and limited in scope (benzene is the main benchmark).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1039/d5cp03907j

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
From the PDF
Fig. 2 (a) Ground state energy and (b) energy evaluations of benzene molecule along the qubitADAPT iterations using COBYLA (green) and Mod-COBYLA (blue). Each point corresponds to the average of 30 simulations using 500 shots. Default tolerance in standard COBYLA is 10 -3 Ha.
Fig. 2 (a) Ground state energy and (b) energy evaluations of benzene molecule along the qubitADAPT iterations using COBYLA (green) and Mod-COBYLA (blue). Each point corresponds to the average of 30 simulations using 500 shots. Default tolerance in standard COBYLA is 10 -3 Ha.From the PDF
Fig. 3 Circuit corresponding to the staircase algorithm for (a) a single exponential operator with one Pauli operator σ z , and (b) a Pauli string ( σ 1 z ⊗ σ 2 z ⊗ σ 3 x ⊗ σ 4 z ), with superindices referring to the qubit index. (c) Two different orientations of the CNOT gates in the staircase algorithm: standard and reverse.
Fig. 3 Circuit corresponding to the staircase algorithm for (a) a single exponential operator with one Pauli operator σ z , and (b) a Pauli string ( σ 1 z ⊗ σ 2 z ⊗ σ 3 x ⊗ σ 4 z ), with superindices referring to the qubit index. (c) Two different orientations of the CNOT gates in the staircase algorithm: standard and reverse.From the PDF
Image 4
From the PDF
Fig. 5 Energy errors (in Ha) of 30 individual experimental measures (a) and in average (b) in the evaluation of benzene ground state energy calculated using one exponential operator with the best and worst layout. Vertical bars in (b) indicate deviation error bars. The energy error is computed as the evaluated energy using IBM-torino with 9000 shots with respect to the exact energy computed in a classical computer.
Fig. 5 Energy errors (in Ha) of 30 individual experimental measures (a) and in average (b) in the evaluation of benzene ground state energy calculated using one exponential operator with the best and worst layout. Vertical bars in (b) indicate deviation error bars. The energy error is computed as the evaluated energy using IBM-torino with 9000 shots with respect to the exact energy computed in a classical computer.From the PDF
Fig. 6 (a) Energy evaluations (in Ha) for the Qubit-ADAPT ansatz with 7 qubit operators. (b) Error with respect to analytical ansatz energy (in Ha) for the Qubit-ADAPT ans¨ atze with 0-7 operators of benzene ground state with IBM-Torino (9000 shots) with (green) and without (blue) error mitigation.
Fig. 6 (a) Energy evaluations (in Ha) for the Qubit-ADAPT ansatz with 7 qubit operators. (b) Error with respect to analytical ansatz energy (in Ha) for the Qubit-ADAPT ans¨ atze with 0-7 operators of benzene ground state with IBM-Torino (9000 shots) with (green) and without (blue) error mitigation.From the PDF
Fig. 7 (a) Energy error (horitzontal line at 1 mHa) (in Ha) of Qubit-ADAPT ans¨ atze with different number of excitation operators for the ground state of benzene as a function of the noise scaling factor ( α ). Simulations carried out using the Qiskit Aer simulator. Each data point corresponds to the average of six computed replicates obtained with 9000 shots each. (b) Decoherence factor necessary to achieve errors smaller than 1 mHa as a function of the number of qubit operators [extrapolated from (a)].
Fig. 7 (a) Energy error (horitzontal line at 1 mHa) (in Ha) of Qubit-ADAPT ans¨ atze with different number of excitation operators for the ground state of benzene as a function of the noise scaling factor ( α ). Simulations carried out using the Qiskit Aer simulator. Each data point corresponds to the average of six computed replicates obtained with 9000 shots each. (b) Decoherence factor necessary to achieve errors smaller than 1 mHa as a function of the number of qubit operators [extrapolated from (a)].From the PDF
Fig. 8 (a) Energy errors along ADAPT iterations (a) and as a function of the coherence parameter α (b) of full Qubit-ADAPT simulations with thermal relaxation noise for the ground state of benzene. Values in (b) correspond to the lowest errors in panel (a). All errors given with respect to the analytical Qubit-ADAPT energy with eight operators.
Fig. 8 (a) Energy errors along ADAPT iterations (a) and as a function of the coherence parameter α (b) of full Qubit-ADAPT simulations with thermal relaxation noise for the ground state of benzene. Values in (b) correspond to the lowest errors in panel (a). All errors given with respect to the analytical Qubit-ADAPT energy with eight operators.From the PDF
Figure S1 Analysis of the impact of Hamiltonian compression on the number of Hamiltonian terms and the ground state energy. On the left, we show the number of Hamiltonian terms as a function of the active space size for different Hamiltonian with three compression thresholds (1e-2, 1e-3, and 1e-4), compared to the uncompressed Hamiltonian. On the right, we plot the ground state (GS) energy as a function of the number of Hamiltonian terms. The GS energies have been computed using the DMRG method (with a bond dimension of 200) for three different compression thresholds (1e-3, 1e-4, and 1e-5). The Hamiltonian is expressed in the basis of natural Orbitals.
