Research map / Large language models / AI and society
Rethink the Role of Deep Learning towards Large-scale Quantum Systems
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.
Benchmarks deep learning against classical ML models for ground-state property estimation and quantum phase classification under equal quantum-measurement budgets (up to 127 qubits), finding ML often matches or beats DL and measurement inputs are largely redundant for DL on GSPE.
Problem
Whether deep learning models are necessary for learning ground-state properties and phases of quantum systems, given that prior comparisons used unequal quantum measurement resources.
Approach
Impose a unified total-measurement budget (n x M) on the training dataset for all ML and DL models, then systematically compare them on Heisenberg, TFIM and Rydberg Hamiltonians, plus a randomization test of measurement inputs.
Key results
- On 127-qubit Heisenberg correlation prediction with n=100, M=512, Lasso and Ridge reach RMSE 0.011 and 0.012 versus 0.017 for LLM4QPE-F.
- On 31-qubit Rydberg phase classification, RF increases test accuracy from 93.57% to 98.36% when measurement outcomes are included as input (n=100, M=256).
- ML models (Lasso/Ridge, RBF kernel, tree ensembles) match or outperform DL models (MLP, CNN, LLM4QPE-T) in most resource-matched settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Lasso instead of LLM4QPE-F | Heisenberg correlation prediction · 127-qubit Heisenberg, n=100, M=512 | RMSE: 0.017 -> 0.011 | Helped |
| Use Ridge instead of LLM4QPE-F | Heisenberg correlation prediction · 127-qubit Heisenberg, n=100, M=512 | RMSE: 0.017 -> 0.012 | Helped |
| Use measurement outcomes as auxiliary input features (ablation) | Ridge · 127-qubit Heisenberg correlation prediction, n=60 | RMSE: 0.032 -> 0.066 | Hurt |
| Use measurement outcomes as auxiliary input features (ablation) | CNN · 127-qubit Heisenberg correlation prediction, n=60 | RMSE: 0.031 -> 0.032 | Hurt |
| Replace real measurement outcomes with randomized outcomes (ablation) | LLM4QPE-T · 127-qubit Heisenberg correlation prediction | RMSE | No clear effect |
| Use measurement outcomes as auxiliary information (ablation) | RandomForest · 31-qubit Rydberg phase classification, n=100, M=256 | accuracy: 93.57% -> 98.36% | Helped |
| Replace real measurement outcomes with randomized outcomes (ablation) | LLM4QPE-T · 31-qubit Rydberg phase classification, n=100, M=64 | accuracy: 85.36% -> 66.93% | Hurt |
| Increase training size and number of snapshots | ML and DL models · Ground-state property estimation and quantum phase classification | prediction performance | Helped |
| Increase MLP model size when regularization is small (ablation) | MLP · 127-qubit Heisenberg correlation prediction, n=100, M=512 | RMSE | Hurt |
| Increase MLP model size when regularization is large (ablation) | MLP · 127-qubit Heisenberg correlation prediction, n=100, M=512 | RMSE | No clear effect |
Surprising findings
- Randomizing measurement outcomes barely changes DL performance on GSPE tasks, suggesting measurement features are largely redundant there.
- Shadow Generator performs only comparably to classical shadows on correlation prediction.
Limitations
- Only three Hamiltonian families and a limited set of DL architectures (MLP, CNN, LLM4QPE) were benchmarked.
- Quantum resource cost is measured only as total number of measurements, ignoring other costs.
- Hyperparameter tuning and training budgets for DL models are only briefly described.
Experimental setup
Datasets: simulated ground states of Heisenberg, transverse-field Ising, and Rydberg atom Hamiltonians with classical-shadow measu
Benchmarks: Heisenberg correlation prediction (up to 127 qubits) TFIM entanglement entropy prediction Rydberg quantum phase classification (31 qubits)
Metrics: RMSE test accuracy
Baselines: Lasso Ridge Dirichlet kernel RBF kernel NTK Random Forest XGBoost LightGBM GradientBoostingTree MLP CNN LLM4QPE
Builds on: classical shadow LLM4QPE Huang et al. provably efficient ML for quantum many-body problems
Closest papers by meaning
- Predicting Ground State Properties: Constant Sample Complexity and Deep Learning Algorithms (NeurIPS 2024 · similarity 0.89)
- Quantum Deep Equilibrium Models (NeurIPS 2024 · similarity 0.88)
- Neural Auto-designer for Enhanced Quantum Kernels (ICLR 2024 · similarity 0.88)
- Reinforced Learning Explicit Circuit Representations for Quantum State Characterization from Local Measurements (ICML 2025 · similarity 0.87)
- One More Time: Revisiting Neural Quantum States from a Reinforcement Learning Perspective (NeurIPS 2026 · similarity 0.87)
- QCircuitBench: A Large-Scale Dataset for Benchmarking Quantum Algorithm Design (NeurIPS 2025 · similarity 0.87)
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.