Artificial intelligence for representing and characterizing quantum systems

arXiv:2509.04923 · quant-ph, cs.AI, cs.LG · Submitted 2025-09-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai, and with me are Mira and Lev, guest researcher.

Mira: Today's paper: "Artificial intelligence for representing and characterizing quantum systems".

Kai: Detailed Research Summary: Artificial Intelligence for Representing and Characterizing Quantum Systems This review meticulously examines the burgeoning field of applying Artificial Intelligence (AI)—specifically machine learning (ML), deep learning (DL),

Mira: First, who's behind it and why it matters.

Title and authors: Kai: So we’ve touched on what this paper is trying to accomplish—representing and characterizing quantum systems using various AI methods—but let's talk about the actual title and who wrote it. The authors are a mix of folks from different strong institutes, which is always interesting in scientific collaboration.

Mira: I noticed the title highlights both "representing" and "characterizing," which suggests they aren't just doing one thing, but aiming for a comprehensive capability to understand these complex quantum states. It frames the entire endeavor as moving toward a more holistic understanding of what’s happening in these systems.

Lev: Collaboration across institutions like Nanyang Technological University and UC San Diego is significant because it implies that the methodology they developed has broad applicability, rather than being tied to one specific lab's hardware setup or theoretical model.

Kai: It really does show that this isn't just a niche idea; these are people from different backgrounds working together on something fundamental to how we can handle quantum complexity.

Mira: The implications of this title for the field is that it suggests AI isn't just a tool for minor tweaks anymore, but a potential new way to approach the entire problem of understanding large quantum systems from scratch.

Lev: That comprehensive framing helps when we think about scaling up; if you can characterize both linear and nonlinear properties using this framework, it opens avenues for testing different physical theories against simulated data.

Kai: So, rather than just looking at one specific property like energy or entanglement, they're aiming to build a system that can map out the landscape of these quantum states itself.

Mira: Precisely; it’s about building a language for quantum states, which is a very ambitious goal when you think about how complex those state spaces are.

Lev: If this framework proves robust, it could be used to rapidly screen candidate Hamiltonians or physical models before we even commit to expensive simulation time on actual quantum hardware.

Kai: It’s moving us toward a future where we can use AI as a primary interface for navigating the vast complexity of quantum physics.

The paper's summary: Kai: Now let's look at what the paper actually summarizes about this work. They outline how they categorize the tasks—predicting linear properties, nonlinear properties, and state reconstruction—which is a very structured way to approach this whole characterization challenge.

Mira: That structure is helpful because it allows them to deploy different AI tools for different parts of the problem; you use one kind of model for simple linear things and another for more intricate nonlinear phenomena.

Lev: I'm particularly interested in how they define those linear properties, because if we can get a provably efficient ML model for those, that gives us a baseline where we know what to expect from any physical system.

Kai: The paper states they've designed provably efficient machine learning models specifically to characterize the linear properties of scalable quantum systems and classify quantum phases. That’s a concrete achievement they are highlighting here.

Mira: And then they point out that deep learning models offer tools for predicting a wide range of properties through representation learning, alongside generative modeling for implicitly reconstructing quantum states using those generative approaches.

Lev: Implicit state reconstruction is where things get tricky; if the model approximates the probability distribution of measurement outcomes without needing to output a full density matrix, it tackles one of the biggest computational bottlenecks in this whole area.

Kai: So, they are showing that AI can handle everything from simple linear predictions to reconstructing complex quantum states using these different machine learning techniques.

Mira: The implication is that we can use these models not just to discover new things, but also for certification and benchmarking of existing systems, which is a very practical application right now.

Lev: That capability to perform direct fidelity estimation based on learned structures would be incredibly useful when we start running experiments on real quantum devices.

Kai: So, the summary really emphasizes the versatility of AI in this context, covering prediction and reconstruction across different levels of complexity in one study.

The paper's improvements: Mira: Moving beyond what they summarized, the paper also lays out some specific improvements they suggest for these different AI paradigms. They are suggesting we embed intrinsic physical knowledge directly into the model structures to get better learning performance and more interpretable results.

Kai: That’s where I see real potential; simply having a large dataset isn't enough; we need the models to actually respect the rules of quantum mechanics, like symmetry and locality when predicting properties.