Figure S1 Analysis of the impact of Hamiltonian compression on the number of Hamiltonian terms and the ground state energy. On the left, we show the number of Hamiltonian terms as a function of the active space size for different Hamiltonian with three compression thresholds (1e-2, 1e-3, and 1e-4), compared to the uncompressed Hamiltonian. On the right, we plot the ground state (GS) energy as a function of the number of Hamiltonian terms. The GS energies have been computed using the DMRG method (with a bond dimension of 200) for three different compression thresholds (1e-3, 1e-4, and 1e-5). The Hamiltonian is expressed in the basis of natural Orbitals.From the PDF
Figure S2 Analysis of the energy error with respect to different Hamiltonian compression thresholds (1e-2, 1e-3, and 1e-4). On the left, we show the ground state energy error as a function of the active space size. On the right, we present the GS energy error as a function of the number of Hamiltonian terms. The ground state energy energies have been computed using the DMRG method with a bond dimension of 200. The Hamiltonian is expressed in the basis of natural Orbitals.
Figure S2 Analysis of the energy error with respect to different Hamiltonian compression thresholds (1e-2, 1e-3, and 1e-4). On the left, we show the ground state energy error as a function of the active space size. On the right, we present the GS energy error as a function of the number of Hamiltonian terms. The ground state energy energies have been computed using the DMRG method with a bond dimension of 200. The Hamiltonian is expressed in the basis of natural Orbitals.From the PDF
Figure S3 Analysis of the impact of Hamiltonian compression on the number of Hamiltonian terms and the ground state energy. On the left, we show the number of Hamiltonian terms as a function of the active space size for different Hamiltonian with three compression thresholds (1e-2, 1e-3, and 1e-4), compared to the uncompressed Hamiltonian. On the right, we plot the ground state (GS) energy as a function of the number of Hamiltonian terms. The GS energies have been computed using the DMRG method (with a bond dimension of 200) for three different compression thresholds (1e-3, 1e-4, and 1e-5). The Hamiltonian is expressed in the basis of canonical orbitals.
Figure S3 Analysis of the impact of Hamiltonian compression on the number of Hamiltonian terms and the ground state energy. On the left, we show the number of Hamiltonian terms as a function of the active space size for different Hamiltonian with three compression thresholds (1e-2, 1e-3, and 1e-4), compared to the uncompressed Hamiltonian. On the right, we plot the ground state (GS) energy as a function of the number of Hamiltonian terms. The GS energies have been computed using the DMRG method (with a bond dimension of 200) for three different compression thresholds (1e-3, 1e-4, and 1e-5). The Hamiltonian is expressed in the basis of canonical orbitals.From the PDF
Figure S4 Analysis of the energy error with respect to different Hamiltonian compression thresholds (1e-2, 1e-3, and 1e-4). On the left, we show the ground state energy error as a function of the active space size. On the right, we present the GS energy error as a function of the number of Hamiltonian terms. The ground state energy energies have been computed using the DMRG method with a bond dimension of 200. The Hamiltonian is expressed in the basis of canonical Orbitals.
Figure S4 Analysis of the energy error with respect to different Hamiltonian compression thresholds (1e-2, 1e-3, and 1e-4). On the left, we show the ground state energy error as a function of the active space size. On the right, we present the GS energy error as a function of the number of Hamiltonian terms. The ground state energy energies have been computed using the DMRG method with a bond dimension of 200. The Hamiltonian is expressed in the basis of canonical Orbitals.From the PDF
Figure S5 Ground state energy (in Hartrees) as a function of the number of iterations for the simulation of the benzene molecule with fermion (green) and qubit (blue) ADAPT-VQE algorithm.
Figure S5 Ground state energy (in Hartrees) as a function of the number of iterations for the simulation of the benzene molecule with fermion (green) and qubit (blue) ADAPT-VQE algorithm.From the PDF
Figure S6 Comparison of the computational cost of VQE algorithms using fermion vs qubit operators for benzene molecule. On the left there is a break down of the total number of CNOTS used for all circuits in the full simulation. We separate this number in energy evaluations and gradient evaluations. On the right there is an analysis of the circuit depths of the circuits used to evaulate the energy and the gradients. We show the maxium circuit depth and the average circuit depth.
