NuTwin: A Zero-Trust, Confidential Cloud Architecture for Real-Time Nuclear Reactor Digital Twins Using Physics-Informed Neural Networks

This is the authors' abstract. We don't add key points for this paper.

Real-time telemetry diagnostics and predictiveanomaly tracking in nuclear reactor cores present a fundamentalcomputational dilemma: high-fidelity numerical neutronics andthermal-hydraulics solvers are too computationally prohibitive forsub-second edge evaluation, whereas conventional black-box deeplearning architectures violate physical conservation bounds andlack operational determinism. Simultaneously, transmitting safetycritical nuclear operational data to public cloud infrastructureintroduces acute cyber-physical security vectors that breach strictoperational air-gaps. This paper introduces NuTwin, a zero-trust,confidential cloud architecture designed for real-time nuclearreactor core digital twins. NuTwin couples a Physics-InformedNeural Network (PINN) surrogate model—constrained directlyby point reactor kinetics equations (PRKE) with six delayedneutron precursor groups and coupled thermal-hydraulic reactivity feedback—with hardware-enforced cryptographic isolation.Live operational telemetry, simulated via Python Point ReactorKinetics (PyRK), is streamed over unidirectional mutual TLS(X.509) connections into AWS IoT Core and processed exclusivelyinside an ephemeral, attestation-verified AWS Nitro Enclave viaan internal virtual socket (vsock) channel devoid of persistentstorage or external network interfaces. Experimental validationdemonstrates that NuTwin bounds the physical residual loss below1.4 × 10−4, achieves sub-second inference latency (42.8 ms perstep), and guarantees zero data-leakage surface under zero-trustoperational constraints.