Simulating strongly correlated electrons in molecules and materials remains a central goal for quantum computing. A promising strategy uses quantum processors to prepare trial wavefunctions that guide classical quantum Monte Carlo (QMC). However, the enormous computational cost of repeatedly processing quantum measurements has severely limited the practical scalability of this approach. Here, we overcome this bottleneck by compressing quantum measurement data into machine-learning models that efficiently provide the wavefunction overlaps needed by QMC. This accelerates quantum-classical QMC by orders of magnitude, enabling chemically accurate simulations on up to 40 qubits, more than doubling the scale of previous hardware demonstrations with comparable accuracy. We demonstrate our approach on hydrogen chains and strongly correlated iron-sulfur clusters using measurements from Google's Willow processor. By enabling efficient reuse of quantum-state information on the classical coprocessor, our framework addresses a fundamental bottleneck in hybrid quantum algorithms that will persist even into the fault-tolerant era.
Introduction. Accurately simulating correlated electrons remains a central challenge in computational science, with applications ranging from molecular discovery to high-temperature superconductivity.[1, 2] Quantum computers offer a fundamentally different way to represent electronic correlation, motivating extensive efforts to develop quantum algorithms for chemistry and materials science.[3, 4] Among these approaches, quantum–classical hybrid quantum Monte Carlo (QC-QMC) has emerged as a promising strategy for achieving high accuracy.[5] In QC-QMC, a quantum processor prepares a correlated trial wavefunction whose overlaps with sampled many-body configurations (walkers) guide the classical Monte Carlo projection and help control the fermionic sign problem. The key promise of QC-QMC is that expressive quantum circuits can provide trial wavefunctions beyond the reach of classical approximations, while the subsequent Monte Carlo projection relaxes the accuracy requirements on quantum state preparation. Realizing this promise, however, requires an efficient interface between quantum overlap evaluations and classical QMC calculations.
To avoid the latency of repeated quantum–classical feedback, QC-QMC collects a reusable ensemble of randomized measurements of the trial state, from which a classical coprocessor estimates the overlaps required by QMC.[5, 6, 7, 8] Yet, repeatedly reconstructing these overlaps for the many walkers encountered throughout QMC propagation incurs enormous classical computational cost.[9, 10] This postprocessing bottleneck has severely limited hardware demonstrations, which have so far reached only 16 qubits[5, 11, 12] and consumed over 13,000 GPU hours for
a single calculation.[11] Meanwhile, a 40-qubit calculation was estimated to require millions of CPU hours.[12]
Here, we overcome this bottleneck by developing a machine learning (ML) framework (Figure 1) that compresses quantum measurement data into a classical surrogate.[13, 14, 15, 16, 17, 18] Rather than reconstructing the full quantum state, we learn its overlaps with the many-body configurations relevant to QMC. Once trained, the surrogate efficiently supplies these overlaps and their derivatives, bypassing repeated classical postprocessing of quantum measurements. This reduces the computational cost by orders of magnitude and makes end-to-end QC-QMC on 40 qubits of Google’s Willow processor affordable, more than doubling the scale of previous hardware demonstrations. At this scale, QC-QMC gives energies within chemical accuracy, for hydrogen chains with a mean-field trial and for an iron–sulfur cluster with a correlated trial.
Figure 1. Learning from quantum data to accelerate QC-QMC. A quantum processor prepares a trial state ∣ΨT⟩ and collects randomized measurements (“classical shadows”), each repeated for multiple shots. Conventional QC-QMC repeatedly processes these measurements to estimate trial-state overlaps with walkers throughout imaginary-time propagation. In our approach, we compress the measurement data into a machine-learning surrogate fθ, which efficiently provides the overlaps QMC requires without repeated shadow postprocessing. In the quantum-guided walk, each path is a QMC walker propagating in imaginary time from the initial state (black circle); the final walkers (dots, sized by weight) form an ensemble (shaded oval) that samples the ground state.
Theory and algorithms. In projector QMC, the exact ground state, ∣Ψ0⟩, is obtained by imaginary-time propagation, which is approximated by a stochastically sampled ensemble of wavefunctions,
The initial state ∣Φ0⟩ is assumed to have nonzero overlap with the ground state, and {∣ϕi⟩} are statistical sample wavefunctions, known as walkers, with associated weights {wi}. In unconstrained QMC, cancellations among walker weights cause the signal-to-noise ratio to decay exponentially with increasing imaginary time and system size, a manifestation of the notorious fermionic sign problem.[19] Practical QMC methods control this instability by guiding the calculation with constraints imposed by a trial wavefunction ∣ΨT⟩. In this work, we use auxiliary-field QMC,[20, 21] but our technique is applicable to other QMC methods.[22, 23] The trial wavefunction enters QMC through its overlaps with the walkers, ⟨ΨT∣ϕi⟩, and their derivatives, which determine the walker evolution and energy estimates (Supplementary Sec. S2 A).
To obtain these overlaps from a quantum-prepared trial state,[5] we perform randomized measurements via Clifford matchgate circuits, yielding classical snapshots that we store and reuse to estimate overlaps with the walkers.[8] Using an improved algorithm for processing matchgate shadows, we can process each snapshot per overlap with O(N3) cost, where N is the number of qubits (Supplementary Sec. S4). This operation must be repeated for Ns measurements, Nw walkers, and Nτ imaginary-time steps, resulting in a cost of O(N3NsNwNτ) for the overlaps alone, before accounting for the derivatives needed for the importance sampling and energy estimates.
