PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence and Machine Learning for Quantum Technologies

Mario Krenn, Jonas Landgraf, Thomas Foesel, Florian Marquardt

arXiv:2208.03836v1quant-phcs.AIcs.LG

TL;DR

This perspective addresses how machine learning and AI can help improve quantum technologies. It surveys selected illustrative applications and concludes that these methods offer useful approaches, while their choice should fit the task.

  • Problem

    How machine learning and AI can be applied to quantum technologies is a rapidly growing interface that the review presents to physicists without requiring prior machine-learning background.

    Source: §I · p. 1

  • Method

    The perspective surveys selected illustrative examples of classical machine learning and AI applications across quantum-technology domains.

    Source: §I · p. 1

  • Results

    The examples span quantum measurement analysis, device-parameter estimation, feedback and control, and discovery of experiments, protocols, and circuits.

    Source: §I · p. 1, A. Measurement data analysis and… · p. 6, B. Parameter estimation: learning the… · p. 9

  • Takeaways & Limitations

    The review emphasizes choosing an AI approach suited to the problem, since simple methods can outperform deep learning when it is misapplied.

    Source: §I · p. 1

    The literature selection is subjective and the review does not claim to provide a comprehensive list of work.

    Source: §I · p. 1

Abstract

from arXiv · show

In recent years, the dramatic progress in machine learning has begun to impact many areas of science and technology significantly. In the present perspective article, we explore how quantum technologies are benefiting from this revolution. We showcase in illustrative examples how scientists in the past few years have started to use machine learning and more broadly methods of artificial intelligence to analyze quantum measurements, estimate the parameters of quantum devices, discover new quantum experimental setups, protocols, and feedback strategies, and generally improve aspects of quantum computing, quantum communication, and quantum simulation. We highlight open challenges and future possibilities and conclude with some speculative visions for the next decade.

I. INTRODUCTION

This perspective surveys selected examples of classical machine learning and AI applied to quantum technologies, while stressing that methods should be matched to the problem. It introduces basic ML approaches and evolutionary algorithms as background for the applications discussed.

I. INTRODUCTION

  • The review tours selected examples at the classical-ML/quantum-technology interface, rather than claiming comprehensive literature coverage.It is aimed at quantum-technology physicists without prior machine-learning background.Source: §I · p. 1
  • The authors stress that the best-suited algorithm matters more than choosing the most advanced AI method for a task.They note that simple methods can outperform modern deep-learning methods when the latter are applied to the wrong task [8].Source: §I · p. 1

II. BASIC TECHNIQUES OF MACHINE LEARNING AND ARTIFICIAL INTELLIGENCE

  • Neural networks receive emphasis because their flexibility has made them popular general-purpose tools, though methods such as support vector machines and decision trees predate deep learning.Source: §II · p. 2

A. Evolutionary Algorithms

  • Evolutionary algorithms iteratively mutate or combine candidate solutions and select the best-performing candidates for the next generation.The review notes that these methods have been applied to quantum-technology tasks.Source: A. Evolutionary Algorithms · p. 2

B. Neural Networks: Structure

Neural networks transform inputs through connected neurons, using weighted sums, biases, and nonlinear activation functions. Different architectures exploit structure in the data, such as spatial invariance or temporal sequence.

B. Neural Networks: Structure

  • Each artificial neuron combines N inputs x_j with weights w_j, adds a bias b, and applies a nonlinear activation function to produce output y.Source: B. Neural Networks: Structure · p. 2
  • Without nonlinear activation functions between layers, a neural network collapses to one linear transformation with limited computational power.ReLU and sigmoid are examples of activation functions.Source: B. Neural Networks: Structure · pp. 2–3
  • Dense layers connect every output of one layer to every neuron in the next, while CNNs use learned kernels to exploit translational invariance in image inputs.Source: B. Neural Networks: Structure · p. 3
  • RNNs process time-series inputs sequentially while retaining information from previous time steps; LSTMs are a commonly used advanced RNN type.Source: B. Neural Networks: Structure · p. 3

C. Neural Networks: Training

Neural-network training fits an input-output mapping by minimizing a cost over examples. The section describes task-dependent losses and stochastic gradient updates, alongside backpropagation as an efficient way to calculate gradients.

