🤖 AI

Artificial intelligence for representing and characterizing quantum systems

This review examines how artificial intelligence, particularly through machine learning, deep learning, and language models, addresses the challenge of efficiently characterizing large-scale quantum systems by enabling quantum property prediction and the construction of state surrogates, while also discussing key applications, challenges, and future prospects at the intersection of AI and quantum science.

Original authors: Yuxuan Du, Yan Zhu, Yuan-Hang Zhang, Min-Hsiu Hsieh, Patrick Rebentrost, Weibo Gao, Ya-Dong Wu, Jens Eisert, Giulio Chiribella, Dacheng Tao, Barry C. Sanders

Published 2026-08-04

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuxuan Du, Yan Zhu, Yuan-Hang Zhang, Min-Hsiu Hsieh, Patrick Rebentrost, Weibo Gao, Ya-Dong Wu, Jens Eisert, Giulio Chiribella, Dacheng Tao, Barry C. Sanders

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe is built from tiny, invisible building blocks called atoms, but at the smallest scale, these blocks don't just sit there; they dance in a strange, magical way. This is the world of quantum physics. In this realm, particles can be in many places at once, and they can be mysteriously linked across vast distances. Scientists have built powerful machines called quantum computers and simulators to study this dance, hoping to solve problems that are impossible for our regular computers. But here's the catch: as these machines get bigger, the "dance" becomes so complex that describing it is like trying to write down every single move of a billion dancers simultaneously. The amount of information needed grows so fast that even the world's fastest supercomputers get overwhelmed.

To make sense of this chaos, scientists are turning to a new kind of helper: Artificial Intelligence (AI). Think of AI as a super-smart detective that is really good at spotting patterns in huge piles of messy data. Instead of trying to write down every single rule of the quantum dance, the AI watches the dancers, learns their habits, and predicts what they will do next. This paper is a massive tour guide through the latest ways scientists are using these AI detectives to understand quantum systems. It explores three different "styles" of AI detectives—some that are good at simple math, some that are deep thinkers capable of learning complex structures, and some that are like super-reading machines that can guess the next step in a story. The goal is to figure out which detective is best for which job, helping us build better quantum computers and discover new secrets of nature.

The Big Picture: Why We Need AI for Quantum Machines

Scientists are currently building quantum machines that are getting bigger and more powerful, with some aiming to handle around 100 logical qubits (the quantum version of computer bits) at depths of about 10,000 steps. This is the "megaquop" era. However, as these machines grow, describing their behavior becomes a nightmare. The space of possibilities grows exponentially, meaning that for every new qubit added, the complexity doubles and doubles again. Traditional methods, like trying to simulate the system on a normal computer, hit a wall because they can't store or process this much information.

This is where AI steps in. Instead of trying to calculate every single possibility, AI models learn from data. They look at the results of experiments and learn to predict what will happen next, or to reconstruct what the quantum system looks like without needing to measure every single part of it. This review paper acts as a map, organizing the different ways AI is being used to tackle these problems. It breaks down the field into three main "paradigms" or styles of learning: Machine Learning (ML), Deep Learning (DL), and Language Models (LM).

The Three AI Detectives

The paper organizes these AI tools into a hierarchy, much like a toolbox where you pick the right tool for the job.

1. The Math Whizzes: Machine Learning (ML)
Think of ML models as the "math whizzes" of the group. They are excellent at finding straight-line relationships in data. If you want to predict a simple property, like how magnetized a material is or how much energy a system has, these models are very efficient.

How they work: They take data from experiments (like snapshots of the quantum system) and use clever math tricks to draw a line that fits the data. They are great at predicting "linear properties," which are things that add up nicely, like average energy or magnetization.
What they found: The paper shows that these models can predict these properties with high accuracy and have strict mathematical guarantees that they won't need too much data to do it. For example, they have successfully predicted properties for systems with up to 51 atoms or 127 qubits.
The Limit: However, the paper is very clear about a major limitation: these "math whizzes" struggle with complex, non-linear problems. If the quantum system behaves in a way that isn't a simple straight line (like certain types of quantum phases or entanglement), these models might fail. In fact, the paper suggests that for some very hard quantum problems, no classical AI model can solve them efficiently, and we might need actual quantum computers to learn the answers.

2. The Deep Thinkers: Deep Learning (DL)
If ML models are math whizzes, Deep Learning models are the "deep thinkers." They use complex neural networks (layers of artificial neurons) that can learn hidden patterns and structures that aren't obvious.

How they work: These models are great at two things. First, they can predict a wide variety of properties at once, not just simple ones. They can guess things like "entanglement" (how linked particles are) or "fidelity" (how close a real experiment is to the perfect theory). Second, they can act as "generative" models. Instead of just predicting a number, they can learn to create new data that looks exactly like the quantum system. This is like a forger who learns to paint a perfect copy of a masterpiece without ever seeing the original canvas.
What they found: These models have been used to reconstruct quantum states for systems up to 100 qubits and to classify different phases of matter (like distinguishing between a magnet and a non-magnet). They are also being used to fix errors in quantum computers, acting like a noise-canceling headphone for quantum data.
The Caveat: While they are powerful, the paper notes that we don't fully understand why they work so well. They are often "black boxes." Also, in some specific tests, simple ML models actually performed just as well as these complex deep learning models, suggesting that for some tasks, the extra complexity might not be necessary.

