Research map / Architectures and efficiency / RNNs and normalization
What is my quantum computer good for? Quantum capability learning with physics-aware neural networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes quantum-physics-aware neural networks (qpa-NNs), GNN-inspired models that predict per-layer Hamiltonian/stochastic error rates and combine them with a parameter-free BCH-based function to predict PST or process fidelity of Clifford circuits, outperforming CNN capability models on 5-qubit IBM data and simulated 4- and 100-qubit devices.
Problem
Predicting which quantum programs (Clifford circuits) a real quantum computer can run reliably; existing off-the-shelf CNN/GNN capability models achieve only modest accuracy, especially with coherent errors, and do not scale well.
Approach
Embed quantum error physics into the network: a convolution-like, graph-structured network predicts rates of local H and S error generators per circuit layer, and a non-trainable function applies Clifford error propagation and a first-order BCH approximation to turn those rates into PST or process fidelity.
Key results
- On 5-qubit experimental data, qpa-NNs reduce MAE vs. CNNs by 50.4% on average (sigma 16.7%) and vs. fine-tuned CNNs by 32.2% on average.
- On 4-qubit simulated coherent-error data, qpa-NNs reduce MAE by 52.4% vs. hyperparameter-tuned CNNs, with Pearson r 0.968 vs 0.749.
- A qpa-NN for a simulated 100-qubit device achieves MAE of 0.097%, the first capability model reported for a 100+ qubit device.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use qpa-NN instead of CNN | Quantum capability prediction on the six experimental 5-qubit IBM datasets · 5-qubit experimental datasets | MAE: 50.4% average reduction | Helped |
| Use qpa-NN instead of fine-tuned CNN | Quantum capability prediction on the six experimental 5-qubit IBM datasets · 5-qubit experimental datasets | MAE: 32.2% on average | Helped |
| Fine-tune CNN on high-PST training data | CNN on 5-qubit experimental datasets · 5-qubit experimental datasets | MAE: 25.1% improvement | Helped |
| Use qpa-NN instead of CNN | Fidelity prediction on 4-qubit random i.i.d. circuits with coherent errors · 4-qubit random i.i.d. circuits | MAE: 52.4% reduction in MAE | Helped |
| Use qpa-NN instead of CNN | Fidelity prediction on 4-qubit random mirror circuits · 4-qubit random mirror circuits | MAE | Helped |
| Train qpa-NN on random circuits and test on random mirror circuits | qpa-NN · 4-qubit random mirror circuits (out of distribution) | MAE: 3.2 × increase in MAE | Hurt |
Surprising findings
- Fine-tuning CNNs on high-PST data only sometimes helps and the qpa-NNs still outperform fine-tuned CNNs on all six datasets, with fine-tuned CNN gains averaging 25.1%.
Limitations
- Assumes time-translation-invariant (Markovian) error rates; non-Markovian errors are not modeled.
- Only H and S error classes are included; amplitude damping and other classes are not modeled.
- Restricted to Clifford circuits; generalization to general circuits requires new approximate error-propagation methods.
Experimental setup
Datasets: 5-qubit IBM Q mirror-circuit datasets (ibmq_london, ibmq_essex, ibmq_burlington, ibmq_vigo, ibmq_ourense, ibmq_yorktown) simulated 4-qubit ring random circuits simulated 4-qubit random mirror circuits simulated 100-qubit ring random circuits
Benchmarks: 5-qubit IBM Q experimental datasets (PST) 4-qubit simulated random and mirror circuits (process fidelity) 100-qubit simulated random circuits (process fidelity)
Metrics: mean absolute error Bayes factor Pearson correlation coefficient
Baselines: CNN (Ghosh et al. capability model) fine-tuned CNN (ft-CNN) hyperparameter-tuned CNN on simulated data
Builds on: convolutional neural networks for capability learning graph convolution elementary error generator framework stim Clifford simulation
Closest papers by meaning
- Rethink the Role of Deep Learning towards Large-scale Quantum Systems (ICML 2025 · similarity 0.85)
- A Unified Theory of Quantum Neural Network Loss Landscapes (ICLR 2025 · similarity 0.83)
- ATP: Adaptive Threshold Pruning for Efficient Data Encoding in Quantum Neural Networks (CVPR 2025 · similarity 0.83)
- Quantum Deep Equilibrium Models (NeurIPS 2024 · similarity 0.83)
- PALQO: Physics-informed model for Accelerating Large-scale Quantum Optimization (NeurIPS 2025 · similarity 0.83)
- Quantum Positional Encodings for Graph Neural Networks (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.