We investigate the potential of supervised machine learning to propagate a quantum system in time. While Markovian dynamics can be learned easily, given a sufficient amount of data, non-Markovian systems are nontrivial and their description requires the memory knowledge of past states. Here we analyze the feature of such memory by taking a simple 1D Heisenberg model as many-body Hamiltonian, and construct a non-Markovian description by representing the system over the single-particle reduced density matrix. The number of past states required for this representation to reproduce the time-dependent dynamics is found to grow exponentially with the number of spins and with the density of the system spectrum. Most importantly, we demonstrate that neural networks can work as time propagators at any time in the future and that they can be concatenated in time forming an autoregression. Such neural-network autoregression can be used to generate long-time and arbitrary dense time trajectories. Finally, we investigate the time resolution needed to represent the system memory. We find two regimes: For fine memory samplings the memory needed remains constant, while longer memories are required for coarse samplings, although the total number of time steps remains constant. The boundary between these two regimes is set by the period corresponding to the highest frequency in the system spectrum, demonstrating that neural network can overcome the limitation set by the Shannon-Nyquist sampling theorem.
We investigate the potential of supervised machine learning to propagate a quantum system in time. While Markovian dynamics can be learned easily, given a sufficient amount of data, non-Markovian systems are nontrivial and their description requires the memory knowledge of past states. Here we analyze the feature of such memory by taking a simple 1D Heisenberg model as many-body Hamiltonian, and construct a non-Markovian description by representing the system over the single-particle reduced density matrix. The number of past states required for this representation to reproduce the time-dependent dynamics is found to grow exponentially with the number of spins and with the density of the system spectrum. Most importantly, we demonstrate that neural networks can work as time propagators at any time in the future and that they can be concatenated in time forming an autoregression. Such neural-network autoregression can be used to generate long-time and arbitrary dense time trajectories. Finally, we investigate the time resolution needed to represent the system memory. We find two regimes: For fine memory samplings the memory needed remains constant, while longer memories are required for coarse samplings, although the total number of time steps remains constant. The boundary between these two regimes is set by the period corresponding to the highest frequency in the system spectrum, demonstrating that neural network can overcome the limitation set by the Shannon-Nyquist sampling theorem.
This paper investigates the potential of quantum neural networks (QNNs) enhanced by non-Markovian dynamics. Traditional Markovian QNNs suffer from limitations in capturing long-range dependencies, a crucial aspect in complex data processing. We propose a novel QNN architecture that incorporates quantum memory and entanglement to facilitate non-Markovian information processing. This approach allows the network to retain and utilize the entire past history of the input, fundamentally altering the dynamics of information flow. Through theoretical analysis and mathematical formulations, we demonstrate how this architecture can significantly improve the network's ability to learn and represent complex patterns, particularly those exhibiting long-range correlations. Specifically, we introduce a modified update rule leveraging the concept of quantum memory and entanglement, expressed as: $ \Delta w_{ij} = \alpha \cdot E[\rho(t)|x_1, x_2, ..., x_K] \cdot \langle \psi_i | \rho(t) | \psi_j \rangle - (1 - \alpha) \cdot w_{ij}$ Where: * $ \Delta w_{ij}$ is the change in the synaptic weight between neurons *i* and *j*. * α is the learning rate. * E[ρ(t)∣x1,x2,...,xK] represents the expected state of the quantum memory at time *t*, given the entire input sequence x1,x2,...,xK. * ρ(t) is the density matrix representing the quantum state at time *t*. * ⟨ψi∣ρ(t)∣ψj⟩ is the inner product between the quantum states ψi and ψj within the density matrix ρ(t). * wij is the current synaptic weight between neurons *i* and *j*. The core of our innovation lies in the utilization of quantum memory, modeled as a quantum state ρ(t), which retains information about the entire past sequence. The entanglement between neurons further enables correlated updates, bypassing the limitations of Markovian models that only consider immediate dependencies. We explore the impact of varying the strength of entanglement and the capacity of the quantum memory on the network's performance using mathematical models. The results suggest a significant improvement in accuracy and convergence speed compared to standard Markovian QNNs when dealing with datasets exhibiting long-range dependencies. Further investigation into the scalability and practical implementation of this architecture is warranted.
