Toward Quantum-Aware Machine Learning: Improved Prediction of Quantum Dissipative Dynamics via Complex Valued Neural Networks

Muhammad Atif, Arif Ullah, Ming Yang · 2026

Read Toward Quantum-Aware Machine Learning: Improved Prediction of Quantum Dissipative Dynamics via Complex Valued Neural Networks by Muhammad Atif, Arif Ullah, Ming Yang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-03964.txt.

Introduction

Muhammad Atif, Arif Ullah, and Ming Yang argue that neural networks for quantum dynamics should themselves speak the language of complex numbers. Modeling how open quantum systems dissipate energy is hard because environments are complex and memory effects linger. Most learning models use real-valued networks, which split the density matrix into real and imaginary parts and can blur the amplitude-phase links that carry quantum coherence.

Their complex-valued networks operate directly on complex inputs, preserving the algebraic structure of quantum states. Benchmarks on the spin-boson model and several variants of a photosynthetic energy-transfer complex showed faster convergence, stabler training, and better physical fidelity than real-valued rivals, including cleaner conservation of trace and Hermiticity. The advantages grew with system size and coherence complexity, supporting complex networks as the natural classical tool for quantum simulation.

Catalog record

CREATORMuhammad Atif, Arif Ullah, Ming Yang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-03964