Our central idea is to decouple quantum-measurement processing from repeated overlap evaluations during QMC. We first estimate complex-valued overlaps for a much smaller set of Ntrain≪NwNτ training determinants and use them to train a differentiable surrogate overlap model fθ(ϕ)∝⟨ΨT∣ϕ⟩. This introduces a one-time labeling cost of O(N3NsNtrain). Subsequent QMC propagation queries only the trained surrogate, whose evaluation cost is independent of Ns and whose derivatives can be evaluated directly without reprocessing the quantum measurements. Further details are provided in Supplementary Sec. S2.
In principle, the surrogate can be any differentiable model of the overlap trained on the measurement data. We train fθ by supervised learning on overlap labels representative of those used by QMC. Because the training labels are obtained from noisy quantum measurements, care must be taken to prevent the model from fitting overlap uncertainty due to the measurement noise and a finite number of measurements.[24, 25] We quantify the relative statistical error of the experimental labels by a noise floor F, estimated from fluctuations across subsets of the measurement circuits (Supplementary Sec. S2 B and Figure S2.2). Since the labels have an uncertainty level on the order of F itself, a learned model whose relative error on the validation labels (Eq. (25)) falls below F may be fitting uncertainty. We therefore use this as a stopping criterion for the training.
Figure S2.2. Convergence of the noise floor with the number of measurement circuits. Noise floor F (Eq. (24)) against the number of circuits Nc. Markers are measured values; lines are power-law fits F∝Nc−1/2; the dotted line shows Nc−1/2. a, Hydrogen chains H4–H20. b, LUCJ circuits of [Fe2S2(SCH3)4]2− at S=0 on hardware: spin-uncoupled and spin-coupled circuits for (18e, 14o), both on 30 qubits, and the spin-uncoupled circuit for (30e, 20o) on 42 qubits. c, LUCJ circuits at S=5, where only the β register is run (14 and 20 qubits): (18e, 14o) on hardware; (30e, 20o) from classically simulated shadows.
In general, the quantum trial states may be prepared through variational quantum algorithms,[26] high-fidelity quantum state preparation,[27] or quantum implementations of sophisticated classical wavefunctions. We demonstrate our ML-surrogate framework in two settings. First, we use mean-field trial wavefunctions for hydrogen chains containing up to 20 atoms (H20, 40 qubits) to benchmark our proposed approach. To enable ML inference for arbitrary QMC walker configurations, we extend neural quantum states (NQS)[13] to accept arbitrary Slater determinants as inputs. Training NQS with randomized quantum measurements has shown promise,[17, 18] but previous models were restricted to bit-string amplitudes, rather than the overlaps with rotated Slater determinants required by auxiliary-field QMC.
Second, for strongly correlated iron–sulfur complexes, preparing an accurate trial wavefunction directly on near-term quantum hardware remains prohibitively expensive. For our hardware demonstration, we instead assume access to an accurate wavefunction and approximate it with shallow quantum circuits, whose parameters we optimize classically to reproduce overlaps with representative training walkers. We then execute these circuits on quantum hardware, collect randomized measurements, and train our ML surrogate based on nonorthogonal configuration interaction (NOCI).[28, 29] This enables chemically accurate QC-QMC calculations for [Fe2S2(SCH3)4]2− in 28- and 40-qubit active spaces, despite hardware noise.
Results and discussion. The experiments in this work were performed on a Google Willow quantum processor,[30] which has over a hundred qubits, two-qubit gate error rates on the order of 10−3 (Supplementary Fig. S1.1), and reported repetition rates of approximately 6×104 shots per second. This enables us to collect millions of classical shadow measurements every few minutes. To suppress the effects of hardware noise, we employ randomized compiling,[31] dynamical decoupling,[32] parity postselection, and qubit-assignment averaging[33] (Supplementary Sec. S1 B). Hydrogen chains are a standard benchmark for strongly correlated electronic structure methods,[34, 35] as the degree of electronic correlation can be tuned by varying the interatomic separation. Increasing the chain length also provides a systematic way to assess how computational methods scale with system size. As a controlled benchmark of our ML framework, we use a spin-unrestricted Hartree–Fock (HF) trial state for hydrogen chains up to H20, performing the experiments and QMC calculations in the HF basis. We show the circuit in Figure 2a: a GHZ-type state is prepared, and each spin register is then measured through its own random Clifford matchgate circuit. Measuring the two registers separately suffices because each QMC walker is a product of an α and a β determinant[36] (Supplementary § 1.6 B).
Figure 2. Accurate and accelerated QC-QMC using experimental randomized measurements. a. Quantum circuit: a GHZ-type state is prepared on the α (spin-up) and β (spin-down) registers. Independent random matchgate circuits then act on each register before computational-basis measurements. The lower diagram illustrates the decomposition of a matchgate block into single-qubit Clifford and CZ gates. b. QC-QMC energy deviations from the noiseless, infinite-shot reference for hydrogen chains from H4 to H20 (8 to 40 qubits) at an interatomic separation of 3.0 Bohr in the STO-6G basis. Results from conventional shadow postprocessing (red) are compared with those from the ML surrogate (blue). The shaded region indicates chemical accuracy (±1 kcal mol−1). ML results are averaged over independently trained models (Figure S3.6), with error bars indicating standard errors of the mean, including reference uncertainty. c. Total classical computational cost of conventional QC-QMC and our ML-accelerated approach. At 8 and 16 qubits, both methods use the same quantum measurement data, yielding speedups of 190- and 4,100-fold, respectively. The ML costs include training-label generation, surrogate training, and QMC propagation. The in-set shows the total number of measurement snapshots (circuits × shots) used at each system size (Supplementary Table S1.1).
state is prepared, and each spin register is then measured through its own random Clifford matchgate circuit. Measuring the two registers separately suffices because each QMC walker is a product of an α and a β determinant[36] (Supplementary § 1.6 B). Together with an improved compilation (Supplementary § 1.6 A), this keeps the circuit sizes compact (Supplementary Fig. S1.2). At every size, our matchgate circuits need under a third of the CZ gates and under half the depth of the full matchgate circuits used in previous experiments.[11] Shallower circuits accumulate less hardware noise per snapshot, which is crucial for extracting useful overlap information (Supplementary Fig. S3.4).
