Liquid Neural Networks for Zero-Retraining Adaptation in Highly Volatile NVMe-oF Storage Fabrics

Modern Storage Area Networks (SANs), particularly those leveraging NVMe over Fabrics (NVMe-oF), are increasingly dynamic, characterized by frequent topological mutations such as dynamic pathing, virtual machine migrations, and elastic scaling of storage targets. Traditional deep learning models applied to SAN traffic engineering or congestion mitigation suffer from severe performance degradation when the underlying topology changes, requiring computationally expensive offline retraining. In this paper, we introduce a novel application of Liquid Neural Networks (LNNs)—a class of continuous-time recurrent neural networks capable of adapting their parameters post-training—to storage fabric management. By deploying LNNs on switch telemetry streams, we present a routing and multipath optimization framework that autonomously adjusts to dynamic Fabric Shortest Path First (FSPF) updates and link volatility in real-time. Our simulated evaluations across complex, multi-tenant NVMe-oF topologies suggest that the LNN-based architecture maintains low I/O latency and stable throughput with zero retraining, outperforming static multipathing algorithms and traditional static-weight neural networks in the tested simulation scenarios by reducing microburst-induced latency by up to 88% during active topology mutations. We state upfront that these findings come from a single, non-repeated run of a custom, currently unreleased discrete-event simulator rather than repeated trials, multiple random seeds, or physical hardware; they should be read as an initial, illustrative demonstration of the architecture rather than a statistically validated or production-confirmed result. Section VII discusses this scope, and the steps needed to strengthen it, in detail.