🔬 condensed matter

A machine-learned dynamical phase diagram of the one-dimensional nonlinear Schrödinger equation with quasiperiodic disorder and a static field

This paper employs a machine-learning framework combining short-time simulations, feature reduction, and Gaussian mixture clustering to efficiently map the dynamical phase diagram of the one-dimensional nonlinear Schrödinger equation with quasiperiodic disorder and a static field, successfully identifying and validating five distinct transport regimes and their dependence on interaction strength, field, and disorder while revealing stark contrasts between delta and Gaussian initial states.

Original authors: Marcos Pérez

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Marcos Pérez

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

In the microscopic world of quantum physics, particles do not always behave like tiny, solid marbles rolling down a hill. Instead, they act more like waves, spreading out and interfering with one another as they move through a material. Scientists have long known that if a material is perfectly ordered, these waves can travel freely and quickly, a state called ballistic transport. However, if the material is messy or disordered, the waves can get stuck in place, a phenomenon known as localization. The question becomes far more complex when the material is neither perfectly ordered nor randomly messy, but instead follows a specific, repeating pattern that never quite repeats itself, known as quasiperiodic disorder. Furthermore, when these particles interact with each other, they can change the rules entirely, sometimes breaking free from the disorder to spread slowly, or in other cases, locking themselves into a tight, unmoving cluster. Understanding how these different forces—disorder, interaction, and external fields—compete to either trap or release a wave is crucial for designing future quantum technologies, yet mapping these outcomes has been a daunting task because the simulations required to see the final result take an impractical amount of time.

A researcher led by Marcos Pérez at the Federal University of Rio Grande do Sul in Brazil has tackled this challenge by using machine learning to create a new kind of map. Instead of running a single, incredibly long simulation for every possible combination of conditions, which would take years of computer time, they ran thousands of very short simulations. They then fed the results of these brief runs into a computer program designed to find patterns and group similar behaviors together. This approach allowed them to quickly identify the boundaries between different transport regimes, effectively creating a dynamical phase diagram that shows how a wave packet will behave under various conditions. The study focused on a one-dimensional chain of particles subject to a quasiperiodic potential, a static electric field, and self-interaction, testing two different starting shapes for the wave: a tight, single-point burst and a broad, smooth hill.

The researcher discovered that the starting shape of the wave matters profoundly. When the wave began as a tight, single-point burst, a large region of the map revealed a phenomenon called self-trapping. In this state, the strong interaction between particles causes the wave to collapse into a tight knot that refuses to spread, even though the disorder and the external field might otherwise allow it to move. This self-trapping region grows larger as the interaction strength increases, invading areas where the wave would normally travel freely or remain localized. In stark contrast, when the wave started as a broad, smooth hill, this self-trapping behavior vanished entirely. Instead, the broad wave followed a different path, entering a slow, subdiffusive spreading regime along the critical line where the disorder is strongest. This difference highlights that the initial state of a quantum system is not a minor detail but a fundamental factor that can change the entire topology of how the system evolves.

To ensure their machine-learned map was accurate, the researcher validated their findings against a smaller set of much longer, more expensive simulations that served as a ground truth. They confirmed that the short simulations, when analyzed through their clustering method, correctly predicted the long-term behavior of the system. The map successfully identified five distinct regimes: ballistic motion where the wave flies freely, localization where it stays put, oscillatory localization where it vibrates in place, slow subdiffusive spreading, and the self-trapped state. For the self-trapped state, the researcher had to look beyond simple spreading measurements, using a separate metric to confirm that a large portion of the wave's energy remained trapped at the center, a detail that standard spreading measurements alone would have missed.

The study also revealed how the external static field interacts with the system. When the field is present, it tends to localize the wave, but the interaction between particles can eventually overcome this, causing the wave to leak out slowly. The researcher found that the boundary between these behaviors shifts depending on the strength of the field and the interaction, creating a complex landscape of possibilities. By using a method that combines short simulations with intelligent pattern recognition, the researcher was able to produce a full phase diagram in a fraction of the time it would have taken using traditional methods. This work demonstrates that machine learning can be a powerful tool for exploring complex physical systems, provided that the results are carefully checked against rigorous, long-term simulations to ensure the patterns found are real and not just artifacts of the short observation window. The resulting map offers a clear, quantitative view of how disorder, interaction, and external forces compete to determine the fate of a quantum wave, providing a solid foundation for future experiments and theoretical work in this field.

Technical Summary: Machine-Learned Dynamical Phase Diagram of the 1D Nonlinear Schrödinger Equation

Problem Statement
The paper addresses the challenge of mapping the transport regimes of a wave packet evolving under the one-dimensional discrete nonlinear Schrödinger (Gross–Pitaevskii) equation with three competing control parameters: self-interaction strength ( $g$ ), static field strength ( $F$ ), and quasiperiodic disorder strength ( $\lambda$ ). The system combines the Aubry–André model (quasiperiodic disorder), Stark localization (static field), and mean-field nonlinearity. Distinguishing between transport regimes—such as subdiffusion ( $\alpha \approx 1/3$ ), strong chaos ( $\alpha \approx 1/2$ ), and true localization—typically requires integrating the equations over many decades in time ( $t \gtrsim 10^6$ ), making a full parameter-space survey computationally prohibitive. Furthermore, the transport phenomenology depends critically on the initial state, with compact (delta) and broad (Gaussian) wave packets exhibiting distinct behaviors, particularly regarding self-trapping.