Figure S6 Comparison of the computational cost of VQE algorithms using fermion vs qubit operators for benzene molecule. On the left there is a break down of the total number of CNOTS used for all circuits in the full simulation. We separate this number in energy evaluations and gradient evaluations. On the right there is an analysis of the circuit depths of the circuits used to evaulate the energy and the gradients. We show the maxium circuit depth and the average circuit depth.From the PDF
Figure S7 Comparison of the circuit depth (of the circuit used to evaluate the energy) vs energy obtained from an ADAPT-VQE simulation (exact). On the left fermion-ADAPTVQE, on the right qubit-ADAPT-VQE.
Figure S7 Comparison of the circuit depth (of the circuit used to evaluate the energy) vs energy obtained from an ADAPT-VQE simulation (exact). On the left fermion-ADAPTVQE, on the right qubit-ADAPT-VQE.From the PDF
Image 16
From the PDF
Image 17
From the PDF
Figure S9 (a) Ground state energy and (b) energy evaluations of benzene molecule along the qubit-ADAPT iterations using COBYLA (green) and COBYLA-MOD (blue). Each point corresponds to the average of 30 simulations using 1000 shots. Default tolerance in standard COBYLA is 10 -3 Hartrees.
Figure S9 (a) Ground state energy and (b) energy evaluations of benzene molecule along the qubit-ADAPT iterations using COBYLA (green) and COBYLA-MOD (blue). Each point corresponds to the average of 30 simulations using 1000 shots. Default tolerance in standard COBYLA is 10 -3 Hartrees.From the PDF
Figure S11 Molecule H 4 ., Comparison of the optimization using COBYLA and COBYLAMOD using a value of 0.1 for rhobeg. Number of shots is 1000. Default tolerance is 1e-3 Ha.
Figure S11 Molecule H 4 ., Comparison of the optimization using COBYLA and COBYLAMOD using a value of 0.1 for rhobeg. Number of shots is 1000. Default tolerance is 1e-3 Ha.From the PDF
Figure S12 Molecule H 4 . Distribution of the energy of the converged ground state. Comparison of the optimization using COBYLA and COBYLA-MOD. Number of shots is 1000. Default tolerance is 1e-3 Ha.
Figure S12 Molecule H 4 . Distribution of the energy of the converged ground state. Comparison of the optimization using COBYLA and COBYLA-MOD. Number of shots is 1000. Default tolerance is 1e-3 Ha.From the PDF
Figure S13 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Fermion vs qubit operators using a value of 0.1 for rhobeg. Number of shots is 1000. Energy as a function of the iterations (left) and evaluations of the energy (right)
Figure S13 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Fermion vs qubit operators using a value of 0.1 for rhobeg. Number of shots is 1000. Energy as a function of the iterations (left) and evaluations of the energy (right)From the PDF
Figure S14 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Fermion vs qubit operators using a value of 0.1 for rhobeg. Number of shots is 1000
Figure S14 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Fermion vs qubit operators using a value of 0.1 for rhobeg. Number of shots is 1000From the PDF
Figure S15 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Exact energy evaluation vs ADAPT iterations (left) and energy error with respect to FCI (right)
Figure S15 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Exact energy evaluation vs ADAPT iterations (left) and energy error with respect to FCI (right)From the PDF
Figure S16 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Energy evaluations vs ADAPT iterations
Figure S16 Molecule H 4 . Comparison of the optimization using COBYLA and COBYLAMOD. Energy evaluations vs ADAPT iterationsFrom the PDF
Figure S17 Maximum job size accepted by various IBM quantum computers, measured as the product of the number of two-qubit gates and the number of measured Pauli operators in the observable. IBM Brussels was tested 3 times
Figure S17 Maximum job size accepted by various IBM quantum computers, measured as the product of the number of two-qubit gates and the number of measured Pauli operators in the observable. IBM Brussels was tested 3 timesFrom the PDF
Figure S18 Energy evaluations (in Ha) with Qubit-ADAPT ans¨ atze with 0-7 qubit operators measured in the IBM Torino computer using 9000 shots per sample. The coefficients of the ansatz are obtained from an exact calculation. Horizontal lines indicate the exact HF (green) and FCI with 4 active orbitals and 4 active electrons (gray). The energy difference between HF and FCI (correlation energy) is 53.9 mHa. Error bars corresponds to 2 σ .
Figure S18 Energy evaluations (in Ha) with Qubit-ADAPT ans¨ atze with 0-7 qubit operators measured in the IBM Torino computer using 9000 shots per sample. The coefficients of the ansatz are obtained from an exact calculation. Horizontal lines indicate the exact HF (green) and FCI with 4 active orbitals and 4 active electrons (gray). The energy difference between HF and FCI (correlation energy) is 53.9 mHa. Error bars corresponds to 2 σ .From the PDF

