Rethink the Role of Deep Learning towards Large-scale Quantum Systems

TL;DR

This work reassesses the necessity of deep learning for learning ground-state properties and quantum phase classification in large-scale quantum systems under practical resource limits. By benchmarking DL against traditional ML across Heisenberg, TFIM, and Rydberg Hamiltonians up to 127 qubits with a unified total measurement budget, it demonstrates that ML frequently matches or surpasses DL in GSPE and QPC tasks. A randomization test reveals measurement inputs are often redundant for GSPE but can enhance QPC performance, suggesting task-dependent roles for input features. The findings guide future QSL model design toward data-efficient, resource-aware strategies, while SSL-based DL offers limited but context-dependent benefits; the work also contributes open-source data-generation tooling for reproducibility and further exploration.

Abstract

Characterizing the ground state properties of quantum systems is fundamental to capturing their behavior but computationally challenging. Recent advances in AI have introduced novel approaches, with diverse machine learning (ML) and deep learning (DL) models proposed for this purpose. However, the necessity and specific role of DL models in these tasks remain unclear, as prior studies often employ varied or impractical quantum resources to construct datasets, resulting in unfair comparisons. To address this, we systematically benchmark DL models against traditional ML approaches across three families of Hamiltonian, scaling up to 127 qubits in three crucial ground-state learning tasks while enforcing equivalent quantum resource usage. Our results reveal that ML models often achieve performance comparable to or even exceeding that of DL approaches across all tasks. Furthermore, a randomization test demonstrates that measurement input features have minimal impact on DL models' prediction performance. These findings challenge the necessity of current DL models in many quantum system learning scenarios and provide valuable insights into their effective utilization.

Figures (12)

• Figure 1: Scaling behavior of learning models when applied to GSPE tasks with $127$-qubit $|{\psi_{\rm{HB}}}\rangle$ and $|{\psi_{\rm{TFIM}}}\rangle$. The upper (lower) two subplots explore the scaling behavior of learning models for $M$ (for $n$) when applied to predicting $\epsilon(\bar{C})$ and $\epsilon(\bar{\mathcal{S}}_2)$, while keeping $n=100$ ($M=512$). The notation "$a$-$b$" means that the explored task is $a$ and the employed model is $b$.
• Figure 2: Role of measurements as input representations on GSPE tasks. RMSE $\epsilon(\bar{C})$ when applied to the employed learning models to predict correlations of $127$-qubit $|{{\psi}_{\rm{HB}}}\rangle$. Left panel: Performance of supervised learning models with varied training size $n$ and fixed snapshots $M=64$. Right panel: Performance of SSL-based model LLM4QPE-T with varied training size $n$ and snapshots $M$. The notation "$a$-A" (or "$a$") refers that the learning model $a$ uses (or does not use) $\boldsymbol{v}$ as auxiliary information.
• Figure 3: Performance of MLPs with varied model sizes on predicting correlation of $127$-qubit $|{\psi}_{\rm{HB}}\rangle$. The x-axis represents the number of parameters in MLPs with 'M' being the abbreviation of million. The notation $\lambda$ refers to regularization weights used in MLPs, and "A ($b$ M)" refers to the performance of model A (i.e., Lasso, and LLM4QPE) with $b$ million parameters.
• Figure 4: Scaling behavior of ML and DL models when applied to QPC tasks with ${|\psi_{\rm{Ryd}}\rangle}$. The left (right) subplots explore the scaling behavior of learning models for $M$ ($n$) when applied to the QPC task, while keeping $n=40$ ($M=256$). The shadow region refers to the standard deviation.
• Figure 5: Role of measurements on QPC tasks. The left and right panels separately exhibit the performance of RF and LLM4QPE-T when applied to learn $31$-qubit $|\psi_{\rm{Ryd}}\rangle$ with varied training size $n$ and snapshots $M$.