What is my quantum computer good for? Quantum capability learning with physics-aware neural networks

TL;DR

The paper tackles the challenge of rapidly predicting which quantum circuits a noisy device can reliably execute. It introduces quantum-physics-aware neural networks (qpa-NNs) that predict local error rates using a graph-inspired structure and then combine these predictions with an efficient BCH-based aggregation to estimate circuit success metrics $PST(c)$ or $F(c)$. Empirically, qpa-NNs significantly outperform convolutional baselines on 5-qubit experiments and 4-qubit/simulated data, and they demonstrate scalability to 100 qubits with feasible accuracy, particularly in modeling coherent errors. The approach offers a principled, scalable method for quantum capability learning and holds promise for fast device characterization and circuit diagnosis in practical quantum computing settings.

Abstract

Quantum computers have the potential to revolutionize diverse fields, including quantum chemistry, materials science, and machine learning. However, contemporary quantum computers experience errors that often cause quantum programs run on them to fail. Until quantum computers can reliably execute large quantum programs, stakeholders will need fast and reliable methods for assessing a quantum computer's capability-i.e., the programs it can run and how well it can run them. Previously, off-the-shelf neural network architectures have been used to model quantum computers' capabilities, but with limited success, because these networks fail to learn the complex quantum physics that determines real quantum computers' errors. We address this shortcoming with a new quantum-physics-aware neural network architecture for learning capability models. Our architecture combines aspects of graph neural networks with efficient approximations to the physics of errors in quantum programs. This approach achieves up to $\sim50\%$ reductions in mean absolute error on both experimental and simulated data, over state-of-the-art models based on convolutional neural networks.

Figures (4)

• Figure 1: Quantum capability learning with quantum-physics-aware neural networks (qpa-NNs). Our qpa-NNs are a novel architecture for learning a quantum computer's capability, i.e., the mapping from quantum circuits (or programs) to how well that imperfect quantum computer can run those circuits. These networks build in physical principles for how errors in quantum circuits occur---which can be expressed in terms of a quantum computer's connectivity graph---and efficient approximations to the physics of how these errors combine to impact a circuit's success rate.
• Figure 2: Prediction accuracy on real quantum computers.(a) The mean absolute error of our qpa-NNs ($\textcolor{blue}{\bm{\bullet}}$), the CNNs from hothem2023learning (o-CNN, $\textcolor{orange}{\bm{+}}$), and fine-tuned CNNs (ft-CNN, $\textcolor{green}{\blacklozenge}$) on the test data. (b) The predictions of the three models for $\mathtt{ibmq\_vigo}$ on the test data, and (c) the distribution of each model's absolute error on the test data, including the 50th, 75th, 95th and 100th percentiles (lines) and the means (points).
• Figure 3: Demonstrating our qpa-NNs' accuracy for hard-to-model coherent errors and at scale.(a) Scatter plot of the prediction errors on test data of a qpa-NN ($\textcolor{blue}{\bm{\bullet}}$) and CNN ($\textcolor{orange}{\bm{\bullet}}$) trained to predict the fidelity $F(c)$ of random circuits run on a hypothetical 4-qubit quantum computer. The qpa-NN significantly outperforms the CNN. The top subplot contains a histogram (green bars) of the ground-truth fidelities. (b) Prediction errors on out-of-distribution test data, from random mirror circuits. The qpa-NN achieves modest prediction accuracy on this out-of-distribution task, suggesting that the qpa-NNs are accurately learning error rates. (c) Prediction errors on the 100-qubit test data, demonstrating that our qpa-NN approach can accurately predict $F(c)$ for circuits run on large-scale quantum computers.
• Figure 4: Device geometries. The connectivity graphs for the (a) 5-qubit $\mathtt{ibmq\_yorktown}$ "bowtie" processor; (b) the remaining 5-qubit experimental "t-bar" processors; and (c) the 4-qubit simulated "ring" processor. The 100-qubit simulated "ring" processor has the same topology, just more qubits.