Verification

Claims the paper makes

8 Valid · 3 Not valid · 4 Not checked · 96 Own results and interpretations

64 more claims

Create a free account to see all 111 claims and to verify the 4 references still open.

See all claims — free account
  • Probably not a substantial claim. 16
  • It may state two separate claims as one. 28
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 1

  1. Sun, 2018, Wiley Interdiscip. Rev.: Comput. Mol. Sci., 8, e1340

Cited by

See all 5

Metadata

Authors
Carreras, A., Orús, R., & Casanova, D.
Affiliations
  • Donostia International Physics Center (ES)
  • Ikerbasque (ES)
Corresponding author
Carreras, Abel
Venue
Physical Chemistry Chemical Physics
Published
2026
Volume
28
Issue
4
Pages
2834-2846
ISSN
  • 1463-9076 (ISSN-L)
  • 1463-9084
OpenAlex
W4416074512
Cited by
5
References
29
Citations by year
2026: 5
FWCI
17.32
Type
journal article
Language
en
Open access
Locations
  • Physical Chemistry Chemical Physics (publishedVersion)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
  • Zenodo (CERN European Organization for Nuclear Research) (cc-by)
Funders
  • Ministerio de Ciencia, Innovación y Universidades (PID2022-136231NB-I00)
  • Eusko Jaurlaritza
  • Ikerbasque, Basque Foundation for Science
  • European Regional Development Fund
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Quantum Computing Algorithms and Architecture
  • Physical Sciences › Physics and Astronomy › Atomic and Molecular Physics, and Optics › Advanced Chemical Physics Studies
  • Physical Sciences › Physics and Astronomy › Atomic and Molecular Physics, and Optics › Quantum many-body systems
Keywords
variational quantum eigensolver, ADAPT-VQE, NISQ devices, quantum noise, benzene, circuit depth reduction, IBM quantum hardware

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.