Nanoscale neural network using non-linear spin-wave interference

The study addresses the challenge of implementing neuromorphic computing in efficient hardware that natively supports high interconnectivity and nonlinearity. Conventional digital platforms (CPUs/GPUs) are inefficient for analog neuromorphic tasks, while many neuromorphic models require dense, often all-to-all interconnections. Wave-based substrates naturally provide high connectivity via interference, but linear interference alone is limited to signal processing and lacks the computational richness needed for general-purpose neuromorphic functions. Prior theoretical work mapped nonlinear wave media to recurrent neural networks (RNNs), but left open whether nonlinear propagation can outperform linear systems and how to realize the assumed nonlinearity physically. This paper proposes and demonstrates a physically realizable spin-wave (magnonic) platform in ferrimagnetic thin films that offers both strong interconnection via interference and intrinsic nonlinearity at moderate amplitudes. The core research question is whether a designed magnetic-field landscape can implement trainable, neuromorphic input–output mappings entirely via spin-wave propagation and interference, and whether leveraging nonlinear spin-wave dynamics materially improves computational performance over linear interference. The work introduces a Pytorch-based inverse-design engine (Spintorch) coupled to a full micromagnetic solver to design such magnetic-field patterns and explores performance in linear and nonlinear regimes.

The work builds on wave-based computing and photonic inverse design, where linear interference provides high connectivity but limited computational power due to lack of intrinsic nonlinearity. Hughes et al. (Sci. Adv. 2019) provided a theoretical framework mapping nonlinear wave equations to RNNs via gradient-based training of spatially varying, intensity-dependent wave speeds, but practical advantages over linear propagation and concrete physical realizations remained unclear. Spin-wave (magnonic) systems in thin magnetic films have been extensively studied for logic, signal processing, and optics-inspired computing, with advantages including room-temperature coherence, sub-100 nm wavelengths at microwave frequencies, and compatibility with electronics. Inverse design has accelerated progress in nanophotonics and metamaterials and has been proposed for magnonics as well. The authors reference prior magnonic devices such as spectrum analyzers, couplers, and logic, and contrast with optical nonlinearities that typically require high intensities and separate components. This study positions spin waves as a medium that naturally exhibits nonlinearity at moderate amplitudes, enabling neuromorphic computation within a single substrate.

Physical platform: a low-damping YIG thin film acts as the spin-wave substrate. A spatially non-uniform magnetic field (bias-field landscape) programs the device by locally modifying spin-wave dispersion and steering waves to create desired interference patterns. In one implementation, arrays of perpendicular magnetic anisotropy (PMA) nanomagnets placed above the YIG produce a reconfigurable, punchcard-like up/down field pattern that serves as the trainable weights; in others, a fine-grained external field grid is directly optimized to increase design degrees of freedom. Inverse-design engine (Spintorch): a custom micromagnetic solver integrated with the Pytorch automatic differentiation framework performs gradient-based optimization of trainable parameters (magnet states or external field values) to achieve target input–output mappings. The solver discretizes the film into 25 nm × 25 nm × 25 nm cells and integrates the Landau–Lifshitz–Gilbert (LLG) equations including damping, fully accounting for demagnetizing (dipolar) fields, exchange interactions, and external fields (bias, time-dependent excitation, and PMA magnet stray fields). Nonlinearity arises naturally via the amplitude-dependent demagnetizing field. Dipolar fields are computed using FFT-accelerated Poisson solutions on GPU; exchange is computed via convolution with a Laplacian kernel. Time integration uses a fourth-order Runge–Kutta scheme. Absorbing boundaries are realized via damping. Solver outputs are time-integrated spin-wave intensities over specified detector areas. Gradients are obtained via Pytorch backpropagation through time. Design tasks and parameters:

• Linear regime inverse design: Spectrum analyzer on a 10 µm × 10 µm area to route f1=3 GHz, f2=3.5 GHz, f3=4 GHz to distinct 300 nm-diameter output regions. Training converged in ~30 epochs using small-amplitude excitations (precession angles of a few degrees; excitation fields in mT range). With magnets, ~300 binary variables; with direct field optimization, up to 16,000 continuous variables.
• Vowel recognition: Input waveforms (male samples from Wavetorch dataset) for vowels “ae,” “ei,” and “iy” (cropped to vowel regions) were frequency-scaled to excite propagating modes compatible with device dimensions. One CPW input launches waves; three spatial outputs correspond to the three vowels. Training set: 4 samples per vowel (12 total). Test set: 41 samples per vowel (123 total). Training ran for 30 epochs per setting. Both linear (~1 mT) and nonlinear (~50 mT) excitation amplitudes were tested; additional amplitudes scanned to map linear, nonlinear, and chaotic regimes.
• Nonlinear separability example: Two-frequency logic-like task with f1=3 GHz and f2=4 GHz. Training targets: focus to o1 for f1 alone; focus to o1 for f2 alone; focus to o2 if and only if f1 and f2 are simultaneously present. Trained at 1 mT (linear), 20 mT (moderately nonlinear), and 50 mT (strongly nonlinear within non-chaotic regime) for 30 epochs on a 10 µm × 10 µm device. Computational constraints: Backpropagation through all time steps requires storing intermediate states, limiting simulated system size and dataset size by GPU memory. Solver correctness was validated against MuMax3 for comparable cases.

