HPC-Driven Modeling with ML-Based Surrogates for Magnon-Photon Dynamics in Hybrid Quantum Systems

TL;DR

The work tackles the multiscale challenge of simulating magnon–photon dynamics in hybrid quantum circuits by integrating a massively parallel GPU Maxwell–LLG solver with a physics-informed ML surrogate. The authors demonstrate an HPC-enabled 3D framework (ARTEMIS) that captures fully coupled electromagnetic and magnetization dynamics while enabling rapid long-horizon predictions through a Long Expressive Memory surrogate trained with physics losses and curriculum learning. They show clear strong-coupling signatures, including anti-crossing and a measured g_mp ≈ 670 MHz, and report up to 5x speedups over purely numerical simulations while preserving key spectral features. The approach is scalable to thousands of GPUs, with robust generalization across probe points, and is poised to accelerate design and prototyping of on-chip magnon–photon devices and broader hybrid quantum platforms. This hybrid HPC-ML paradigm offers a pathway to efficient, accurate multi-physics modeling in complex quantum systems, supporting rapid exploration and optimization of next-generation quantum-spintronic architectures.

Abstract

Simulating hybrid magnonic quantum systems remains a challenge due to the large disparity between the timescales of the two systems. We present a massively parallel GPU-based simulation framework that enables fully coupled, large-scale modeling of on-chip magnon-photon circuits. Our approach resolves the dynamic interaction between ferromagnetic and electromagnetic fields with high spatiotemporal fidelity. To accelerate design workflows, we develop a physics-informed machine learning surrogate trained on the simulation data, reducing computational cost while maintaining accuracy. This combined approach reveals real-time energy exchange dynamics and reproduces key phenomena such as anti-crossing behavior and the suppression of ferromagnetic resonance under strong electromagnetic fields. By addressing the multiscale and multiphysics challenges in magnon-photon modeling, our framework enables scalable simulation and rapid prototyping of next-generation quantum and spintronic devices.

Figures (9)

• Figure 1: Magnon-photon hybrid quantum circuit modeling setup and simulated field profiles. (a) Perspective view of the magnon-photon coupling circuit, comprising a CPW resonator and a ferromagnetic thin film. The CPW resonator consists of a center conductor and lateral ground conductors (yellow) positioned on a silicon substrate (blue). The ferromagnet, shown as a red stripe, is placed on top of the center conductor. Material and geometrical parameters are detailed in the main text. (b)–(d) Cross-sectional views of the circuit, overlaid with the simulated magnetic field (H) distribution at different locations along the longitudinal direction of the CPW resonator, as indicated in (a). (e)–(g) Top views of the circuit, overlaid with the simulated electric and magnetic field distributions in the z-plane at different observation points, as marked in (a). (h) Theoretical field distributions along the longitudinal direction of the circuit, where the electric field $E_x$ is concentrated in the lateral air gap, and the magnetic field $H_x$ is observed above and beneath the center conductor.
• Figure 1: The theoretical analysis of the frequencies and linewidths of the hybrid polaritons agrees well with the numerical simulations. The right-hand figures show the color map of the simulated electric field (top) and the theoretically calculated electric field spectrum (bottom). In the left figure, the theoretical results are plotted as contour lines (red dashed) for clearer visualization. See Supplementary Information for details.
• Figure 1: 1D hybrid magnon-photon cavity.
• Figure 2: The HPC and ML frameworks for magnon-photon hybrid quantum modeling. (a) Simulation domain setup and GPU partitioning in the Maxwell–LLG coupled numerical model, ARTEMIS. The GPU parallelization leverages the grid-structured GPU library AMReX AMReX_JOSS, achieving nearly flat weak-scaling performance. (b) Top view of the magnon-photon hybrid quantum circuit; the cross marks indicate probing points used for data collection. (c) Example time-series data of the magnetization M recorded at a single probing point. (d) Detailed architecture of the LEM cell. The cell updates a latent state ($C_t$) and a hidden state ($H_t$) using the input $X_t$, previous states, and learnable time constants ($\Delta t$). $C_t$ captures short-term dynamics, while $H_t$ integrates transformed latent features to maintain temporal context. (e) The curriculum learning strategy used to train the ML model, with a staged schedule with progressively adjusted learning rates, batch sizes, sequence lengths, and physics-based loss weights to stabilize training and enhance convergence. (f) Overall ML framework for time-series prediction, built out of the LEM-based architecture. The encoder consists of sequential LEM Cells that process the input sequence (X) into two evolving internal states: the latent state (C); and the hidden state (H). These states are used to initialize an auto-regressive decoder, which also consists of stacked LEM Cells. At each step, the decoder outputs a hidden state (H) that is passed through a shared fully connected (FC) layer to produce the predicted output.
• Figure 2: Electric-field spectra as the excitation intensity is increased. Stronger microwave magnetic fields suppress the magnon dynamics and leave only the cavity photon mode. Such behaviour arises because, at high powers, nonlinear spin-wave interactions -- particularly three-magnon splitting -- induce a Suhl instability that transfers energy from the uniform ferromagnetic resonance to non-uniform spin waves; this saturation of the Kittel mode causes the magnon–photon anticrossing gap to close and yields a single photon-like resonance, a phenomenon sometimes described as the high-power cavity magnon-polaritons regime nonlinear2023. See Supplementary Information for details.