We train an expressive NQS surrogate directly on noisy measurement data, starting from a random initialization with no classical initial guess. Figure 2b shows the QC-QMC energies obtained with the trained surrogate as the QMC trial, relative to the ideal limit (infinite number of shots and noiseless devices) of QMC with HF trial. At every size, the deviation averaged over networks trained with different random seeds (Figure S3.6) lies within chemical accuracy (1 kcal mol−1). Figure 2c compares the cost of our approach with that of conventional QC-QMC, which evaluates every overlap, force bias, and local energy directly from the snapshots, using our optimized GPU implementation (Supplementary Fig. S2.1). At 8 and 16 qubits, conventional QC-QMC needs 100 and 2,500 circuits, respectively, at 500 shots each, to reach chemical accuracy. With the same data, our ML model reduces QC-QMC cost by 190-fold (8 qubits) and 4,100-fold (16 qubits). This comparison includes the cost of labeling the training data, training the NQS model, and running QMC end-to-end (Table S3.1). At 24 qubits and beyond, the estimated cost of processing all measurement data using conventional QC-QMC exceeds 105 GPU hours, making direct benchmarking prohibitively expensive. At these sizes, the one-time cost of producing the training labels dominates the total cost of our ML approach (>90%), which is still orders of magnitude smaller than a conventional QC-QMC run.
System
Qubits
Circuits
Shots
Total
walkers
Labeling
Training
Run
Total
(GPU hours)
Speedup
H₄
8
100
500
50,000
11,520
0.004
0.40
0.76
1.17
218
187×
H₈
16
2,500
500
1,250,000
11,520
0.33
0.35
1.13
1.80
7,367
4,082×
H₁₂
24
10,000
1,000
10,000,000
43,520
37.9
0.45
1.76
40.1
–
–
H₁₆
32
20,000
1,380
27,600,000
43,520
189.9
0.46
2.55
192.9
–
–
H₂₀
40
50,000
1285
64,250,000
81,920
1,188.2
0.40
4.21
1,192.8
–
–
Table S3.1. TABLE S3.1. Cost of conventional QC-QMC (optimized algorithm) and of QC-QMC with the neural surrogate for hydrogen chains. GPU hours are hours on one graphics compute die (GCD) of an AMD Instinct MI250X GPU on the Frontier supercomputer at the Oak Ridge Leadership Computing Facility (OLCF). An AFQMC run has 640 walkers and 5,000 blocks, each block being 25 propagation steps (overlap and force bias) and one local-energy evaluation. Shadows: number of circuits, shots per circuit, and their product. The surrogate is trained on N walker pairs plus 2,560 stopping-set pairs (N=3,200, 3,200, 19,200, 19,200 and 38,400 from 8 to 40 qubits), so 2(N+2560) walkers are labeled, each with its overlap estimated from all the shadows. The cost of this work covers the full protocol: Training is the sum of ten trainings (one per network seed), Run is ten AFQMC runs, one with each trained network as the trial, and Total is their sum with labeling. Speedup is the conventional cost over the total cost of this work (10 independent trained models and AFQMC).
Figure S3.6. Training-seed dependence of the hydrogen-chain energies. Deviation of QC-QMC with the NQS surrogate for each network that passes the health checks, with the statistical error of its AFQMC run.
We next turn to a widely studied strongly correlated system: the iron–sulfur cluster. This class of transition-metal complexes facilitates electron transfer and catalysis across a wide range of biological processes.[37] Their electronic structure is challenging owing to competing spin couplings among iron d electrons, which produce many nearly degenerate spin states,[38, 39, 40] and charge fluctuations between iron and sulfur atoms. Achieving high accuracy at low computational cost for this compound and related systems has been challenging for both classical QMC[41] and near-term quantum methods such as sample-based quantum diagonalization (SQD).[42, 43, 44] This provides a more demanding test of our framework than the hydrogen chain benchmark.
For this demonstration, we use a shallow, hardware-efficient ansatz based on the local unitary cluster Jastrow (LUCJ),[5, 45] as in the SQD study[42] but with a different parameterization (Supplementary Sec. S1 C). We optimize the LUCJ circuit parameters to reproduce overlaps between the reference wavefunction and training walkers. Interestingly, direct variational energy minimization of the LUCJ wavefunction does not yield accurate QC-QMC results (Table S1.3). Here, we use a compact NOCI expansion[28, 29] as the surrogate, initialized from an inexpensive classical calculation and subsequently trained on quantum measurements. This provides a direct comparison with a classically optimized NOCI trial containing the same number of determinants, allowing us to assess the benefit of learning from quantum data. We study two active spaces, (18-electron, 14-orbital) and (30-electron, 20-orbital), hereafter denoted as (18e, 14o) and (30e, 20o), respectively, which map onto 28 and 40 qubits.
