SPINONet: Scalable Spiking Physics-informed Neural Operator for Computational Mechanics Applications

The architectural separation strategy is theoretically sound and practically well-executed. Confining discontinuities to the branch network while maintaining differentiable trunk pathways for coordinate derivatives (Eq. 9) correctly addresses the incompatibility between spiking dynamics and physics-informed residual computation. The hardware-agnostic energy analysis in Section 3.2 rigorously derives how synaptic MAC operations scale linearly with input spiking activity $\alpha_{\text{in}}$, establishing that energy parity with dense ANNs is approached only at high activity levels ($\sim 90\%$ for $T_s=1$). The forward-mode automatic differentiation analysis (Section 3.4) correctly exploits the scalar-input structure of separable trunks to reduce Jacobian computation costs from exponential to linear in dimensionality.

Despite the theoretical elegance, several empirical and methodological limitations warrant caution. First, the accuracy trade-off is non-negligible: SPINONet achieves 0.016 relative $L^2$ error versus 0.007 for the vanilla baseline on the Eikonal equation, and 0.07 versus 0.06 on Burgers (Table 2), suggesting the spiking regularization or surrogate gradient approximation introduces approximation error. Second, the energy analysis, while thorough, is hardware-agnostic and assumes ideal event-driven execution without validating actual power consumption on neuromorphic or edge hardware. Third, the degenerate convergence modes observed in the Eikonal example (erroneous and flipped-sign solutions in Figure 7) necessitate supervised data augmentation, which partially undermines the "physics-only" training paradigm for ill-posed problems. Finally, spiking activity rates of 28%–65%, while sparse, remain far from the ultra-sparse regimes ($<10\%$) where event-driven architectures achieve transformative efficiency gains.

The experimental evidence supports the scalability claims but reveals modest predictive trade-offs compared to dense separable baselines. The separable architecture demonstrates clear computational advantages over non-separable PI-DeepONet, with memory and runtime scaling favorably with grid resolution (Figure 3). However, SPINONet shows nearly identical runtime to the vanilla separable baseline ($\sim 0.013$s vs $\sim 0.012$s per epoch for Burgers, Table 2), suggesting the separable structure dominates computational costs rather than the neuron sparsity. The comparison would benefit from inclusion of other energy-efficient baselines such as quantized or pruned ANNs to contextualize whether spiking offers advantages beyond conventional compression. The hybrid training strategy (adding supervised loss) improves stability for the Eikonal equation, but the paper does not quantify how much labeled data is required for other problems, limiting generalizability of this fix.

The paper provides detailed algorithmic descriptions (Algorithms 1–3) and comprehensive architectural specifications in Table 1, including network depths, latent dimensions, and separable ranks. The mathematical formulation of VSN dynamics (Eqs. 10–11) and surrogate gradient training (Eq. 34 with slope parameter $K_s$) is sufficiently detailed for implementation. However, the authors do not mention code availability or open-source release, which is critical for verifying the energy calculations and spiking activity metrics. Specific hyperparameters such as the exact firing threshold $\mathcal{T}_h$, leakage parameter $\beta_l$, and surrogate gradient slope $K_s$ values used in experiments are not explicitly reported. Additionally, the random seeds, initialization schemes for VSN membrane potentials, and specific optimizer learning rate schedules would be necessary for exact reproduction of the reported errors.