• Linear inverse design: The micromagnetic inverse-design approach successfully configured a spin-wave scatterer to perform linear RF signal processing, including frequency demultiplexing of 3.0/3.5/4.0 GHz components to distinct 300 nm outputs on a 10 µm × 10 µm substrate. Designs converged in ~30 epochs and were robust to moderate magnet-state errors. Direct field optimization with 16,000 continuous variables improved selectivity over ~300 binary magnet variables at similar compute cost.
• Vowel recognition: Both linear (1 mT) and nonlinear (50 mT) regimes achieved 100% training accuracy on the small training set (4 samples/vowel), but generalization differed markedly on the 123-sample test set. • Linear regime: Lower test performance; frequent misidentification of all three vowels, with particularly poor accuracy for “ae” (about 60%). • Nonlinear regime: Significantly better test performance; confusion matrix was close to diagonal, with only 7 misclassifications out of 123 (≈94.3% accuracy). Nonlinear training also converged faster and to lower loss than linear. • Amplitude sweep revealed three regions: (i) linear regime with modest, amplitude-independent accuracy; (ii) nonlinear regime where accuracy jumps and peaks around 50 mT; (iii) strongly nonlinear/chaotic regime at higher amplitudes where accuracy degrades below linear.
• Nonlinear separability (two-frequency task): The target mapping (o1 for single tones, o2 only for simultaneous tones) is impossible for a linear system due to superposition. Results matched this: at 1 mT the device failed (outputs were linear combinations), while at 20 mT and 50 mT the device correctly focused to o2 for simultaneous tones and suppressed o1, with 50 mT yielding the strongest separation and fastest convergence.
• Energy and speed estimates: The magnetic energy stored in representative nonlinear operation patterns is ~1000 eV (~1×10⁻¹⁶ J), with build-up times ~10 ns, indicating potentially low-energy, high-speed operation intrinsic to the magnetic domain. However, system-level energy is dominated by electrical I/O (especially readout), estimated at ≈1×10⁻¹¹ J per output at GHz rates due to required microwave amplification of sub-µV signals.

The findings demonstrate that spin-wave substrates can implement neuromorphic computation entirely via wave propagation and interference, with the programmable magnetic-field landscape functioning as trainable weights. Linear interference suffices for certain RF signal-processing tasks (e.g., frequency demultiplexing), but its computational power is limited; it struggles to generalize in pattern recognition tasks. Introducing nonlinear spin-wave dynamics substantially enhances computational capability, enabling tasks incompatible with linear systems and improving generalization on previously unseen data, consistent with behavior akin to multilayer networks or RNNs. The system realizes weighted interconnections and distributed activation within a single film, avoiding repeated I/O conversions that would negate wave-computing advantages. Energy analyses suggest orders-of-magnitude lower intrinsic computing energy than digital approximate operations; yet, practical system energy is constrained by transducers and readout electronics, where current microwave amplification overheads can dominate. Compared to optical reservoir computing, the proposed device can achieve similar or better energy per operation with a smaller footprint and integrated nonlinearity, though strictly linear optical systems can be more energy-efficient for linear tasks. The primary practical bottleneck identified is design scalability: inverse design with full micromagnetics and backpropagation is compute- and memory-intensive, limiting device size and dataset scope in this study. Nevertheless, the results support the hypothesis that nonlinear wave interference provides RNN-like computational capabilities and establish a path toward compact, low-power, integration-friendly magnonic neural hardware.

This work introduces Spintorch, a Pytorch-integrated micromagnetic inverse-design framework, and demonstrates experimentally realistic spin-wave scatterers that implement neuromorphic functions. In the linear regime, the approach automates the design of compact RF signal processors (e.g., spectrum analyzers). In the nonlinear regime, it realizes neural-network-like computation within a single magnetic film, combining interconnections and activation via nonlinear spin-wave interference. Empirically, nonlinear operation significantly improves generalization in vowel recognition and enables tasks impossible with linear superposition. Energy estimates indicate substantial intrinsic efficiency and high speed, although current I/O electronics dominate system-level energy. Future directions include scaling to larger substrates and richer datasets via more efficient training or surrogate models, improving on-chip transduction and readout (e.g., leveraging inverse spin-Hall effect or optimized antenna geometries), exploring alternative materials and programming methods (e.g., FIB-tuned films or metallic ferromagnets), and exploiting recurrent dynamics and fading memory for temporal processing. Experimental realization and benchmarking against state-of-the-art neural accelerators will further clarify performance and energy advantages.

• Computational scalability: Backpropagation through time with full micromagnetics requires storing all intermediate states, limiting simulated device size and dataset size by available GPU memory. Larger and more complex networks could not be explored here.
• Training data size: The vowel-recognition training set was very small (4 samples per class), making 100% training accuracy unsurprising and complicating broader generalization claims.
• Operating regime sensitivity: At high excitation amplitudes, spin waves exhibit chaotic dynamics that degrade performance; robust operation requires staying within a moderate nonlinear range.
• System-level energy dominated by I/O: Readout of small, high-frequency signals necessitates substantial microwave amplification (≈10 mW), yielding ~1×10⁻¹¹ J per output at GHz rates, which can overshadow intrinsic magnetic-domain efficiency.
• Physical assumptions and integration: PMA magnets were assumed bistable and unaffected by underlying spin waves; device-to-device variability, switching errors, and fabrication tolerances may impact performance. While designs appear robust to some magnet-state errors, large-scale integration and reconfigurability require further experimental validation.
• Scope: The study focuses on simulation; larger-scale comparisons with state-of-the-art neural models and comprehensive experimental demonstrations remain for future work.