Methodology
The author proposes a machine-learning pipeline that replaces expensive long-time simulations with a high-volume, short-time approach:

Sampling: The author performs 3,600 short simulations ( $t_{\max} = 2.5 \times 10^3$ $t_{m a x} = 2.5 \times 1 0^{3}$ ) across two 15 $\times$ $\times$ 15 parameter slices:
- Slice A: Interaction ( $g$ ) vs. Quasiperiodic disorder ( $\lambda$ ) at $F=0$ .
- Slice B: Interaction ( $g$ ) vs. Static field ( $F$ ) at $\lambda=1$ .
- Simulations are run for two initial states (single-site delta and broad Gaussian) and averaged over four disorder-phase realizations.
Feature Extraction: Each short trajectory is reduced to an eight-dimensional vector of physically motivated dynamical features, including:
- Spreading exponents ( $\alpha$ ) for the second moment ( $m_2$ ) and participation number ( $P$ ).
- Final values of $\log_{10} m_2$ and $\log_{10} P$ .
- Peak-drop (fraction of initial central norm lost).
- Bloch score (normalized center-of-mass excursion).
- Growth ratios and final center-of-mass position.
Unsupervised Clustering: A Gaussian Mixture Model (GMM) is fitted to the standardized feature space. The number of components ( $k=6$ ) is selected via the Bayesian Information Criterion (BIC). This step partitions the parameter space into clusters without supervision.
Semi-Supervised Labeling: While the boundaries are learned, the physical names of the regimes (ballistic, localized, oscillatory-localized, subdiffusive, self-trapped) are assigned via a hand-calibrated threshold rule applied to the cluster members by majority vote. The author explicitly distinguishes this from "discovering" phases, noting the taxonomy is pre-defined.
Validation Strategy:
- Lattice Scaling: To ensure valid comparison, the long-time validation runs use a lattice size ( $N=16,384$ ) eight times larger than the training lattice ( $N=2,048$ ). This prevents ballistic packets from hitting the periodic boundary during the validation window, a failure mode that would occur if lattice sizes were matched without adjustment.
- Ground Truth: 100 independent long-time runs ( $t_{\max} = 10^4$ ) are performed on a subgrid to validate the short-time clusters.
- Dual Observables: Validation uses both the asymptotic spreading exponent ( $\alpha_\infty$ ) and the long-time residual peak norm ( $\Pi_0$ ), the latter being crucial for diagnosing self-trapping where $\alpha_\infty$ is ambiguous.

Key Results

Phase Diagrams: The pipeline successfully generates 15 $\times$ $\times$ 15 phase maps for both initial states.
- Gaussian State: Shows a "clean" transition from ballistic (metal) to localized (insulator) regimes, with a subdiffusive corridor pinned to the Aubry–André critical line ( $\lambda \approx 2$ ) and an oscillatory-localized band. No self-trapping is observed.
- Delta State: Exhibits a broad "self-trapping wedge" that invades both ballistic and localized regions as $g$ increases. The onset of self-trapping at $\lambda=0$ is precisely located at $g = 3.907 \pm 0.017$ .
Validation Metrics:
- The short-time clusters accurately predict long-time asymptotic exponents: Ballistic ( $\alpha_\infty \approx 2.01$ ), Localized ( $\approx 0.12$ ), Oscillatory-localized ( $\approx 0.23$ ), and Subdiffusive ( $\approx 0.28$ , consistent with weak chaos $\alpha=1/3$ ).
- For the Self-trapped regime, the spreading exponent is non-diagnostic ( $\alpha_\infty \approx 1.47$ ) because it measures the radiated tail. However, the independent observable $\Pi_0$ confirms self-trapping with a high retention value ( $0.72 \pm 0.32$ ), distinct from delocalized phases.
Clustering Efficacy: The clustering step significantly smooths the phase boundaries compared to applying the threshold rules pointwise, reducing spatial "roughness" (neighbor disagreement) by factors of 2 to 4, particularly in the field-driven slice.
Stability: Bootstrap resampling indicates cluster label stability of 75.5% (delta) and 84.2% (Gaussian), quantifying the uncertainty in boundary locations.

Significance and Claims
The paper claims to provide a computationally efficient method (approx. 13 minutes per slice-state on a laptop) to map complex dynamical phase diagrams that would otherwise be inaccessible due to the cost of direct asymptotic simulation ( $t \gtrsim 10^6$ ).

Key contributions include:

Validation of Short-Time Proxies: Demonstrating that short-time dynamical features, when clustered, can reliably predict long-time asymptotic behavior across diverse regimes.
Initial State Dependence: Highlighting that the choice of initial state (delta vs. Gaussian) fundamentally alters the topology of the phase diagram, specifically the existence of the self-trapping wedge, which has implications for experimental protocols involving different quench conditions.
Methodological Rigor: The author is explicit about the limitations of their approach. They clarify that the names of the phases are not discovered by the machine learning but are assigned via human-defined rules; the ML only learns the partitioning of the parameter space. They also explicitly report "crossover cells" where the rule-based labeling disagrees with the cluster majority, rather than hiding these uncertainties.
Robust Validation Design: The use of a significantly larger lattice for validation ensures that the comparison between short-time training and long-time ground truth is physically meaningful, avoiding artifacts caused by boundary effects in the fastest-spreading regimes.

The work concludes that while the method has limits (e.g., confidence decreases near boundaries and for the ballistic regime's very long-time behavior), it offers a practical, uncertainty-quantified framework for exploring dynamical phase diagrams in driven, disordered, nonlinear systems.

Technical Summary: Machine-Learned Dynamical Phase Diagram of the 1D Nonlinear Schrödinger Equation

More like this