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

Jialin Song, Yingheng Tang, Pu Ren, Shintaro Takayoshi, Saurabh Sawant, Yujie Zhu, Jia-Mian Hu, Andy Nonaka, Michael W. Mahoney, N. Benjamin Erichson, Zhi Yao

Abstract

Simulating the interaction between microwave photons and magnons in hybrid quantum systems is computationally intensive due to the multi-scale nature of the underlying physics. This paper introduces a hybrid framework that integrates a massively parallel GPU-based solver (ARTEMIS) with a physics-informed machine learning surrogate (LEM). By training on high-fidelity data generated by ARTEMIS, the LEM surrogate achieves a 5x speedup in predicting long-term system dynamics while maintaining a resonance-frequency error below 2.3%. This approach enables efficient modeling of complex nonlinear phenomena and scales effectively to thousands of GPUs, paving the way for rapid prototyping of next-generation quantum and spintronic devices.

Introduction: The Challenge of Hybrid Quantum Systems

Hybrid quantum systems aim to leverage the complementary strengths of different physical platforms, such as superconducting circuits for processing and magnetic materials for storage or sensing. A prominent example is cavity magnonics, where microwave photons in a resonator couple with magnons—collective excitations of electron spins—in a ferromagnetic material.

Schematic of the hybrid quantum system: a coplanar waveguide resonator coupled with a ferromagnetic (FM) strip.

Accurately modeling these systems requires solving the coupled Maxwell equations for electromagnetic fields and the Landau–Lifshitz–Gilbert (LLG) equation for magnetization dynamics. The computational cost of high-precision 3D simulations is prohibitive for iterative design, often forcing researchers to rely on simplified models that fail to capture critical nonlinear effects.

Methods: A Hybrid HPC-ML Framework

To overcome these limitations, the researchers developed a two-stage approach combining high-performance computing (HPC) with machine learning (ML).

The ARTEMIS Solver

The foundation of the framework is ARTEMIS, a massively parallel solver built on the AMReX library. It employs an explicit Finite-Difference Time-Domain (FDTD) scheme to solve the coupled Maxwell-LLG equations. Optimized for GPU architectures using MPI+CUDA, ARTEMIS demonstrates near-perfect weak scaling up to 2,048 GPUs on the Perlmutter supercomputer.

The LEM Surrogate

To accelerate the design cycle, the team trained a Long Expressive Memory (LEM) neural network as a surrogate model. Unlike standard RNNs, LEM is specifically designed to handle the multi-scale differential equations characteristic of physical systems.

Comprehensive overview of the modeling framework, showing (a) the device schematic, (b-d) cross-sectional field distributions, (e-g) spatial field distributions, and (h) field intensity profiles.

The training process incorporates several key innovations:

  • Physics-Informed Loss: The model's loss function includes a term derived from the LLG equation, ensuring that predictions adhere to physical laws like the conservation of magnetization.
  • Curriculum Learning: Training proceeds in stages, starting with simple, short sequences and progressing to complex, long-term dynamics.
  • Extrapolation Capability: The surrogate is trained to predict the full timeline of a simulation using only the first 20% of the high-fidelity data as input.

Key Results and Insights

The hybrid framework delivers significant performance gains and physical insights:

  • 5x Acceleration: The ML surrogate provides a five-fold speedup compared to standalone HPC simulations, drastically reducing the time required for long-term dynamic analysis.
  • High Fidelity: The model maintains a resonance-frequency error of less than 2.3%, even when predicting dynamics at spatial locations not seen during training.
  • Nonlinear Dynamics: The framework successfully captures complex phenomena such as Suhl instability and the characteristic "anti-crossing" behavior of strong magnon-photon coupling.
  • Scalability: The underlying ARTEMIS solver's ability to scale to thousands of GPUs allows for full-chip modeling of large-scale quantum circuits.

Conclusion and Future Directions

The integration of HPC-driven data generation with physics-informed ML surrogates represents a powerful paradigm for modeling hybrid quantum systems. While current models treat spatial points independently, future work will focus on spatiotemporal architectures to better capture the coupled evolution of 2D and 3D fields. This framework is extensible to other multi-physics domains, including systems involving phonons or superconducting qubits, offering a versatile tool for the development of future quantum technologies.

Summary updated

Abstract

We introduce a hybrid computational framework that merges HPC-based numerical solvers with physics-informed ML surrogates for efficient modeling of magnon-photon interactions. By running short-duration, high-fidelity Maxwell-LLG simulations and feeding their results into an ML model, we substantially cut simulation time while achieving accurate predictions across larger spatiotemporal domains.
BibTeX
@misc{song2025hpc,
  author = {Jialin Song and Yingheng Tang and Pu Ren and Shintaro Takayoshi and Saurabh Sawant and Yujie Zhu and Jia-Mian Hu and Andy Nonaka and Michael W. Mahoney and N. Benjamin Erichson and Zhi Yao},
  title = {{HPC-Driven Modeling with ML-Based Surrogates for Magnon-Photon Dynamics in Hybrid Quantum Systems}},
  year = {2025},
  eprint = {2510.22221},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.22221},
}
Figure