C. Neural Networks: Training

  • Supervised learning trains a network approximation fθ(x) on input-target pairs by choosing parameters θ to minimize the sample-averaged cost.Source: C. Neural Networks: Training · p. 3
  • For regression, mean-square error is the canonical cost; for classification, cross-entropy compares probability distributions over the predefined classes.Source: C. Neural Networks: Training · p. 3
  • Stochastic gradient descent updates parameters using gradients estimated from randomly selected batches, with learning rate η controlling the update size.Source: C. Neural Networks: Training · p. 3
  • Backpropagation calculates gradients layer by layer from the output and is computationally no more demanding than evaluating the network forward.GPUs support efficient manipulation of the large matrices used in neural-network training.Source: C. Neural Networks: Training · p. 4

D. Unsupervised Learning

Unsupervised learning finds structure in unlabeled datasets, including compact representations, clusters, and new samples resembling the observed data.

D. Unsupervised Learning

  • Unsupervised learning discovers structure in unlabeled data, including feature learning and clustering.Source: D. Unsupervised Learning · p. 4
  • Feature learning seeks compact data representations, while clustering groups samples with similar properties without supplied class labels.Source: D. Unsupervised Learning · p. 4
  • Generative models stochastically create new samples following the distribution of previously observed data.Autoencoders can use supervised-learning techniques for self-supervised feature learning, whereas clustering and generative models typically use distinct techniques.Source: D. Unsupervised Learning · p. 4

E. Reinforcement Learning

Reinforcement learning (RL) seeks effective action strategies when the optimal strategy is unknown and supervised learning cannot be applied. The section introduces the Markov decision process and outlines policy-gradient, Q-learning, and actor-critic approaches.

E. Reinforcement Learning

  • An MDP pairs an agent with an environment: the agent observes states, chooses actions, receives rewards, and seeks to maximize cumulative reward over T time steps.The policy gives the probability of choosing an action given the observed state.Source: E. Reinforcement Learning · p. 4
  • Policy-gradient methods optimize a parameterized policy by estimating how its parameters affect average return, reinforcing actions common in high-reward trajectories.The gradient can be computed through the policy function without knowing the system model.Source: E. Reinforcement Learning · p. 4
  • Q-learning learns a Q function that estimates expected future return for a state and action, while actor-critic methods combine policy-gradient and Q-learning benefits.Actor-critic methods estimate a state-value function with a critic network and compare action outcomes against that value.Source: E. Reinforcement Learning · pp. 4–5

F. Automatic differentiation and gradient-based optimization

Automatic differentiation extends efficient gradient calculation beyond neural-network training. The section describes its use in continuous optimization and previews applications to quantum experiments and circuits.

F. Automatic differentiation and gradient-based optimization

  • Automatic differentiation obtains exact gradients with respect to variables in numerical calculations, distinct from symbolic differentiation in computer algebra.Modern neural-network frameworks offer multiple modes of automatic differentiation.Source: F. Automatic differentiation and… · p. 5
  • It can support gradient-based continuous optimization, particularly when many parameters make computational efficiency important.Source: F. Automatic differentiation and… · p. 5
  • The article previews using automatic differentiation to discover quantum experiments and circuits, among other applications.Source: F. Automatic differentiation and… · p. 5

G. How to get started

The authors point readers beginning with machine learning toward practical frameworks and an online collection of learning resources.

G. How to get started

A. Measurement data analysis and quantum state representation

Machine learning can interpret quantum measurements and represent or predict quantum states and dynamics. The examples range from device readout and state reconstruction to simulation, while future work includes improving data efficiency as devices scale.