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Toward Quantum-Aware Machine Learning: Improved Prediction of Quantum Dissipative Dynamics via Complex Valued Neural Networks Muhammad Atif, Arif Ullah, Ming Yang Abstract Accurately modeling quantum dissipative dynamics remains challenging due to environmental complexity and non-Markovian memory effects. Although machine learning provides a promising alternative to conventional simulation techniques, most existing models employ real-valued neural networks (RVNNs) that inherently mismatch the complex-valued nature of quantum mechanics. By decoupling the real and imaginary parts of the density matrix, RVNNs can obscure essential amplitude–phase correlations, compromising physical consistency. Here, we introduce complex-valued neural networks (CVNNs) as a physics-consistent framework for learning quantum dissipative dynamics. CVNNs operate directly on complex-valued inputs, preserve the algebraic structure of quantum states, and naturally encode quantum coherences. Through numerical benchmarks on the spin-boson model and few variants of the Fenna–Matthews–Olson complex, we demonstrate that CVNNs outperform RVNNs in convergence speed, training stability, and physical fidelity—including significantly improved trace conservation and Hermiticity. These advantages increase with system size and coherence complexity, establishing CVNNs as a robust, scalable, quantum-aware classical approach for simulating open quantum systems in the pre-fault-tolerant quantum era. I Introduction Open quantum systems provide a fundamental framework for understanding how quantum states evolve under the influence of external environments (baths). Such system–environment interactions are central to numerous physical phenomena, including quantum information processing 8 , quantum memory 29 , quantum transport 16 , photosynthetic energy transfer 78 , and biochemical processes such as proton tunneling in DNA 55 . Conceptually, in the system-bath model, an open quantum system consists of a primary subsystem of interest coupled to an external bath. The dynamics of the subsystem are encoded in the reduced density matrix (RDM), which captures both intrinsic quantum evolution and the influence of environmental degrees of freedom. Modeling environmental effects, however, is highly challenging due to the vast and often continuous degrees of freedom associated with the bath. Existing approaches span two broad categories: mixed quantum–classical treatments and fully quantum formulations. Mixed quantum–classical schemes 45 ; 14 ; 34 ; 50 ; 51 ; 41 ; 42 ; 43 ; 60 ; 44 ; 15 ; 47 treat the system quantum mechanically while representing the environment through classical phase-space variables such as molecular vibrations or solvent motion, whose evolution follows Newtonian or Hamiltonian dynamics. This hybridization dramatically reduces computational cost by avoiding an explicit representation of the environment’s Hilbert space. Fully quantum approaches, in contrast, place both the system and the environment within the quantum Hilbert space, enabling a complete treatment of coherence, dissipation, and entanglement. Path-integral-based methods such as the hierarchical equations of motion (HEOM) 59 ; 38 ; 57 ; 74 ; 21 ; 73 ; 4 ; 71 and the quasiadiabatic propagator path integral (QUAPI) 38 ; 39 rigorously capture non-Markovian memory effects. Similarly, quantum master equation approaches–including the stochastic equation of motion (SEOM) 22 ; 23 ; 66 ; 10 ; 17 ; 56 and the generalized master equation 37 ; 35 —provide alternative fully quantum frameworks for describing environmental interactions. Despite their success, both categories face well-known limitations. Mixed quantum–classical approaches may break detailed balance 52 ; 3 ; 2 or fail to capture subtle quantum correlations 40 . Fully quantum methods, while theoretically rigorous, often become prohibitively expensive for strong coupling regimes, long memory times, or large environmental modes, where fine temporal resolution is required to ensure numerical stability. In response to these challenges, machine learning (ML) has recently emerged as a promising alternative for modeling complex spatiotemporal dynamics in quantum systems. 67 ; 63 ; 64 ; 62 ; 49 ; 25 ; 20 ; 84 ; 72 ; 32 ; 5 ; 75 ; 33 ; 58 ; 53 ; 31 ; 77 ; 36 ; 9 ; 82 ; 80 ; 26 ; 65 ; 81 ; 69 ; 79 ; 85 Existing work, which rely exclusively on real-valued neural networks (RVNNs), have demonstrated strong predictive capabilities in forecasting quantum dynamics from the past history of the system and in learning direct mappings between system parameters, time, and the corresponding quantum states. Despite these successes, a fundamental representational tension exists: quantum mechanics is intrinsically formulated within a complex-valued Hilbert space. This complex structure is not merely a mathematical convenience but encodes essential physical information, particularly in the phase relationships that define quantum coherence and interference. A parallel and compelling development across diverse scientific and engineering domains—from advanced signal processing and computational imaging to materials discovery and even emerging architectures for large language models 48 ; 11 ; 30 ; 70 ; 13 ; 27 ; 61 ; 7 ; 83 ; 18 ; 6 ; 76 —underscores that complex-valued neural networks (CVNNs) possess a unique, intrinsic capacity to capture phase-dependent structures and continuous rotational symmetries. These are features that RVNNs, by their architectural design, inherently struggle to represent. The standard practice in RVNNs is to decouple the real and imaginary components of complex-valued data, processing them through separate, parallel real-valued pathways. This bifurcation risks discarding or obscuring the critical correlations that are holistically encoded within the complex geometry of quantum states, potentially limiting physical fidelity and generalization. While quantum neural networks would, in principle, offer the most natural framework for quantum dynamics learning, current noisy intermediate-scale quantum (NISQ)-era devices remain too limited in scale and fidelity to outperform classical approaches. This motivates the search for classical architectures that can emulate essential quantum features without requiring quantum hardware. Given that complex numbers serve as a compact and faithful representation of quantum structure, 48 ; 11 ; 30 CVNNs emerge as a promising intermediary between RVNNs and fully quantum models. In this work, we conduct a systematic comparison between CVNNs and RVNNs for learning and forecasting quantum dissipative dynamics. Using the spin-boson (SB) model and several variants of the Fenna–Matthews–Olson (FMO) complex, we show that CVNNs converge faster, exhibit greater training stability, and yield predictions with improved physical fidelity—including superior trace conservation and Hermiticity—relative to RVNNs. Importantly, these advantages become increasingly pronounced with growing system size and coherence complexity, highlighting the scalability of the CVNN framework. Together, our results establish CVNNs as a powerful classical surrogate for quantum-aware learning and a practical pathway for modeling open quantum systems in the pre-fault-tolerant quantum era. II Theory and Methodology We consider an open quantum system \mathrm{S} with Hilbert space \mathcal{H}_{\mathrm{S}} of dimension n , interacting with an environment \mathrm{E} with Hilbert space \mathcal{H}_{\mathrm{E}} of dimension d . The joint system \mathrm{S}+\mathrm{E} evolves unitarily under the Liouville–von Neumann equation (\hbar=1) \dot{\rho}(t)=-\mathrm{i}[H,\rho(t)], (1) where H is the total Hamiltonian acting on the tensor product space \mathcal{H}_{\mathrm{S}}\otimes\mathcal{H}_{\mathrm{E}} , and \rho(t)\in\mathbb{C}^{(nd)\times(nd)} denotes the full density matrix of system plus environment. Given an initially factorized state \rho(0)=\rho_{\mathrm{S}}(0)\otimes\rho_{\mathrm{E}}(0), (2) the unitary propagator U(t)=e^{-iHt}\in\mathbb{C}^{(nd)\times(nd)} generates the reduced dynamics of the subsystem through the partial trace, \rho_{\mathrm{S}}(t)=\operatorname{Tr}_{\mathrm{E}}\big[U(t)\rho(0)U^{\dagger}(t)\big], (3) where \rho_{\mathrm{S}}(t)\in\mathbb{C}^{n\times n} represents the RDM of the system. The map \rho_{\mathrm{S}}(0)\mapsto\rho_{\mathrm{S}}(t) is typically non-Markovian and cannot, in general, be expressed as a closed local evolution equation. Instead, one may write a generalized master equation of the form \dot{\rho}_{\mathrm{S}}(t)=-\mathrm{i}[H_{\mathrm{S}},\rho_{\mathrm{S}}(t)]+\mathcal{R}\left[{\rho_{\mathrm{S}}(s)}_{0\leq s\leq t}\right], (4) where H_{\mathrm{S}} acts on the system alone, and \mathcal{R} is a superoperator encoding dissipation, decoherence, and memory. Abstractly, the reduced dynamics can therefore be represented as a history-dependent quantum map

Read the complete text