Lev: Incorporating symmetry explicitly into network architectures would make sense from a physical perspective because it enforces constraints that should hold true regardless of how much data you feed it.

Mira: They suggest embedding fundamental symmetries such as permutation symmetry, gauge symmetries, locality constraints, and Lie-algebraic properties directly into the network architectures to enforce invariance in measurement outcomes across multiple shots.

Kai: That sounds like a necessary step for making these models reliable for real experimental data; if the model doesn't respect those rules, its predictions are just noise dressed up nicely.

Lev: For my research, I think this structural incorporation of locality assumptions is crucial because in many physical systems, the way interactions are local dictates the overall behavior.

Mira: They also address generalization challenges by noting that a general-purpose foundation model capable of learning from diverse data types—like circuits and Hamiltonians—is still something they haven't fully realized yet.

Kai: That limitation is important to acknowledge because it means we can’t just expect one single AI to solve every quantum problem perfectly across all possible systems.

Lev: If the paper manages to show knowledge transfer between related tasks, that would be a big step toward building those more general models you mentioned.

Conclusion: Kai: So, wrapping up this discussion on "Artificial intelligence for representing and characterizing quantum systems," we've seen how they've laid out the structure of using AI across linear prediction, nonlinear prediction, and state reconstruction. The paper shows a lot of promise in its ability to handle the complexity that comes with large quantum systems.

Mira: Ultimately, what I see is that this work moves us toward a more sophisticated representation where AI can act as a surrogate for computationally expensive simulations while also providing insights into phase competition and order parameters.

Lev: From my side, I think the explicit integration of physical knowledge into the models is the most critical piece because it’s what will make these AI tools reliable enough to use in any kind of experimental setting.

Kai: It sounds like we're looking at a framework that can help us predict things faster and give us a better picture of what's happening in complex quantum systems, which is exactly the kind of practical application we need right now.

Mira: If this approach scales as they hope, it implies that AI will become an indispensable part of the toolkit for anyone trying to understand and control these increasingly large quantum systems.

Lev: I just want to say that while the paper shows great potential, we have to keep pushing the discussion on how these models can handle real-world noise and experimental imperfections in our error correction protocols.

Kai: So, we've explored what this paper proposes regarding "Artificial intelligence for representing and characterizing quantum systems" and its path forward. It’s clear that the direction is toward smarter, more structured AI tools to manage the scale of quantum physics.

College of Computing and Data Science, Nanyang Technological University, Singapore · QICI Quantum Information and Computation Initiative, Department of Computer Science, The University of Hong Kong · Department of Physics, University of California, San Diego · Hon Hai (Foxconn) Research Institute · Centre for Quantum Technologies, National University of Singapore · Department of Computer Science, National University of Singapore

quant-ph, cs.AI, cs.LG

Submitted: 2025-09-05

Updated: 2025-09-05

Comments: 32 pages. Comments are welcome

Journal ref: Nature Reviews Physics 8, 579 (2026)

DOI: 10.1038/s42254-026-00962-5

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: This review meticulously examines the burgeoning field of applying Artificial Intelligence (AI)—specifically machine learning (ML), deep learning (DL), and language models (LMs)—to efficiently

Key concepts

Hilbert Space Scaling
The number of possible states in a quantum system grows exponentially with the size of the system. This makes simulating or fully describing large systems computationally impossible for classical computers, necessitating AI methods to manage this massive complexity.
Quantum Property Prediction
This involves training AI models to estimate specific physical characteristics of a quantum state, such as its energy levels or phase. ML models are used for linear properties, while DL and LMs are explored for predicting more complex non-linear behaviors.
Neural Quantum States (NQS)
These are quantum states represented by deep learning neural networks. They can be explicitly defined or learned implicitly by a model, allowing the AI to approximate the true quantum state distribution based on experimental measurement outcomes.

Terminology

Summary

This review meticulously examines the burgeoning field of applying Artificial Intelligence (AI)—specifically machine learning (ML), deep learning (DL), and language models (LMs)—to efficiently characterize large-scale quantum systems, which are increasingly generated by quantum analog simulators and megaquop quantum computers. The central difficulty lies in the exponential scaling of the Hilbert space with respect to system size. AI has emerged as a potent tool to tackle this challenge by leveraging its strengths in high-dimensional pattern recognition and function approximation.