Featuring memory of past inputs is a fundamental requirement for machine learning models processing time-dependent data. In quantum reservoir computing, all architectures proposed so far rely on Markovian dynamics, which, as we prove, inherently lead to an exponential decay of past information, thereby limiting long-term memory capabilities. We demonstrate that non-Markovian dynamics can overcome this limitation, enabling extended memory retention. By analytically deriving memory bounds and supporting our findings with numerical simulations, we show that non-Markovian reservoirs can outperform their Markovian counterparts, particularly in tasks that require a coexistence of short- and long-term correlations. We introduce an embedding approach that allows a controlled transition from Markovian to non-Markovian evolution, providing a path for practical implementations. Our results establish quantum non-Markovianity as a key resource for enhancing memory in quantum machine learning architectures, with broad implications in quantum neural networks.
Conventional approaches to simulating quantum many-body dynamics produce a single trajectory: if the Hamiltonian or the initial state is changed, the computation must be re-performed. Recent efforts toward foundation models have begun to address this limitation, yet existing methods transfer across either Hamiltonians or initial states, but not both. In this work, we introduce the Universal Neural Propagator (UNP), a single, unified model that learns the functional mapping from driving protocols to time-evolution propagators. Trained in an entirely self-supervised way, a single UNP model predicts dynamics across a function space of driving protocols and an exponentially large Hilbert space of initial states simultaneously. We benchmark on a two-dimensional driven Ising model and demonstrate the UNP's accuracy and transferability across product and entangled initial states, as well as for both in- and out-of-distribution driving protocols. The UNP remains accurate at system sizes beyond exact diagonalization, and can be efficiently fine-tuned across all initial states using observable data. By shifting the object of learning from quantum states to operators, this work opens a route toward transferable simulation of driven quantum matter.
We present a \emph{novel} and scalable supervised machine learning framework to predict open-quantum system dynamics and detect non-Markovian memory using only local ancilla measurements. A system qubit is coherently coupled to an ancilla via a symmetric XY Hamiltonian; the ancilla interacts with a noisy environment and is the only qubit we measure. A feedforward neural network, trained on short sliding windows of supplementary data from the past, forecasts the observable system ⟨Z(S)(t)⟩ without state tomography or knowledge of the bath. To quantify memory, we introduce a normalized revival-based metric that counts upward 'turn-backs' in \emph{predicted} ⟨Z(S)(t)⟩ and reports the fraction of evaluated samples that exceeds a small threshold. This bounded score provides an interpretable, model-independent indicator of non-Markovianity. We demonstrate the method on two representative noise channels, non-unital amplitude damping and unital dephasing from random telegraph noise (RTN). Under matched conditions, the model accurately reproduces the dynamics and flags memory effects, with RTN exhibiting a larger normalized revival score than amplitude damping. Overall, the approach is experimentally realistic and readily extensible, enabling real-time, interpretable non-Markovian diagnostics from accessible local measurements.
Machine learning (ML) architectures such as convolutional neural networks (CNNs) have garnered considerable recent attention in the study of quantum many-body systems. However, advanced ML approaches such as gated recurrent neural networks (RNNs) have seldom been applied to such contexts. Here we demonstrate that a special class of RNNs known as long short-term memory (LSTM) networks is capable of learning and accurately predicting the time evolution of one-dimensional (1D) Ising model with simultaneous transverse and parallel magnetic fields, as quantitatively corroborated by relative entropy measurements and magnetization between the predicted and exact state distributions. In this unsupervised learning task, the many-body state evolution was predicted in an autoregressive way from an initial state, without any guidance or knowledge of any Hamiltonian. Our work paves the way for future applications of advanced ML methods in quantum many-body dynamics without relying on the explicit form of the Hamiltonian.
Understanding the long-time dynamics of strongly correlated quantum matter remains a major challenge: as quantum entanglement grows during time evolution, conventional numerical methods become increasingly costly, limiting access to the long-range correlations and low-energy excitations that govern gapless systems. Here we show that these dynamical correlations can be extended beyond the accessible numerical window without explicitly evolving the underlying many-body wavefunction. We introduce an autoregressive machine-learning framework that enables the extrapolation of dynamical spin correlations in both time and space beyond the reach of conventional numerical methods. Trained on time-dependent density matrix renormalization group simulations of the gapless XXZ model, our approach is benchmarked against exact solutions available for this analytically solvable system. Combined with physics-informed spatial extension, a multi-layer perceptron model using ReLU activation functions has been shown to be superior to convolutional neural networks and linear regressions for longer time extrapolation. A study of error accumulation further demonstrates that our autoregressive neural network extrapolations are highly robust to perturbations, suggesting stable and reliable predictions. This work offers a complementary route to studying the dynamics of gapless quantum many-body systems, in which machine learning extends and complements the capabilities of state-of-the-art numerical approaches.