A. Measurement data analysis and quantum state representation

  • Machine-learning classifiers improve quantum-device readout: an SVM outperformed classical clustering, and the analysis identified physical bitflips as the main noise contribution.Without such events, SVM classification becomes near-perfect; neural networks also improved trapped-ion readout, especially with more data per measurement.Source: A. Measurement data analysis and… · p. 6
  • A recurrent neural network can use noisy measurement traces to predict arbitrary measurement statistics during qubit evolution, effectively mapping measurement data to a quantum state.The network also receives the initial preparation state and final measurement basis as inputs.Source: A. Measurement data analysis and… · p. 6
  • PAC learning produced hypotheses close to the dataset with linear measurement scaling, confirmed in quantum-optics experiments with up to six photons despite realistic noise.This contrasts with tomography that seeks a hypothesis for every possible measurement and requires exponentially many measurements.Source: A. Measurement data analysis and… · p. 6
  • Joint measurements on copies of many-particle states can yield an exponential improvement over classical learning algorithms [36].The result was demonstrated with 40 superconducting qubits using a classical RNN, including on noisy, uncorrected quantum computers.Source: A. Measurement data analysis and… · p. 6
  • Because full-state storage grows exponentially with particle number, neural quantum states offer memory-efficient wave-function approximations, training a separate network for each state.The approach is described as a neural-network-based version of quantum state tomography.Source: A. Measurement data analysis and… · pp. 6–7
  • RBM-style tomography represents both probability and phase and matches experimental distributions across measurement bases; a Rydberg-atom study used the learned wave function to reconstruct unmeasured observables.The multi-basis approach evaluates transformed wave functions in a single reference basis.Source: A. Measurement data analysis and… · p. 7
  • Neural quantum states extend to mixed states and open systems, while other networks approximate many-body dynamics or predict observables without explicitly storing the wave function.The cited dynamics examples include dissipative evolution, driven systems, and neural ODEs.Source: A. Measurement data analysis and… · p. 7
  • Improving data efficiency for training and interpretation remains a challenge as quantum devices scale, with co-discovery of measurement and data-interpretation strategies identified as a future direction.Source: A. Measurement data analysis and… · p. 7

B. Parameter estimation: learning the properties of quantum systems

Machine learning approaches estimate quantum-system parameters through adaptive measurement choices, learned calibration maps, and models of system dynamics. The examples span quantum-device calibration, metrology, and Hamiltonian learning, with feedback speed and experimental design remaining important considerations.

Quantum metrology

  • Quantum metrology includes non-adaptive resource-efficient measurement and adaptive feedback; AI has contributed to both, while new experimental setups have also been explored.Adaptive approaches change inputs or measurement settings based on prior outcomes and require more advanced implementations.Source: B. Parameter estimation: learning the… · p. 8
  • A neural network trained on photonic calibration measurements can estimate unknown phase shifts from detected-photon counts.The calibration data were augmented to account for statistical noise.Source: B. Parameter estimation: learning the… · p. 8
  • Particle swarm optimization found slightly better Mach–Zehnder feedback strategies than greedy Bayesian BWB, but neither matched the optimal strategy found by exhaustive computation for small photon numbers.Other early machine-learning approaches used evolutionary algorithms to approximate the ideal feedback strategy [66, 67].Source: B. Parameter estimation: learning the… · p. 9
  • For qubit frequency estimation under dephasing, a neural-network reinforcement-learning agent using Bayesian updates and non-greedy measurement sequences performed very competitively.The agent received the current Bayesian parameter distribution and previous measurement choices, then was rewarded for reducing total parameter variance.Source: B. Parameter estimation: learning the… · p. 9

Device calibration

  • A constrained variational autoencoder reduced quantum-dot characterization measurements by a factor of 4 by sampling maps compatible with prior measurements and selecting informative voltage points.The naive 100-by-100 voltage scan would require 10,000 measurements; efficient cVAE sampling may scale to larger devices.Source: B. Parameter estimation: learning the… · pp. 9–10
  • A classifier can recognize charge transitions in multi-quantum-dot stability diagrams, enabling automatic tuning of the double-dot charge occupation.The diagrams are images, making convolutional neural networks a suitable classification approach.Source: B. Parameter estimation: learning the… · p. 10

Quantum Hamiltonian Learning

  • Quantum Hamiltonian Learning uses a quantum simulator with classical Bayesian estimation, while QMLA iteratively refines the Hamiltonian structure that QHL cannot adapt.An experimental QHL implementation estimated nitrogen-vacancy-center dynamics and indicated when its initial Hamiltonian model had deficits; QMLA was demonstrated in a hybrid quantum system.Source: B. Parameter estimation: learning the… · p. 10

Future Challenges and Opportunities

  • Adaptive approaches face a speed–sophistication trade-off because measurement-to-feedback time is often short, and future work could co-design the experiment and feedback strategy.The section also describes active learning for selecting informative device measurements.Source: B. Parameter estimation: learning the… · pp. 9–10

C. Discovering strategies for hardware-level quantum control

This section surveys reinforcement learning for hardware-level quantum control, from continuous-time pulse design to feedback strategies. It contrasts model-free black-box learning with model-based optimization and highlights practical challenges in moving from simulations to experiments.