The overarching goal of this research is to enable AI models to represent and characterize scalable quantum systems in a data-driven manner, focusing on two primary tasks: quantum property prediction and the construction of surrogates for quantum states (both implicit and approximate reconstruction). These capabilities underpin critical applications ranging from quantum certification, benchmarking, and algorithm enhancement to gaining deeper insights into strongly correlated phases of matter.

The integration of AI into this characterization space is categorized into three synergistic paradigms: ML, DL, and LMs. The review structures the progress along a methodological hierarchy based on these models and the specific tasks they address: (1) predicting linear properties of quantum systems, (2) predicting instances of non-linear properties, and (3) reconstructing quantum states and processes.

1. Machine Learning (ML): Predicting Linear Properties

ML models are primarily focused on characterizing the linear properties of quantum states. A general scheme involves transforming raw quantum data into a labeled dataset (TML) where (i) represents the estimated physical property. Training often utilizes regression-based or kernel-based methods, such as h ML(x; w) = w, phi(x). A significant advantage of ML models is their ability to operate in a measurement-agnostic protocol, allowing for efficient characterization without requiring direct quantum data input. Concrete examples include models designed for predicting linear properties of Hamiltonian ground states and classifying these states into different quantum phases.

2. Deep Learning (DL): Property Prediction and State Reconstruction

Deep learning models offer powerful capabilities across both property prediction and state reconstruction, utilizing a discriminative learning framework for the former and a generative learning paradigm for the latter.

  • Property Prediction: DL models are trained on datasets (TDL) derived either from direct measurement outcomes (s(i)) or by incorporating auxiliary information (z(i)).

  • State Reconstruction: For state reconstruction, DL models frequently employ Neural Quantum States (NQS), which can be explicitly defined or implicitly learned. Implicit reconstruction is achieved through generative modeling, where the model approximates the probability distribution of measurement outcomes Q(s; theta) by minimizing a negative log-likelihood loss function (L(theta)).

3. Language Models (LMs): Foundation Models for Quantum Systems

Language models, building upon the GPT architecture, provide a flexible framework for auto-regressive representation of large families of quantum states. Their learning protocol typically involves two stages:

  • Pre-training: Self-supervised training to capture generalizable patterns across diverse quantum states, often by emulating measurement outcomes.

  • Fine-tuning: Adaptation to specific property prediction tasks using labeled datasets.

LMs represent a significant leap, paving the way for foundation models for quantum systems.

A critical theme emerging in the literature is the necessity of integrating intrinsic physical knowledge into model architectures. The review strongly emphasizes that existing approaches are often solely data-driven, lacking explicit leverage of underlying physical structures. To enhance learning performance and improve interpretability, research is focusing on:

  • Symmetry Incorporation: Explicitly embedding fundamental symmetries such as permutation symmetry, gauge symmetries, locality constraints, and Lie-algebraic properties directly into network architectures to enforce invariance in measurement outcomes across multiple shots.

  • Generalization Challenges: Despite progress on specific quantum families (e.g., knowledge transfer between related tasks), a general-purpose foundation model capable of learning from diverse data types (measurement outcomes, circuits, Hamiltonians) remains unrealized.

The research identifies several pressing challenges:

  1. Determining the existence and efficiency of ML models for broader tasks like non-linear property prediction.

  2. Establishing whether advanced DL methods offer demonstrable advantages over classical ML models for specific quantum tasks.

  3. Understanding the theoretical foundations of classical learning models within measurement-based protocols.

Looking forward, the transition will move from conceptual development to large-scale implementation targeting realistic and experimentally relevant systems.

Improvements for AI systems

Here are specific, actionable improvements to existing AI systems based on the insights from this review, categorized by paradigm:

) Improvements for Machine Learning (ML) Paradigm:

  1. Implement geometrically-informed feature maps (e.g., Lasso regression or Dirichlet kernels) tailored to the specific geometry of the Hamiltonian family being studied.

  2. Develop provably efficient ML models that explicitly exploit structural constraints like locality, smoothness, and intrinsic symmetry when predicting linear properties of Hamiltonian ground states.