Active space
Circuit
Variational
AFQMC
(18e, 14o)
Energy-optimized
+16.73
−11.23(11)
Taught
+30.66
−0.80(11)
(30e, 20o)
Energy-optimized
+26.89
−10.88(10)
Taught
+52.84
−0.43(11)
Table S1.3. Energy-optimized versus taught spin-uncoupled LUCJ circuits. Error of the variational energy of each circuit and of phaseless AFQMC with the exact (noiseless) circuit state as the trial, in mEh, each relative to FCI for its own active-space Hamiltonian. Teaching starts from the energy-optimized circuit; the taught circuits are the ones run on Willow. Teaching raises the variational energy but brings the AFQMC energy close to FCI.
In Figure 3a, we show the quantum circuit for (18e, 14o): a GHZ-type preparation followed by a shallow LUCJ ansatz, measured with spin-blocked randomized measurements[36] (matchgate shadows). Each spin register runs along a two-
Figure 3. Chemically accurate QC-QMC for [Fe2S2(SCH3)4]2− in a 28-qubit active space. The active space contains 18 electrons in 14 orbitals (18e, 14o). a. Quantum circuit consisting of GHZ-type state preparation, a shallow local unitary cluster Jastrow (LUCJ) ansatz, and randomized matchgate measurements. We consider spin-coupled and spin-uncoupled variants. The former includes α–β Jastrow interactions, whereas the latter operates on two independent 15-qubit spin registers, each containing one ancilla to ensure even fermion parity. Full circuits in Supplementary Figs. S1.5 and S1.6. b. Qubit layout on the Willow processor. Each spin register (α, blue; β, purple) runs along a two-row snake of dark orange couplers, which both circuits use; black crosses mark the ancilla qubit of each register. The light orange couplers between the registers are used only by the spin-coupled circuit, for its α–β Jastrow pairs. Grey qubits and couplers are unused. Error rates in Supplementary Fig. S1.1. c. Singlet ground-state energy errors relative to the exact result, comparing classical trials (red) with quantum-circuit trials (blue and yellow). The upper panel shows variational trial-state errors, while the lower panel shows the corresponding QMC errors. Experimental QC-QMC results (blue) use measurements from Willow; ideal results (yellow) use noiseless circuits in the infinite-shot limit. The shaded region indicates chemical accuracy (±1 kcal mol−1). d. Energies of the singlet (S=0) and high-spin (S=5) states relative to the exact singlet energy, comparing unrestricted coupled-cluster methods, experimental and ideal QC-QMC, and exact results. Vertical lines indicate the corresponding spin gaps.
row snake on Willow (Figure 3b). The spin-coupled circuit places its α–β Jastrow pairs on the couplers (light orange) between the two snakes, whereas the spin-uncoupled circuit omits them and runs as two independent 15-qubit circuits including ancillas.
Figure 3c compares classical QMC and QC-QMC for the singlet ground state. To illustrate the difficulty of this problem, we show the coupled-cluster (CC) hierarchy up to singles (S), doubles (D), triples (T), and quadruples (Q), where the trial energy improves as one adds more excitations, but the QMC error does not quickly converge to chemical accuracy.[41] Even CCSDTQ, whose own energy is within chemical accuracy, leaves an AFQMC error of 2.8 kcal mol−1 when its linearized wavefunction is used as the trial. In contrast, our quantum trial achieves chemical accuracy on Willow both with and without the α–β Jastrow couplings, with errors of −0.86(6) and −0.85(11) kcal mol−1, respectively (Supplementary Figure S3.8). This experiment collected 1.3×108 and 2.5×107 postselected snapshots for the spin-coupled and -uncoupled circuits, respectively (Supplementary Table S1.1). By contrast, a classical NOCI trial with the same number of determinants (four) as our spin-coupled surrogate, optimized for its variational energy, has a lower variational energy than our quantum trials yet yields an AFQMC energy about 5 kcal mol−1 below the exact result, and classical NOCI trials with 2–8 determinants all miss chemical accuracy (Supplementary Fig. S3.7). This highlights the role of quantum data in generating an accurate QMC trial wavefunction.
Figure S3.8. QC-QMC for [Fe₂S₂(SCH₃)₄]²⁻ with surrogates trained on the LUCJ hardware data (a), Absolute AFQMC error relative to FCI, on a logarithmic axis, for the three LUCJ circuits. Classical AFQMC results are from classical VAP-NOCI trials. QC-QMC: AFQMC with the surrogate trained from that experimental data. Red dashed line: chemical accuracy (1.6 mEh). (b), QC-QMC error of the shadow surrogates against their validation error. Each point is a QC-QMC surrogate, trained with different seeds or starting states. Horizontal axis: validation error v (Eq. (25)) over the noise floor F of the labels, with a separate range in each row; the dashed black line is v=F, below which the labels cannot show a surrogate to be closer to the circuit state.
In Figure 3d, we calculate the energy gap between the singlet and the S=5 high-spin state. CC methods systematically improve and approach the exact value at CCSDTQ. QC-QMC, with both spin states guided by the measurement-trained trial surrogate, gives 20.2(1) kcal mol−1 against the exact 19.3 kcal mol−1, within chemical accuracy and on par with CCSDTQ. Our experimental QC-QMC closely reproduces the ideal (noiseless, infinite-shot) QC-QMC results, highlighting the noise robustness of QC-QMC combined with the ML surrogate. Notably, the spin-uncoupled trial is as accurate as the spin-coupled one, so the α–β Jastrow pairs are not needed here. Without explicit opposite-spin Jastrow couplings in the trial, the circuit has no gates between the spin registers and splits into two independent halves, enabling experiments with reduced circuit complexity. For S=5, the α-spin sector is fully occupied, so a nontrivial α-spin circuit is unnecessary (Supplementary Sec. S3 F).
