- Jupyter Notebook 92.2%
- Python 7.8%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| assets | ||
| Fig2 | ||
| Fig3 | ||
| Fig4 | ||
| lib | ||
| module | ||
| LICENSE | ||
| README.md | ||
Demonstration of Efficient Predictive Surrogates for Large-scale Quantum Processors
This repository contains the code and data for reproducing the experimental results in "Demonstration of Efficient Predictive Surrogates for Large-scale Quantum Processors". We introduce and validate two provably efficient predictive surrogates that emulate the mean-value behavior of noisy quantum processors with drastically reduced measurement overhead:
- VQE pre-training for transverse-field Ising models, achieving ground-state energy estimates with only 0.023% of conventional measurement cost while surpassing the accuracy of direct quantum approaches
- Identification of Floquet symmetry-protected topological phases using surrogate-emulated dynamics, reducing measurement requirements by over 100× compared to direct quantum processor access
These results demonstrate that predictive surrogates enable accurate, resource-efficient emulation of large-scale quantum processors for practical quantum simulation tasks.
Overview
Advanced quantum processors are valuable but scarce resources. To extend their utility, we develop predictive surrogates—classical models that emulate quantum processor behavior with provable efficiency guarantees. We introduce two complementary methods tailored to different quantum simulation scenarios:
| Method | Input Parameter Distributions | Measurement Operators | Application Scenarios | Theoretical Efficiency Conditions |
|---|---|---|---|---|
h_cs |
$\mathbf{x} \sim \text{Unif}[-\pi,\pi]^d$ Parameters are independent |
Multiple observables (Pauli-based classical shadows) |
• VQA pre-training • Device benchmarking |
• Small gradient norm • Large noise level • Local bounded observables |
h_qs |
$\mathbf{x} \in [-R,R]^d$ Arbitrary distribution Parameters may be correlated |
Single observable (Arbitrary measurement operator) |
• Hamiltonian simulation • Non-equilibrium dynamics |
• Small range $R$ • Large noise level • Few rotation gates |
Key experimental findings:
✅ >99.97% reduction in measurement overhead compared to conventional VQEs
✅ Scalable performance demonstrated across system sizes from 6 to 42 qubits
✅ Enables efficient classical inference for quantum processor emulation
Repository Structure
This repository contains the core code and data required to reproduce the experimental results from the paper:
Demonstration_of_Efficient_Predictive_Surrogates_for_Large-scale_Quantum_Processors/
├── Fig2/ # Pre-training VQEs for transverse-field Ising models
│ ├── Plot_fig2.ipynb # Main visualization script for Figure 2
│ ├── pretrain_by_hcs_ablation.py # Ablation studies for h_cs pre-training
│ └── expdata/ # Data files for Figure 2
├── Fig3/ # Relative error comparison across qubit counts
│ ├── Plot_fig3.ipynb # Main visualization script for Figure 3
│ ├── pretrain_by_hcs.py # Pre-training experiments using h_cs
│ ├── pretrain_by_hqs.py # Pre-training experiments using h_qs
│ └── expdata/ # Data files for Figure 3
├── Fig4/ # Floquet phase identification with DTC dynamics
│ ├── Plot_fig4.ipynb # Main visualization script for Figure 4
│ ├── predict_Rabi_oscillations.py # Rabi oscillations prediction for Figure 4a,b
│ ├── predict_phase_transition.py # Phase transition prediction for Figure 4c
│ └── expdata/ # Data files for Figure 4
├── lib/ # Core modules for predictive surrogates
├── module/ # Additional utility modules
├── LICENSE
└── README.md
Each figure folder contains:
- Code scripts to generate the corresponding figures in the paper
- Experimental data needed to run the scripts
- Jupyter notebooks for interactive visualization
The lib/ and module/ directories contain the core implementations of the predictive surrogates (h_cs and h_qs) and supporting utilities for quantum circuit simulation and data processing.
Requirements
python>=3.10.4
numpy>=1.21.0
scipy>=1.7.0
jax>=0.4.30
optax>=0.2.4
pennylane>=0.41.1
qiskit>=0.46.3
h5py>=3.7.0
matplotlib>=3.6.0
seaborn>=0.12.0
Usage
1. Reproducing Fig 2: Using Predictive Surrogate h_cs to Pre-train VQE for 1D TFIMs
To generate the visualizations for Figure 2, run the Jupyter notebook Plot_fig2.ipynb in the Fig2 directory.
What this figure shows:
- 2a: Quantifies the agreement between the quantum processor outputs and the
h_cssurrogate predictions using MAE, R², and Pearson correlation. - 2b: Compares the normalized energy deviation for randomly initialized parameters and parameters optimized by
h_cswith different training dataset sizes and truncation thresholds. - 2c: Shows the optimization dynamics of
h_cspre-training compared to conventional VQE, including post-training fine-tuning and verification on the quantum processor. - 2d: Analyzes the normalized MAE between quantum processor outputs and
h_cspredictions for various Ising couplings. - 2e: Demonstrates the normalized deviation for initialized vs. pre-trained parameters across a range of coupling strengths.
2. Reproducing Fig 3: Scaling Behavior of h_cs and h_qs for VQE Pre-training
To generate the visualizations for Figure 3, run the Jupyter notebook Plot_fig3.ipynb in the Fig3 directory.
What this figure shows:
- 3a: Presents the normalized deviation
𝔯(x)for theh_cssurrogate, comparing randomly initialized parameters and pre-trained parameters across different qubit counts (8, 12, 16, 20) with a fixed training data size of 1024. - 3b: Shows the normalized deviation
𝔯(x)for theh_qssurrogate, comparing initialized and pre-trained parameters across larger system sizes (18, 24, 30, 36, 42 qubits) with a fixed training size of 160.
3. Reproducing Fig 4: Characterizing Floquet Phase Diagrams with Predictive Surrogate h_qs
To generate the visualizations for Figure 4, run the Jupyter notebook Plot_fig4.ipynb in the Fig4 directory.
What this figure shows:
- 4a: Displays the time evolution of disorder-averaged local magnetizations predicted by
h_qsin both Floquet symmetry-protected topological (FSPT) and thermal phases. - 4b: Shows the Fourier transform of local magnetizations, highlighting the subharmonic oscillations at boundary qubits in the FSPT phase.
- 4c: Plots the subharmonic peak height as a function of disorder strength δ, comparing results from the quantum processor and the
h_qssurrogate, with the inset showing the standard deviation and identified phase transition region.
Data Availability & Reproducibility
The expdata/ subdirectories contain the processed experimental data used to generate each figure in the paper. Running the provided Jupyter notebooks with these data files will reproduce the exact figures as they appear in the manuscript.
Reproducibility Guarantee:
All figures regenerated from the provided data are pixel-identical to the versions published in the paper.
The data files encapsulate the essential results from experiments conducted on a 66-qubit superconducting quantum processor, ensuring that the findings can be faithfully validated.
Citation
If you find this work useful, please cite our paper:
@article{Liao2025demonstration,
author = "Liao, Wei-You and Du, Yuxuan and Wang, Xinbiao and Tian, Tian-Ci and Luo, Yong and Du, Bo and Tao, Dacheng and Huang, He-Liang",
title = "{Demonstration of Efficient Predictive Surrogates for Large-scale Quantum Processors}",
journal={arXiv preprint arXiv:2507.17470},
year = "2025"
}
License
This project is released under the MIT License. See LICENSE for details.



