Research map / Probabilistic and scientific modeling / Differential equations and PDEs
QuanONet: Quantum Neural Operator with Application to Differential Equation
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 QuanONet, a pure-quantum-circuit neural operator (and trainable-frequency variant TF-QuanONet) for learning differential equation solution operators, backed by a quantum universal approximation theorem and tested in simulation plus a 2-qubit IBM hardware run.
Problem
Learning nonlinear operators that map input functions to differential equation solutions is not well studied in quantum machine learning; existing quantum approaches solve specific equations or use hybrid quantum-classical neural operators.
Approach
A purely quantum branch-trunk circuit with hardware-efficient ansatz whose output is an inner product of input-dependent quantum states (justified by a new quantum universal approximation theorem for operators), plus trainable encoding frequencies (TF) so the circuit can adapt its spectrum to the dominant operator frequency.
Key results
- TF-QuanONet reaches test error about 1.1e-4 on the homogeneous operator and 1.1e-4 to 1.2e-4 on the ODE-family tasks, versus about 6e-4 for DeepONet at similar parameter counts (Table 1).
- On the diffusion-reaction system, TF-QuanONet achieves about 5.1e-5 to 5.5e-5 test error versus 8.8e-5 for DeepONet and 6e-4 for HEA-based QNN.
- A 2-qubit TF-QuanONet run on IBM brisbane hardware achieved MSE 1.57e-3 for the anti-derivative operator, compared with 1.5e-5 in simulation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use trainable-frequency encoding in place of fixed-frequency encoding | QuanONet · Anti-derivative operator | test error: 0.001341 ± 0.000190 -> 0.000113 ± 0.000010 | Helped |
| Use trainable-frequency encoding in place of fixed-frequency encoding | QuanONet · Homogeneous operator | test error: 0.002005 ± 0.000253 -> 0.000192 ± 0.000028 | Helped |
| Use trainable-frequency encoding in place of fixed-frequency encoding | QuanONet · Nonlinear ODE | test error: 0.059582 ± 0.008323 -> 0.039072 ± 0.007277 | Helped |
| Use trainable-frequency encoding in place of fixed-frequency encoding | QuanONet · Diffusion-reaction system | test error: 0.000347 ± 0.000015 -> 0.000601 ± 0.000033 | Hurt |
| Change λ from 0.001 to 0.1 (ablation) | QuanONet · Anti-derivative operator | test error: 0.131876 ± 0.003809 -> 0.001341 ± 0.000190 | Helped |
| Change λ from 0.001 to 0.1 (ablation) | QuanONet · Homogeneous operator | test error: 0.330244 ± 0.005671 -> 0.002005 ± 0.000253 | Helped |
| Change λ from 0.001 to 0.1 (ablation) | QuanONet · Nonlinear ODE | test error: 0.869680 ± 0.031546 -> 0.059582 ± 0.008323 | Helped |
| Change λ from 0.001 to 0.1 (ablation) | QuanONet · Diffusion-reaction system | test error: 0.000601 ± 0.000033 -> 0.001802 ± 0.001053 | Mixed |
| Change λ across 0.001, 0.01, and 0.1 (ablation) | TF-QuanONet · Anti-derivative operator | test error: 0.000113 ± 0.000010; 0.000121 ± 0.000010; 0.000112 ± 0.000008 | No clear effect |
| Change λ across 0.001, 0.01, and 0.1 (ablation) | TF-QuanONet · Homogeneous operator | test error: 0.000192 ± 0.000028; 0.000198 ± 0.000013; 0.000181 ± 0.000021 | No clear effect |
| Change λ across 0.001, 0.01, and 0.1 (ablation) | TF-QuanONet · Nonlinear ODE | test error: 0.039072 ± 0.007277; 0.039469 ± 0.008081; 0.039074 ± 0.007452 | No clear effect |
| Increase λ to 0.1 (ablation) | TF-QuanONet · Diffusion-reaction system | test error: 0.000051 ± 0.000010 -> 0.000145 ± 0.000028 | Hurt |
Surprising findings
- Plain QuanONet with small coefficient lambda (0.001) performs much worse than with lambda=0.1, showing sensitivity to input scaling, which trainable frequency removes.
- On the nonlinear ODE, quantum methods generally underperform classical baselines, which the authors attribute to a more spread-out frequency spectrum requiring deeper circuits.
Limitations
- Experiments are noiseless simulation for most results; hardware validation is limited to one 2-qubit anti-derivative case with a large error gap.
- Quantum methods underperform classical baselines on the nonlinear ODE task.
- Performance depends on the input scaling coefficient lambda for the non-trainable-frequency version.
Experimental setup
Datasets: Synthetic Gaussian random field input functions with numerical PDE/ODE solutions
Benchmarks: Anti-derivative operator Homogeneous operator Nonlinear ODE Diffusion-reaction system
Metrics: test error (MSE) standard deviation over 5 runs
Baselines: HEA QNN TF-HEA DeepONet FNN quantum DeepONet (Xiao et al., 2024a)
Builds on: DeepONet Trainable-frequency quantum models (Jaderberg et al., 2024) Quantum neural tangent kernel frequency principle (Xu & Zhang, 2024) Universal approximation theorem for operators (Chen & Chen, 1995) Schuld et al. 2021 quantum models as Fourier series
Closest papers by meaning
- Neural Quantum Spectral Operator Learning for Solving Partial Differential Equations (NeurIPS 2026 · similarity 0.88)
- Q-MAML: Quantum Model-Agnostic Meta-Learning for Variational Quantum Algorithms (AAAI 2025 · similarity 0.84)
- PALQO: Physics-informed model for Accelerating Large-scale Quantum Optimization (NeurIPS 2025 · similarity 0.84)
- One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective (NeurIPS 2026 · similarity 0.83)
- On the Relation between Trainability and Dequantization of Variational Quantum Learning Models (ICLR 2025 · similarity 0.83)
- Exploring channel distinguishability in local neighborhoods of the model space in quantum neural networks (ICLR 2025 · 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.