Dual-Model Machine Learning for Predictive Leak Forecasting and Detection in Liquid-Cooled AI Data Centers

Abstract

Modern GPU data centers supporting AI training workloads have increas- ingly adopted direct-to-chip liquid cooling systems to manage thermal loads exceeding 50 kW per rack, far beyond air cooling capabilities. However, coolant leaks in these high-density facilities result in substantial energy waste through unplanned shutdowns, extended repair periods, and preventive isolation of adjacent racks. We present a novel smart IoT monitoring system combining LSTM neural networks for probabilistic time- to-leak forecasting with Random Forest classifiers for real-time binary detection. The dual-model architecture provides both advance warning (2-4 hours) for planned mainte- nance and immediate alerts (sub-minute latency) for sudden failures. Validation using simulation-based data generation following ASHRAE 2021 specifications demonstrates strong performance: 96.5% F1-score for binary detection and 87% forecasting accu- racy at 90% probability within ±30-minute windows. The dataset comprises 72 hours of minute-resolution monitoring with realistic leak scenarios incorporating documented industry patterns. Statistical analysis reveals strong predictive signals from humidity (r = 0.70, p ¡ 0.001), pressure (r = -0.50), and flow rate, while temperature shows minimal immediate response (p = 0.236) due to thermal inertia, guiding optimal sen- sor deployment. The integrated system achieves 98.4% coverage with 850ms end-to-end latency. Energy analysis shows this approach could prevent approximately 1,500 kWh annual waste for a 47-rack facility, supporting sustainable operations. The complete implementation is provided to facilitate validation in operational environments, estab- lishing a foundation for intelligent leak management as liquid cooling becomes standard in AI infrastructure