Finally, we use the spin-uncoupled trial wavefunction for the (30e, 20o) active space, corresponding to a 40-qubit total system run as two 21-qubit circuits (Figure 4). Using 3.0×108 postselected snapshots and a NOCI surrogate, QC-QMC achieves chemical accuracy for the singlet ground state (0.91(8) kcal mol−1) relative to the exact energy. Unlike the singlet, where nearest-neighbor Jastrow pairs suffice, the S=5 state requires pairs beyond nearest neighbors. Routing these pairs along the qubit line makes the circuit too deep to run at sufficiently high fidelity on the current device, so we classically emulate the quantum data using fast quantum circuit simulators.[46] We train an ML model on these emulated data as we do for the experimental data (Supplementary Table S1.1 and Figure S2.2). Combining the experimental singlet results with the simulated high-spin data yields the spin gap within chemical accuracy.
Figure 4. QC-QMC spin gap of [Fe2S2(SCH3)4]2− in a 40-qubit active space. a, Qubit layout for the singlet (S=0) calculation on Willow. The 40 spin orbitals map to 40 qubits, with one additional ancilla per spin register to ensure even fermion parity. We implement the spin-uncoupled ansatz as two independent 21-qubit circuits operating in parallel on 42 physical qubits. Error rates in Supplementary Fig. S1.1. b, Energies of the singlet (S=0) and high-spin (S=5) states relative to the exact singlet energy. We compare unrestricted coupled-cluster results with QC-QMC and exact results. Vertical lines indicate the spin gaps. The QC-QMC (Exp.) result combines experimental classical shadows for the singlet state with classically simulated shadows for the high-spin state (dashed line). The QC-QMC (ideal) result uses noiseless circuits with the infinite-shot limit for both states.
Conclusions and outlook. Our work overcomes a major computational bottleneck in hybrid quantum–classical computing by compressing quantum measurement data into reusable machine-learning models. This accelerates QC-QMC by orders of magnitude, enabling chemically accurate simulations on up to 40 qubits, more than doubling the scale of previous hardware demonstrations. Because the classical postprocessing bottleneck persists even on fault-tolerant quantum computers, our approach addresses a fundamental challenge for both near-term and future quantum computing.
A remaining challenge is the statistical measurement cost of resolving overlaps that decay exponentially with system size.[47, 48] Although this limitation has not prevented our present calculations, even for strongly correlated systems such as iron–sulfur clusters, understanding its impact at larger scales will be crucial for identifying problems where QC-QMC can achieve practical quantum advantage. More broadly, our results establish a pathway toward scalable hybrid algorithms in which quantum information can be efficiently exploited by classical many-body methods.
Acknowledgments. This work was primarily supported by the U.S. Department of Energy, Office of Fusion Energy Sciences, “Foundations for quantum simulation of warm dense matter” project (TJ, AZ, ADB, JL). TJ and JL were partially supported by the startup funds from Harvard University and Wellcome Leap through the Quantum for Bio (Q4Bio) program. TJ also acknowledges support from the Gordon and Betty Moore Foundation Fellowship. We thank Pedram Roushan for valuable feedback on the presentation of this work, Jinghong Zhang and Nishad Maskara for their engagement at the beginning of the project, and Ryan Byrne, Mohammed Alghadeer, and Ricky Oliver for their support in running the quantum experiments.
This research used computational resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory through the INCITE Project No. MAT281, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DEAC05-00OR22725, and the FAS Research Computing cluster supported by the Faculty of Arts and Sciences (FAS) Division of Science Research Computing Group at Harvard, and National Nuclear Security Administration’s Advanced Simulation and Computing Program. JN, TJ and JL acknowledge the support of NSF ACCESS through allocation CHE250121 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.
We collected the hardware data presented in this work remotely on a 105-qubit Willow processor, with access provided via Google’s Quantum Engine through the Willow Early Access Program. Calibration and support were provided by the Quantum Hardware Residency Program. We thank the Google Quantum AI team for providing the quantum systems and support that enabled these results. The views expressed in this work are solely those of the authors and do not reflect the policy of Google or the Google Quantum AI team.
This article has been authored by an employee of National Technology & Engineering Solutions of Sandia, LLC under Contract No. DE-NA0003525 with the U.S. Department of Energy (DOE). The employee owns all right, title, and interest in and to the article and is solely responsible for its contents. The United States Government retains, and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this article or allow others to do so, for United States Government purposes. The DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan https://www.energy.gov/downloads/doe-public-access-plan.
AI statement. Claude (Fable 5.1 and Opus 5.5) and Codex (Astra 6.0 and Sol 6.1) assisted the authors with algorithmic implementations, programming, critical review, and manuscript preparation. The authors supervised all use and validated the outputs.
Code and data availability. The data supporting the findings of this study are provided in the main text and Supplementary Information. The code and raw measurement records are available from the corresponding authors upon request and will be made publicly available in the future.
Author contributions. TJ and JL conceived the idea and designed the study. TJ performed classical numerics and QC-QMC calculations with contributions from AZ, JN, and ADB. AZ developed improved algorithms for classical shadows and circuit compilation, and performed the quantum experiments with contributions from TJ, SK, and ER. TJ and AZ developed and optimized the GPU implementation of the QC-QMC postprocessing. TJ, AZ, and JL drafted the manuscript. JL supervised the overall project. All authors participated in discussions, manuscript writing, and data analysis.
