Virtual Sensing to Enable Real-Time Monitoring of Inaccessible Locations & Unmeasurable Parameters

The article introduces MIMONet, a neural-operator framework designed for operator-based virtual sensing in nuclear-grade thermal-fluid systems, enabling real-time monitoring of interior states where physical instrumentation is impractical. MIMONet employs multi-modal branch encoders, multiplicative branch fusion, and shared-latent multi-field decoding, achieving less than 5% relative error and sub-millisecond inference times on data-center accelerators, even under 50% sensor noise. This advancement is significant for practitioners as it offers a method to maintain observability in complex energy systems, paving the way for improved safety and efficiency in scenarios lacking direct measurement capabilities.