3. The Story Tellers: Language Models (LM)
The newest and most exciting group are the "Story Tellers," based on the same technology that powers chatbots like GPT. These are Large Language Models (LLMs) adapted for quantum physics.

How they work: Imagine a quantum system as a story. The "characters" are the qubits, and the "plot" is how they change over time. Language models are trained to read thousands of these "stories" (quantum data) and learn the grammar of quantum mechanics. They use a "pre-training" phase to learn the general rules of the universe, and then a "fine-tuning" phase to become experts on specific tasks.
What they found: These models are incredibly flexible. They can be trained on a huge variety of quantum systems and then quickly adapted to predict new things with very little extra data. Some models have been trained to predict the energy of complex systems or to learn the "Hamiltonian" (the rulebook of how the system evolves) just by looking at measurement data.
The Potential: The paper suggests these models could become "foundation models" for quantum science—universal tools that can be applied to almost any quantum problem, from simulating new materials to debugging quantum computers.

The Big Questions and Open Mysteries

The paper doesn't just list what works; it also highlights what we don't know yet. It poses several big questions that scientists are still trying to answer:

Can we do better than simple math? The paper asks if there are provably efficient AI models that can handle complex, non-linear quantum problems, or if we are stuck with simple predictions.
Do we need the quantum machine to learn? It explores whether we can learn about a quantum system just by looking at the data it produces (measurement-based) or if we need to interact with the machine in real-time. The paper suggests that for some tasks, interacting with the machine might be necessary to get the right answer.
Are the complex models actually better? While deep learning and language models sound fancy, the paper points out that in some head-to-head tests, they didn't always beat the simpler models. We need more rigorous testing to see when the complex models are truly worth the extra effort.
Can we build a universal quantum AI? The ultimate goal is a general-purpose AI that can learn any quantum system, but we aren't there yet. Current models are often good at specific families of systems but struggle to generalize to everything.

The Road Ahead

The paper concludes with an optimistic outlook. We are moving from the "toy model" phase, where scientists test ideas on small, simple systems, to the "real world" phase, where these tools are applied to massive, real quantum computers. The future involves building better datasets, creating standardized tests to compare AI models, and integrating AI directly into the hardware and software of quantum machines.

Ultimately, this review suggests that AI is not just a helper but a fundamental part of the future of quantum science. Whether it's a simple math model predicting a magnet's strength or a giant language model simulating a new material, these tools are essential for unlocking the secrets of the quantum world. As quantum machines grow to the "megaquop" scale, AI will be the bridge that allows us to understand and control them, turning the impossible complexity of the quantum dance into a story we can finally read.

Technical Summary: Artificial Intelligence for Representing and Characterizing Quantum Systems

Problem Statement

The characterization of large-scale quantum systems, particularly those generated by quantum analog simulators and "megaquop" quantum computers (systems with $\sim 100$ logical qubits and depths $\sim 10,000$ ), faces a fundamental barrier: the exponential scaling of the Hilbert space with system size. Classical simulators, such as tensor networks, fail to model highly entangled states, and extensions to non-Clifford circuits incur exponential computational costs. Consequently, describing and characterizing these systems using traditional methods is intractable. The paper addresses the challenge of developing efficient, data-driven methods to represent and characterize these scalable quantum systems, focusing on two core tasks: predicting quantum properties (linear and nonlinear) and constructing surrogates for quantum states.

Methodology

The review organizes recent advances (post-2022) into a methodological hierarchy of three synergistic AI paradigms: Machine Learning (ML), Deep Learning (DL), and Language Models (LMs). The authors adopt a task-oriented perspective, analyzing how each paradigm addresses specific characterization tasks.

1. Learning Paradigms and Workflow

All protocols generally follow a three-stage workflow (Fig. 3):

Data Collection: Gathering training datasets $\mathcal{T} = \{(x^{(i)}, s^{(i)}, z^{(i)})\}$ , where $x$ represents classical parameters (e.g., Hamiltonian couplings), $s$ represents measurement outcomes (shots), and $z$ is auxiliary information (e.g., gate layouts). Data collection strategies include measurement-agnostic (relying only on $x$ ) and measurement-based (requiring $s$ ) protocols.
Model Implementation & Optimization: Training models to learn patterns.
Model Prediction: Applying trained models to unseen systems.

2. Machine Learning (ML) Paradigm

Focus: Primarily linear property prediction (e.g., expectation values of observables) and phase classification.
Mechanism: Relies on feature engineering and kernel methods. Models map input parameters $x$ to a high-dimensional feature space where relationships become linear.
Key Techniques:
- Classical Shadows: Utilizing randomized measurements (e.g., Pauli-based) to construct "shadows" of quantum states, enabling efficient estimation of local observables.
- Kernel Machines: Using truncated Dirichlet kernels or positive good kernels to predict linear properties of Hamiltonian ground states.
- Fourier Analysis: For gate-based states (e.g., VQAs), models exploit the fact that quantum neural networks (QNNs) can be expanded as truncated Fourier series, allowing linear regression models to serve as classical surrogates.
Theoretical Guarantees: These models offer provable sample complexity bounds (often polynomial in system size $N$ ) for predicting linear properties of gapped, geometrically local Hamiltonians.