References
Supplementary Material
Quantum-state preparation and measurements
Hardware implementation
Qubit layouts and error rates. Figure S1.1 shows the qubit layouts used on the Willow processor for the 40-qubit hydrogen chain and for [Fe2S2(SCH3)4]2− in (18e, 14o) and (30e, 20o), together with their single-qubit, two-qubit (CZ) and readout error rates. Single-qubit and readout errors are similar for the three layouts. CZ errors are lower for the hydrogen chain than for the two [Fe2S2(SCH3)4]2− layouts, since a one-dimensional path can avoid poor couplers more easily than a compact two-dimensional grid.
References
[1]A. Nandy, C. Duan, M. G. Taylor, F. Liu, A. H. Steeves, and H. J. Kulik, Computational discovery of transition-metal complexes: From high-throughput screening to machine learning, Chem. Rev. 121, 9927 (2021).DOI
[2]L. P. Gor’kov and V. Z. Kresin, Colloquium: High pressure and road to room temperature superconductivity, Rev. Mod. Phys. 90, 011001 (2018).DOI
[3]B. Bauer, S. Bravyi, M. Motta, and G. K.-L. Chan, Quantum Algorithms for Quantum Materials Science, Chemical Reviews 120, 12685 (2020).DOI
[4]R. W. Chien, M. Chiew, B. Harrison, J. Necaise, W. Wang, M. M. Madsen, C. McLauchlan, T. M. Henderson, G. E. Scuseria, S. Strelchuk, et al., Simulating fermions with a digital quantum computer, Nature Reviews Physics 8, 131 (2026).DOI
[5]W. J. Huggins, B. A. O’Gorman, N. C. Rubin, D. R. Reichman, R. Babbush, and J. Lee, Unbiasing fermionic quantum Monte Carlo with a quantum computer, Nature 603, 416 (2022).
[6]H.-Y. Huang, R. Kueng, and J. Preskill, Predicting many properties of a quantum system from very few measurements, Nature Physics 16, 1050 (2020).DOI
[7]A. Zhao, N. C. Rubin, and A. Miyake, Fermionic partial tomography via classical shadows, Physical Review Letters 127, 110504 (2021).DOI
[8]K. Wan, W. J. Huggins, J. Lee, and R. Babbush, Match-gate shadows for fermionic quantum simulation, Communications in Mathematical Physics 404, 629 (2023).
[9]M. Kiser, A. Schroeder, G.-L. R. Anselmetti, C. Kumar, N. Moll, M. Streif, and D. Vodola, Classical and quantum cost of measurement strategies for quantum-enhanced auxiliary field quantum Monte Carlo, New Journal of Physics 26, 033022 (2024), arXiv:2312.09872.arxiv.org/abs/2312.09872
[10]B. Huang, Y.-T. Chen, B. Gupt, M. Suchara, A. Tran, S. McArdle, and G. Galli, Evaluating a quantum-classical quantum Monte Carlo algorithm with Matchgate shadows, Physical Review Research 6, 043063 (2024).DOI
[11]L. Zhao, J. J. Goings, W. Aubramud, A. Arrasmith, L. Calderin, S. Churchill, D. Gabay, T. Harvey-Brown, M. Hiles, M. Kaja, M. Keesan, K. Kulesz, A. Maksymov, M. Maruo, M. M. noz, B. Nijholt, R. Schiller, Y. de Sereville, A. Smidutz, F. Tripier, G. Yao, T. Zaveri, C. Collins, M. Roetteler, E. Epifanovsky, A. Kovyrshin, L. Tornberg, A. Broo, J. R. Hammond, Z. Chandani, P. Kha- late, E. Kyoseva, Y.-T. Chen, E. M. Kessler, C. Y.-Y. Lin, G. Ramu, R. Shaffer, M. Brett, B. Huang, M. R. Hugues, and T. Y. Takeshita, Quantum-classical auxiliary field quantum Monte Carlo with matchgate shadows on trapped ion quantum computers, Physical Review Research 8, 033061 (2026).DOI
[12]F. Nappi, M. Kiser, and F. Simkovic, IV, Quantum-Classical Auxiliary-Field Quantum Monte Carlo at the Edge of Practicability (2026), version 1, arXiv:2606.19239.arxiv.org/abs/2606.19239
[13]G. Carleo and M. Troyer, Solving the quantum many-body problem with artificial neural networks, Science 355, 602 (2017).DOI
[14]K. Choo, A. Mezzacapo, and G. Carleo, Fermionic neural-network states for ab-initio electronic structure, Nature Communications 11, 2368 (2020).DOI
[15]J. Hermann, Z. Schätzle, and F. Noé, Deep-neural-network solution of the electronic Schrödinger equation, Nature Chemistry 12, 891 (2020).
[16]D. Pfau, J. S. Spencer, A. G. D. G. Matthews, and W. M. C. Foulkes, Ab initio solution of the many-electron Schrödinger equation with deep neural networks, Physical Review Research 2, 033429 (2020).
[17]V. Wei, W. A. Coish, P. Ronagh, and C. A. Muschik, Neural-shadow quantum state tomography, Physical Review Research 6, 023250 (2024).DOI
[18]H.-Y. Huang, J. Preskill, and M. Soleimanifar, Certifying almost all quantum states with few single-qubit measurements, Nature Physics 21, 1834 (2025).
