Sparse Neural Operators Enable Real-Time Edge Virtual Sensing

Virtual sensing is crucial for digital twins and safety-critical systems, enabling real-time reconstruction and forecasting of spatio-temporal physics. However, deploying these capabilities on edge devices faces significant challenges in terms of generalization, latency, and energy efficiency. This research proposes a solution through Sparse-Activation-ReLU (SAR) layers integrated into a trunk-based NOMAD architecture. SAR layers promote activation sparsity without requiring surrogate-gradient training, making them compatible with event-based computing and improving efficiency. The SAR approach demonstrates over a fivefold improvement in the combined Latency-Error-Energy (LEE) metric compared to existing spiking neuron implementations. Further enhancements are achieved through synthetic knowledge distillation, which more than halves the LEE score. The work also refines other spiking neuron models, significantly reducing L2 error and improving efficiency. Overall, this research represents a substantial step towards energy-efficient virtual sensing, providing a framework for neuromorphic or other edge device integration that could become a benchmark for future sparsity- or brain-inspired designs.