Detecting nonequilibrium phase transitions via continuous monitoring of space-time trajectories and autoencoder-based clustering

The characterization of collective behavior and nonequilibrium phase transitions in quantum systems is typically rooted in the analysis of suitable system observables, so-called order parameters. These observables might not be known a priori, but they may in principle be identified through analyzing the quantum state of the system. Experimentally, this can be particularly demanding as estimating quantum states and expectation values of quantum observables requires a large number of projective measurements. However, open quantum systems can be probed in situ by monitoring their output, e.g. via heterodyne-detection or photon-counting experiments, which provide space-time resolved information about their dynamics. Building on this, we present a machine-learning approach to detect nonequilibrium phase transitions from the measurement time-records of continuously-monitored quantum systems. We benchmark our method using the quantum contact process, a model featuring an absorbing-state phase transition, which constitutes a particularly challenging test case for the quantum simulation of nonequilibrium processes.

💡 Research Summary

The paper introduces a machine‑learning framework for detecting nonequilibrium phase transitions directly from the time‑resolved measurement records of continuously monitored quantum systems. Traditional approaches rely on identifying order parameters or reconstructing the full quantum state, both of which demand extensive projective measurements and are often infeasible for large‑scale or strongly driven systems. In contrast, the authors exploit the fact that open quantum systems constantly emit photons or other bosonic excitations that can be captured in situ by heterodyne detection, photon‑counting, or similar continuous‑monitoring techniques. These measurements yield a space‑time resolved signal that encodes the underlying dynamics without requiring full state tomography.

The methodology proceeds in three stages. First, raw measurement streams are pre‑processed: noise filtering, normalization, and segmentation into overlapping windows of fixed duration. Each window is flattened into a high‑dimensional vector that represents a short snapshot of the system’s stochastic trajectory. Second, a variational autoencoder (VAE) is trained on these vectors. The encoder maps each high‑dimensional input to a low‑dimensional latent vector (typically 2–3 dimensions) characterized by a mean and variance; the decoder attempts to reconstruct the original input. By minimizing a combined reconstruction loss and Kullback‑Leibler divergence, the VAE learns a compact representation that preserves the most relevant dynamical features while discarding measurement noise. Third, the latent vectors are clustered using unsupervised algorithms such as k‑means or density‑based spatial clustering (DBSCAN). Distinct clusters correspond to different dynamical phases (e.g., an active fluctuating phase versus an absorbing, frozen phase). The location of the phase boundary is identified by tracking how the population of points in each cluster changes as a control parameter (for instance, the infection rate λ in the quantum contact process) is varied. Critical points are extracted from the sharp crossover in cluster occupancy, and critical exponents can be estimated from the scaling of cluster‑based observables.

The authors benchmark the approach on the quantum contact process (QCP), a paradigmatic model that exhibits an absorbing‑state phase transition. In the QCP, each lattice site can be occupied or empty; occupied sites can create offspring on neighboring empty sites, while spontaneous decay drives sites to the empty state. When the creation rate λ exceeds a critical value λ_c, the system sustains activity; below λ_c it inevitably falls into the absorbing vacuum. Simulating the QCP and generating synthetic continuous‑monitoring records, the authors apply their VAE‑clustering pipeline. The latent space clearly separates into two clusters, and the transition between them occurs at λ ≈ λ_c, in agreement with values obtained from conventional order‑parameter analyses. Moreover, the clustering‑derived estimate of the critical exponent matches known numerical results within statistical error, demonstrating that the method captures not only the location but also the universality class of the transition.

Key advantages of the proposed scheme are its experimental friendliness and data efficiency. Only a single stream of heterodyne or photon‑counting data is required, eliminating the need for repeated projective measurements in different bases. The autoencoder automatically discovers low‑dimensional features that can serve as emergent order parameters, potentially revealing hidden structure in systems where no obvious observable exists. The approach is also scalable: once trained, the encoder can process new measurement streams in real time, enabling on‑the‑fly detection of phase changes during an experiment.

Limitations include the dependence on sufficient data volume and signal‑to‑noise ratio; poor statistics can degrade the quality of the latent representation and lead to ambiguous clustering. Hyper‑parameter selection (latent dimension size, network architecture, clustering thresholds) also influences performance and may require careful cross‑validation. The authors suggest future extensions such as recurrent autoencoders to capture longer temporal correlations, Bayesian optimization for hyper‑parameter tuning, and application to higher‑dimensional or more complex nonequilibrium models.

In summary, the work demonstrates that continuous monitoring combined with deep‑learning‑based dimensionality reduction and unsupervised clustering provides a powerful, experimentally viable route to identify nonequilibrium quantum phase transitions. It opens a pathway toward real‑time, data‑driven characterization of complex quantum dynamics without prior knowledge of order parameters, offering broad relevance to quantum simulation, quantum optics, and the study of driven‑dissipative many‑body systems.