3. Deep Learning (DL) Paradigm

Focus: Nonlinear property prediction (e.g., entanglement entropy, fidelity), multi-property prediction, and implicit quantum state reconstruction.
Mechanism: Utilizes Deep Neural Networks (DNNs) to automatically extract latent representations from data, bypassing manual feature engineering.
Key Techniques:
- Discriminative Learning: For property prediction, using architectures like CNNs, GNNs, and RNNs to map measurement data or circuit parameters to physical properties.
- Generative Learning: For state reconstruction, employing Neural Quantum States (NQS). These include:
  - Autoregressive Models: (e.g., RNNs, Transformers) that factorize the joint probability of measurement outcomes.
  - Energy-Based Models: (e.g., RBMs) that assign energy to configurations.
- Multi-task Learning: Learning latent representations that can be transferred to predict diverse properties (e.g., using a model trained on spin correlations to distinguish topological phases).
- Applications: Quantum error mitigation (predicting error-mitigated observables), Hamiltonian learning, and quantum system benchmarking.

4. Language Model (LM) Paradigm

Focus: Foundation models for quantum systems, leveraging the pre-training and fine-tuning strategy of Generative Pre-trained Transformers (GPTs).
Mechanism:
- Pre-training: A large-scale, self-supervised stage where the model learns general structural patterns of quantum states (e.g., approximating measurement outcome distributions) from vast, unlabeled datasets.
- Fine-tuning: Adapting the pre-trained model to specific downstream tasks (e.g., predicting energy or magnetization) using smaller, labeled datasets.
Key Techniques:
- ShadowGPT / RydbergGPT: Models trained to emulate classical shadows or specific physical systems.
- Variational Foundation Models: Using transformers as variational ansätze to minimize variational energy without direct measurement data.
- Diffusion Models: Emerging approaches (e.g., QuaDiM) for state synthesis conditioned on Hamiltonian parameters.

Key Contributions and Results

The paper synthesizes a broad spectrum of literature, categorizing results by paradigm and task:

ML Results: Demonstrated that ML models with classical shadows can efficiently predict linear properties (e.g., local expectation values, correlation functions) for ground states of 51-atom Rydberg systems and 25-qubit Heisenberg models with rigorous sample complexity bounds. They also successfully classify topological phases in 200-qubit systems.
DL Results: Showed that DL models can predict nonlinear properties (e.g., entanglement entropy in 100-qubit systems) and perform implicit state reconstruction. DL models have been applied to benchmark quantum processors (50 qubits), mitigate errors in noisy circuits, and learn Hamiltonian parameters for superconducting processors (27 qubits).
LM Results: Introduced foundation models (e.g., GPT-based) capable of predicting linear properties of 45-qubit random Heisenberg models and ground states of Rydberg atom arrays. These models demonstrate strong transferability, allowing fine-tuning on specific tasks with minimal data.
Comparative Analysis: The review highlights that while DL models offer flexibility and handle nonlinearities, they do not always outperform simpler ML models (e.g., Random Forests) in specific tasks like Quantum Error Mitigation (QEM) when data is limited. The paper emphasizes the need for systematic benchmarking to determine the specific regimes where DL and LMs provide genuine advantages.

Significance and Claims

The paper positions itself as a consolidation of central findings to guide future research at the intersection of AI and quantum science. Its significance lies in:

Methodological Unification: It provides a structured taxonomy (ML, DL, LM) for understanding how different AI tools address the curse of dimensionality in quantum systems, moving beyond domain-specific applications to a unified theoretical framework.
Bridging Theory and Practice: It connects theoretical guarantees (e.g., sample complexity bounds for ML) with experimental realizations (e.g., IBM quantum computers, Rydberg simulators), showing that these methods are not merely theoretical but have immediate experimental impact.
Identifying Open Questions: The paper explicitly outlines six critical open questions, including:
- The existence of provably efficient ML models for nonlinear property prediction.
- The theoretical foundations of measurement-based protocols.
- The conditions under which DL and LMs strictly outperform traditional ML.
- The integration of physical symmetries (gauge, locality) into model architectures.
- The development of general-purpose foundation models for diverse quantum data.
Future Outlook: It argues that AI is essential for the "Megaquop" era, enabling the certification, benchmarking, and optimization of fault-tolerant quantum computers. The authors claim that AI will eventually shape quantum computing from software (algorithm design, compilation) to hardware (error correction, control), serving as an indispensable tool for unlocking the full computational power of quantum technologies.

The paper maintains a modest tone regarding the superiority of advanced models, acknowledging that simpler ML approaches often suffice for linear tasks and that the theoretical underpinnings of DL and LMs in quantum contexts require further rigorous investigation.