[19]S. Zhang, J. Carlson, and J. E. Gubernatis, Constrained path monte carlo method for fermion ground states, Phys. Rev. B 55, 7464 (1997).DOI
[20]S. Zhang and H. Krakauer, Quantum Monte Carlo Method using Phase-Free Random Walks with Slater Determinants, Physical Review Letters 90, 136401 (2003).pubmed.ncbi.nlm.nih.gov/12689312
[21]M. Motta and S. Zhang, Ab initio computations of molecular systems by the auxiliary-field quantum Monte Carlo method, Wiley Interdisciplinary Reviews: Computational Molecular Science 8, e1364 (2018).DOI
[22]T. Jiang, J. Zhang, M. K. Baumgarten, M.-F. Chen, H. Q. Dinh, A. Ganeshram, N. Maskara, A. Ni, and J. Lee, Walking through hilbert space with quantum computers, Chemical Reviews 125, 4569 (2025).DOI
[23]N. S. Blunt, L. Caune, and J. Quiroz-Fernandez, Quantum computing approach to fixed-node monte carlo using classical shadows, Journal of Chemical Theory and Computation 21, 1652 (2025).DOI
[24]G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, and G. Carleo, Neural-network quantum state tomography, Nature Physics 14, 447 (2018).DOI
[25]J. Carrasquilla, G. Torlai, R. G. Melko, and L. Aolita, Reconstructing quantum states with generative models, Nature Machine Intelligence 1, 155 (2019).DOI
[26]M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, Variational quantum algorithms, Nature Reviews Physics 3, 625 (2021).DOI
[27]D. W. Berry, Y. Tong, T. Khattar, A. White, T. I. Kim, G. H. Low, S. Boixo, Z. Ding, L. Lin, S. Lee, G. K.-L. Chan, R. Babbush, and N. C. Rubin, Rapid initial-state preparation for the quantum simulation of strongly correlated molecules, PRX Quantum 6, 020327 (2025).DOI
[28]C. A. Jiménez-Hoyos, R. Rodríguez-Guzmán, and G. E. Scuseria, Multi-component symmetry-projected approach for molecular ground state correlations, The Journal of chemical physics 139, 204106 (2013).
[29]C. Giuliani, J. Nys, R. Martinazzo, G. Carleo, and R. Rossi, Precise quantum chemistry calculations with few slater determinants, Nature Communications 17, 5613 (2026).DOI
[30]Google Quantum AI and Collaborators, Quantum error correction below the surface code threshold, Nature 638, 920 (2025).
[31]J. J. Wallman and J. Emerson, Noise tailoring for scalable quantum computation via randomized compiling, Physical Review A 94, 052325 (2016).DOI
[32]L. Viola, E. Knill, and S. Lloyd, Dynamical decoupling of open quantum systems, Physical Review Letters 82, 2417 (1999).DOI
[33]Google AI Quantum and collaborators, Observation of separated dynamics of charge and spin in the Fermi-Hubbard model, arXiv:2010.07965 (2020).
[34]M. Motta, D. M. Ceperley, G. K.-L. Chan, J. A. Gomez, E. Gull, S. Guo, C. A. Jiménez-Hoyos, T. N. Lan, J. Li, F. Ma, A. J. Millis, N. V. Prokof’ev, U. Ray, G. E. Scuseria, S. Sorella, E. M. Stoudenmire, Q. Sun, I. S. Tupitsyn, S. R. White, D. Zgid, and S. Zhang (Simons Collaboration on the Many-Electron Problem), Towards the solution of the many-electron problem in real materials: Equation of state of the hydrogen chain with state-of-the-art many-body methods, Physical Review X 7, 031059 (2017).
[35]M. Motta, C. Genovese, F. Ma, Z.-H. Cui, R. Sawaya, G. K.-L. Chan, N. Chepiga, P. Helms, C. Jiménez-Hoyos, A. J. Millis, U. Ray, E. Ronca, H. Shi, S. Sorella, E. M. Stoudenmire, S. R. White, and S. Zhang (Simons Collaboration on the Many-Electron Problem), Ground-state properties of the hydrogen chain: Dimerization, insulator-metal transition, and magnetic phases, Physical Review X 10, 031058 (2020).DOI
[36]A. Zhao and A. Miyake, Group-theoretic error mitigation enabled by classical shadows and symmetries, npj Quantum Information 10, 57 (2024).DOI
[37]H. Beinert, R. H. Holm, and E. Münck, Iron-Sulfur Clusters: Nature’s Modular, Multipurpose Structures, Science 277, 653 (1997).pubmed.ncbi.nlm.nih.gov/9235882
[38]S. Sharma, K. Sivalingam, F. Neese, and G. K.-L. Chan, Low-energy spectrum of iron–sulfur clusters directly from many-particle quantum mechanics, Nature Chemistry 6, 927 (2014).DOI
[39]Z. Li, S. Guo, Q. Sun, and G. K.-L. Chan, Electronic landscape of the P-cluster of nitrogenase as revealed through many-electron quantum wavefunction simulations, Nature Chemistry 11, 1026 (2019).DOI
[40]H. Zhai, C. Li, X. Zhang, Z. Li, S. Lee, and G. K.-L. Chan, Classical quantum chemical solution of the femo-cofactor model to chemical accuracy and its implications, Science Advances 12, eaeug7106 (2026).DOI
[41]E. F. Kjønstad, H. Zhai, J. Shee, S. Sharma, and G. K.-L. Chan, Can phaseless auxiliary-field quantum Monte Carlo with broken symmetry trials describe iron-sulfur clusters? (2026), arXiv:2605.03270.
[42]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).DOI
[43]P. Reinholdt, K. M. Ziems, E. R. Kjellgren, S. Coriani, S. P. Sauer, and J. Kongsted, Critical limitations in quantum-selected configuration interaction methods, Journal of Chemical Theory and Computation 21, 6811 (2025).