  3. "For phase classification tasks (e.g., topologically ordered phases), integrate nonlinear classifiers with feature maps that incorporate arbitrarily large reduced density matrices to ensure a rigorous guarantee for phase distinction."

  4. Design measurement-based ML models that are computationally efficient, ensuring they achieve sample and computational efficiency in the prediction stage, even when compared against quantum learning models.

) Improvements for Deep Learning (DL) Paradigm:

  1. "For property prediction tasks, develop modular DL architectures (e.g., Graph Neural Networks - GNNs or specialized CNNs) that can efficiently handle multi-task learning by learning shared latent representations capable of predicting multiple physical properties simultaneously."

  2. "For quantum state reconstruction, prioritize implicit reconstruction using autoregressive models (like Transformers) to learn the underlying probability distribution of measurement outcomes, rather than relying on explicit density matrix output which suffers from exponential scaling."

  3. "In quantum error mitigation (QEM), develop DL models based on measurement-based protocols that utilize auxiliary information to create task-specific classifiers capable of mitigating noise errors in near-term experimental implementations."

  4. "For Hamiltonian learning, employ specialized DL architectures (e.g., FCNNs trained on local measurements) to learn and verify the structure of specific Hamiltonians (like stabilizer Hamiltonians) under geometric locality assumptions."

  5. "Integrate transfer learning strategies into DL models to enable them to generalize from small-scale, classically simulatable systems to larger, more computationally challenging quantum regimes (e.g., predicting phase diagrams of larger systems not seen during initial training)."

) Improvements for Language Model (LM) Paradigm:

  1. Utilize Foundation Models (GPT-like architectures) pre-trained on diverse quantum data corpora to capture generalizable structural features across various quantum states and measurement settings.

  2. "Implement a robust pre-training/fine-tuning pipeline where the foundation model is fine-tuned on smaller, labeled datasets containing specific task information (e.g., entanglement entropy data or noise profiles) to specialize in precise property prediction."

  3. Develop specialized GPT variants (e.g., ShadowGPT, RydbergGPT) that are pre-trained solely to emulate the behavior of Pauli-based classical shadows for predicting ground state energy and correlation functions.

  4. Employ diffusion models conditioned on Hamiltonian parameters to unify property prediction with state synthesis, allowing a single model to approximate ground states across an entire phase diagram.

) Overall System Capability Improvements:

The improved AI system can perform the following tasks:

  1. Predict linear properties (energy, magnetization, correlation functions) of quantum ground states and digital circuit output states with provable efficiency bounds derived from the learned structure.

  2. Classify complex quantum phases of matter (including topological phases) by leveraging both linear and nonlinear property prediction capabilities.

  3. Implicitly reconstruct high-dimensional quantum states (like GHZ or Ising models) by learning the underlying distribution of measurement outcomes, bypassing the exponential cost of explicit density matrix reconstruction.

  4. Act as a surrogate optimizer for Variational Quantum Algorithms (VQAs) by predicting gradient trajectories or suggesting high-quality initial parameters to accelerate optimization efficiency.

  5. Perform quantum system benchmarking and predict processor performance metrics (fidelity, bond dimension) based on physical system characteristics and simulation results.

Abstract

Efficient characterization of large-scale quantum systems, especially those produced by quantum analog simulators and megaquop quantum computers, poses a central challenge in quantum science due to the exponential scaling of the Hilbert space with respect to system size. Recent advances in artificial intelligence (AI), with its aptitude for high-dimensional pattern recognition and function approximation, have emerged as a powerful tool to address this challenge. A growing body of research has leveraged AI to represent and characterize scalable quantum systems, spanning from theoretical foundations to experimental realizations. Depending on how prior knowledge and learning architectures are incorporated, the integration of AI into quantum system characterization can be categorized into three synergistic paradigms: machine learning, and, in particular, deep learning and language models. This review discusses how each of these AI paradigms contributes to two core tasks in quantum systems characterization: quantum property prediction and the construction of surrogates for quantum states. These tasks underlie diverse applications, from quantum certification and benchmarking to the enhancement of quantum algorithms and the understanding of strongly correlated phases of matter. Key challenges and open questions are also discussed, together with future prospects at the interface of AI and quantum science.

Sources

Related papers