C. Discovering strategies for hardware-level quantum control

  • RL is applied to continuous-time hardware control, including pulse design and feedback strategies based on noisy measurement traces.The section distinguishes this from higher-level protocols and experimental setups built from discrete elements.Source: C. Discovering strategies for… · p. 10
  • Earlier RL work spans discrete-pulse state preparation, deep-RL dynamical decoupling and state preparation, and unitary-gate fidelity optimization.Table-based Q-learning was used for restricted state and action spaces; the first deep-RL open-loop work used discrete and continuous controls with a recurrent agent.Source: C. Discovering strategies for… · p. 11
  • Deep RL was demonstrated on a real quantum computing platform for open-loop unitary-gate synthesis, with policy updates based on collected data and rewards.The lack of real-time device access was not a concern because the task required only open-loop control.Source: C. Discovering strategies for… · p. 11
  • For open-loop tasks, an RL agent may use simulated state information during training, then store the learned control sequence for later experiments without real-time control.Many model-free RL applications are numerical and rely on device simulations; state information from a simulation is unavailable in a real experiment.Source: C. Discovering strategies for… · pp. 11–12
  • Feedback expands the strategy space exponentially, and RL has been used for quantum error correction, state engineering, and control using measurement-induced nonlinearity.The authors also identify reliable reward extraction and genuinely black-box training as challenges for experimental model-free RL.Source: C. Discovering strategies for… · p. 11
  • Model-free RL can treat an experimental device as a black box without prior knowledge of its parameters, but training spends effort learning an implicit device model.When a calibrated model is available, model-based methods can optimize through its dynamics; automatic differentiation has been applied to quantum control tasks.Source: C. Discovering strategies for… · pp. 11–12
  • Feedback-GRAPE extends gradient-based model optimization to arbitrary strong discrete measurements by accounting for their stochastic effects [100].A prior automatic-differentiation example handled feedback based on weak linear measurements.Source: C. Discovering strategies for… · p. 12
  • Experimental training is difficult when measurements yield rewards only at run end; superconducting-qubit feedback may allow only a few hundred nanoseconds per network evaluation, and simulation-based transfer approaches require calibration.Immediate per-step rewards would improve training by crediting actions preceding high rewards.Source: C. Discovering strategies for… · p. 12

D. Discovering quantum experiments, protocols, and circuits

The section surveys AI methods for discovering quantum experiments, protocols, and circuits, from discrete building blocks to continuous parameter optimization. Examples include new experimental designs and strategies, alongside challenges such as computational cost, noise, and interpretability.

Discovery of Quantum Experiments

  • Automated experiment discovery produced blueprints for new quantum phenomena and ideas, while reinforcement learning rediscovered setups and automated their simplification.Melvin reuses discovered macro components; reported examples include a new approach to multiphoton interactions.Source: D. Discovering quantum experiments,… · p. 13
  • Theseus maps experiments to colored weighted graphs for gradient-based search and simplification, and its authors report it can outperform genetic algorithms or reinforcement learning when a simulator is available and feedback is unnecessary.Simplifying the graph topology can make a solution easier for human researchers to understand.Source: D. Discovering quantum experiments,… · p. 13, Figure 6 · p. 13
  • Tachikoma combines evolutionary search with neural-network classification to find quantum-metrology setups with previously unachieved advantage, though laboratory buildability and quality remain uncertain.The neural network classifies quantum states to address the costly computation of state fidelity.Source: D. Discovering quantum experiments,… · p. 13
  • Logic-AI SAT solving can quickly reject unsatisfiable candidates to guide continuous optimization, while variational-autoencoder latent spaces reveal discrete setup structures linked to experimental properties.The review describes both approaches as early or promising directions for investigating and understanding machine-learned representations.Source: D. Discovering quantum experiments,… · p. 14

Superconducting Circuits

  • SCILLA searches discrete superconducting-circuit topologies and then optimizes parameters; its 4-local coupler outperforms the other hand-crafted design in noise resilience and coupling strength.The search addresses a design space that grows exponentially with the number of circuit elements.Source: D. Discovering quantum experiments,… · p. 14