[44]D. Danilov, J. Robledo-Moreno, K. J. Sung, M. Motta, and J. Shee, Enhancing the accuracy and efficiency of sample-based quantum diagonalization with phaseless auxiliary-field quantum Monte Carlo, Journal of Chemical Theory and Computation 21, 11585 (2025).
[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
[46]S. V. Isakov, D. Kafri, O. Martin, C. V. Heidweiller, W. Mruczkiewicz, M. P. Harrigan, N. C. Rubin, R. Thompson, M. Broughton, K. Kissell, et al., Simulations of quantum circuits with approximate noise using qsim and cirq, arXiv:2111.02396 (2021).arxiv.org/abs/2111.02396
[47]G. Mazzola and G. Carleo, Exponential challenges in unbiased quantum Monte Carlo algorithms with quantum computers (2022), version 1, arXiv:2205.09203.arxiv.org/abs/2205.09203
[48]J. Lee, D. R. Reichman, R. Z. Babbush, N. C. Rubin, F. D. Malone, B. O’Gorman, and W. J. Huggins, Response to “Exponential challenges in unbiased quantum Monte Carlo algorithms with quantum computers”, arXiv:2207.13776 (2022).arxiv.org/abs/2207.13776
[49]S. Chen, W. Yu, P. Zeng, and S. T. Flammia, Robust shadow estimation, PRX Quantum 2, 030348 (2021).
[50]R. Brieger, M. Heinrich, I. Roth, and M. Kliesch, Stability of classical shadows under gate-dependent noise, Physical Review Letters 134, 090801 (2025).DOI
[51]H. Belagali, T. Van 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].arxiv.org/abs/2607.21337
[52]T. Van Voorhis and M. Head-Gordon, A geometric approach to direct minimization, Molecular Physics 100, 1713 (2002).
[53]T. Jiang, B. O’Gorman, A. Mahajan, and J. Lee, Unbiasing fermionic auxiliary-field quantum Monte Carlo with matrix product state trial wavefunctions, Physical Review Research 7, 013038 (2025).DOI
[54]C. Paige and C. Van Loan, A Schur decomposition for Hamiltonian matrices, Linear Algebra and its Applications 41, 11 (1981).DOI
[55]C. B. Moler and G. W. Stewart, An algorithm for generalized matrix eigenvalue problems, SIAM Journal on Numerical Analysis 10, 241 (1973).DOI
[56]D. J. Thouless, Stability conditions and nuclear rotations in the hartree-fock theory, Nuclear Physics 21, 225 (1960).DOI
[57]T. Jiang, M. K. A. Baumgarten, P.-F. Loos, A. Mahajan, A. Scemama, S. F. Ung, J. Zhang, F. D. Malone, and J. Lee, Improved modularity and new features in ipie: Toward even larger AFQMC calculations on CPUs and GPUs at zero and finite temperatures, J. Chem. Phys. 161, 162502 (2024).DOI
[58]Z. Li and G. K.-L. Chan, Spin-projected matrix product states: Versatile tool for strongly correlated systems, Journal of chemical theory and computation 13, 2681 (2017).DOI
[59]B. O’Gorman, Fermionic tomography and learning, arXiv:2207.14787 (2022).arxiv.org/abs/2207.14787
[60]V. Heyraud, H. Chomet, and J. Tilly, Unified framework for matchgate classical shadows, npj Quantum Information 11, 65 (2025).DOI
[61]M. Wimmer, Algorithm 923: Efficient numerical computation of the Pfaffian for dense and banded skew-symmetric matrices, ACM Transactions on Mathematical Software 38, 1 (2012).
[62]J. Rubow and U. Wolff, A factorization algorithm to compute Pfaffians, Computer Physics Communications 182, 2530 (2011).DOI
[63]C. Van Loan, A symplectic method for approximating all the eigenvalues of a Hamiltonian matrix, Linear Algebra and its Applications 61, 233 (1984).DOI
[64]S. Okada, Pfaffian formulas and Schur Q-function identities, Advances in Mathematics 353, 446 (2019).DOI
[65]J. H. Wilkinson, The Algebraic Eigenvalue Problem, Vol. 662 (Oxford, Clarendon, 1965).
[66]G. H. Golub and C. F. Van Loan, Matrix Computations (JHU Press, 2013).
[67]C. Baer, The Faddeev-Leverrier algorithm and the Pfaffian, Linear Algebra and its Applications 630, 39 (2021).DOI
[68]R. A. Horn and C. R. Johnson, Matrix Analysis (Cambridge University Press, 2012).
[69]J. von zur Gathen and J. Gerhard, Modern Computer Algebra, 3rd ed. (Cambridge University Press, 2013).
[70]R. C. Thompson, Pencils of complex and real symmetric and skew matrices, Linear Algebra and its Applications 147, 323 (1991).DOI
[71]A. Dmytryshyn, Structure preserving stratification of skew-symmetric matrix polynomials, Linear Algebra and its Applications 532, 266 (2017).DOI
[72]A. N. Habermann, Parallel neighbor-sort (or the glory of the induction principle), Carnegie Mellon University Technical Report No. AD-759-248 (1972).
[73]NIST, NIST Digital Library of Mathematical Functions, §26.13: Permutations: Cycle notation, https://dlmf.nist.gov/26.13.
[74]P. Braccia, N. Diaz, M. Larocca, M. Cerezo, and D. García-Martín, Optimal Haar random fermionic linear optics circuits, arXiv:2505.24212 (2025).arxiv.org/abs/2505.24212