This study presents a proof-of-concept coolant-leak monitoring system for direct-to-chip liquid cooling in GPU-focused data centers. Its architecture combines an LSTM model for probabilistic forecasting of time to leak with a Random Forest model for rapid detection of leaks that have already occurred, using MQTT, InfluxDB, and Streamlit for data flow and visualization. The evaluation relies solely on synthetic data aligned with ASHRAE 2021, meaning that the reported performance and energy benefits require validation at live sites.
Key Takeaways
The proposed approach for liquid-cooled GPU data centers combines LSTM-based leak forecasting with Random Forest-based real-time detection, rather than relying only on conventional reactive threshold alarms.
Humidity, coolant pressure, and flow rate are identified as the most informative early indicators of a leak; temperature is less suitable for immediate detection because of the thermal inertia of equipment and the rack air environment.
In the synthetic simulation, the system forecasts leaks several hours in advance and detects sudden events within a minute, potentially allowing controlled isolation and workload migration before an emergency shutdown.
For a site of several dozen racks, the estimated benefit is roughly one and a half megawatt-hours of avoided annual energy loss; however, this is a modeled estimate rather than a demonstrated operational result.
The work is not yet a basis for production automation without multi-site validation, stability testing over more than six months, BMS/DCIM integration, and operator training.
Key Figures
Leak detection accuracy (Random Forest F1 score)
96.5 %
Result obtained using synthetic data.
Leak forecasting accuracy
87 %
Measured at the specified probability level and within the defined time window; based on synthetic data.
Forecast probability
90 %
Probability threshold or confidence level used to assess forecasting accuracy.
Forecast evaluation window
±30 минут
Time window around the predicted leak event.
Early-warning forecast horizon
2–4 часа
Claimed lead time before a leak occurs.
Sudden-leak detection time
в пределах 1 минуты
Claimed Random Forest event-identification speed.
Synthetic monitoring duration
7 дней
Minute-by-minute data used for the proof of concept.
Number of IoT sensors
4 датчика на стойковую оболочку
Sensors per rack enclosure used in the synthetic model.
Modern GPU rack power
30–50 kW
Typical rack load cited in the document.
Size of the assessed facility
47 стоек
Representative facility used to calculate the energy effect.
Potentially avoidable energy losses
около 1 500 kWh в год
Modeled annual estimate for the representative facility under proactive maintenance.
Failure reduction with predictive maintenance
50 %
General introductory claim, referenced to engineering-equipment maintenance practices; not a result of the authors' experiment.
Mean time to repair
6 часов
Assumption used in the energy-loss calculation.
Energy loss during repair
240 kWh на стойку
Estimated loss per rack during the assumed downtime.
Preventively shut down neighboring racks
2–3 стойки
Typical operator response practice described in the study.
Energy loss per incident
около 600 kWh на инцидент
Calculated aggregate loss when neighboring racks are shut down preventively.
Leak frequency under reactive maintenance
3–5 инцидентов на 100 стоек в год
Industry benchmarks used by the authors.
Expected leaks at the assessed facility
около 2.5 события в год
Calculated from the assumed incident rate.
Modeled coverage of the two-model architecture
98.4 %
Claimed practical estimate, corresponding in the model to prevention of up to the stated number of incidents per year.
Practical Value for Data Center Owners
For data center owners and project teams, the study outlines an architecture that could be taken to pilot deployment on direct-to-chip cooling loops: humidity, pressure, and flow sensors combined with streaming analytics that can integrate with DCIM/BMS platforms. Its practical value lies in shifting from detecting the consequences of a leak to providing early warning, enabling isolation of a loop or rack and workload migration. However, the reported accuracy and savings estimates should not be treated as a guaranteed business case: they are based on a seven-day synthetic dataset and require field validation that accounts for sensor drift, equipment differences, and real-world failure scenarios.
This strategic white paper from the Open Compute Project outlines its Open Data Center for AI program, focused on deploying gigawatt-scale AI data centers. It calls for open reference site designs, standardized grid-interconnection practices, SLAs for third-party operators, integrated telemetry across IT and facility infrastructure, and consistent power-assessment methodologies. The initiative aims to accelerate AI data center construction and grid connection, increase infrastructure reuse, and reduce the risk of supply-chain misalignment.
AI data centersgigawatt-scale data centersdata center reference designsliquid cooling
An ASHRAE technical white paper on the factors accelerating the transition of data centers from predominantly air-based to liquid cooling. The document links rising CPU, GPU, FPGA, and memory power levels to the limitations of air-based systems in terms of airflow, temperature, noise, rack density, and personnel safety. It examines ASHRAE water classes, direct liquid cooling and immersion cooling, and the implications for PUE, TCO, power delivery, heat reuse, and deployment in enterprise, HPC, and colocation data centers.
Liquid coolingDirect Liquid CoolingImmersion coolingData center cooling