Discovering Quantum Protocols and Discrete Feedback Strategies

  • Reinforcement learning rediscovered quantum communication protocols by learning action sequences, finding action-efficient strategies with high probability over 100k trials.Tasks included teleportation, state purification, and entanglement swapping; direct gradient-based methods are also possible when a system model exists and feedback is absent.Source: D. Discovering quantum experiments,… · pp. 14–15

Quantum Circuits

  • Genetic algorithms rediscovered a teleportation circuit with fewer evaluations than exhaustive search, while hardware-aware circuit design outperformed textbook solutions in fidelity under realistic noise.Reinforcement learning has also been used for finite-gate synthesis, and circuit-simplification RL was reported superior to simulated annealing for circuits up to 50 qubits.Source: D. Discovering quantum experiments,… · pp. 15–16
  • Parameterized-circuit optimization faces exponentially vanishing gradients in randomly initialized circuits; proposed alternatives include reinforcement learning, predicted initial parameters, gradient-free methods, and quantum-computer-aided hardware design.The latter was experimentally demonstrated as a proof of principle in a three-qubit superconducting circuit; supervised simulation approximations also face a training-data versus computational-benefit trade-off.Source: D. Discovering quantum experiments,… · pp. 15–16

E. Quantum Error Correction

The section reviews machine learning for interpreting syndromes, searching codes, and discovering complete quantum-error-correction strategies. These approaches range from modifying existing codes to ab-initio discovery, with simulation scale and deployment speed remaining practical challenges.

Syndrome interpretation

  • Neural decoders learn to infer likely error locations from syndromes, including approaches trained on simulated syndrome–error pairs and reinforcement learning for faulty measurements.The review frames syndrome interpretation as supervised learning when the encoding and syndrome set are fixed.Source: E. Quantum Error Correction · p. 16

Code search

  • Reinforcement learning optimized surface-code connectivity for up to 70 data qubits, finding nontrivial designs for biased or spatially localized noise.The search modifies an existing code and uses efficient surface-code simulations to assess performance.Source: E. Quantum Error Correction · pp. 16–17

Autonomous quantum error correction

  • AutoQEC found a 3-code with a longer lifetime than previously studied systems under the same experimental constraints, then researchers derived and analyzed its logical state.The algorithm maximizes average logical-qubit fidelity while restricting its search to experimentally capable systems.Source: E. Quantum Error Correction · p. 17

Full QEC protocol discovery

  • Ab-initio reinforcement learning discovered error-correction and adaptive-measurement mitigation strategies for arbitrary small-system settings, but simulation effort has limited demonstrations to a handful of qubits.The method targets preservation of quantum information under a given platform and noise model; direct experimental learning remains a future challenge.Source: E. Quantum Error Correction · p. 17

Future challenges

  • Deployment requires syndrome-interpreting networks to run fast enough on devices, while ab-initio strategy discovery still faces exponentially scaling simulation costs.The section also points to co-discovering autonomous QEC experiments and feedback strategies.Source: E. Quantum Error Correction · p. 17

IV. OUTLOOK

Looking toward 2035, the authors envision machine learning helping manage the complexity of larger quantum systems through adaptive control, experiment discovery, and error correction. They also emphasize goal-driven discovery and the need to understand what machines learn.

Fully controlled and error-corrected quantum systems

  • For increasingly large quantum platforms, machine learning is envisioned as a toolbox for calibration, experiment discovery, extracting information from indirect measurements, control, and error correction.The authors anticipate this role for systems with thousands of qubits or quantum simulators with more degrees of freedom.Source: §IV · p. 17

Specifying goals, not algorithms

  • Rather than specifying algorithms, researchers could specify goals such as high operational fidelity, entanglement, or signal sensitivity and let computers discover how to achieve them.The authors suggest machines might eventually use scientific literature to propose goals and experiments autonomously.Source: §IV · pp. 17–18

Discovering new algorithms

  • The authors propose autonomous discovery of higher-level quantum programs, including quantum algorithms, as a future direction for machine learning.They note that large language models have recently been applied to classical algorithm discovery.Source: §IV · p. 18

How can the human learn?

  • Understanding what a machine has learned is presented as crucial for making its discoveries more widely applicable; discrete-action and logic-based approaches may be easier to interpret than deep-learning methods.The authors also identify symbolic regression as a possible route to extracting discrete explanations from neural-network predictions.Source: §IV · p. 18
Loading 2208.03836v1…