Neural Network Tuning Reveals Which Cooling Metrics Rule High-Density Data Centers

Data center cooling is being transformed by a quiet revolution in numbers. A new study published in Discover Networks applies neural networks to a set of thermal and hydraulic metrics describing liquid-to-air and liquid-to-liquid cooling architectures, and asks a deceptively simple question: which of those metrics actually matter most, and does the answer change when the neural network itself is reconfigured? The findings, based on a synthetic thermo-hydraulic dataset of 100 simulated operational instances, point to a remarkably consistent answer. Temperature differences, airflow, and mass flow dominate system performance across nearly every model configuration tested, regardless of how many hidden layers or neurons the network contains.

The research arrives at a moment when thermal management in computing facilities has become a defining constraint on artificial intelligence. According to the study, rack densities that once ranged from 5 to 10 kilowatts per rack have climbed to 30, 50, or even 100 kilowatts, driven by AI training, high-performance computing, and compact edge servers. Conventional air-based cooling, the authors argue, cannot sustainably meet these demands. Liquid cooling, once considered a niche or experimental approach, is now widely adopted by hyperscale operators, colocation providers, and enterprise data centers worldwide, particularly in edge environments supporting latency-sensitive mobile and Internet of Things applications.

Methodologically, the study is distinctive in how it interrogates its own models. The researchers built two synthetic databases, one for liquid-to-air cooling performance and one for liquid-to-liquid performance, using representative parameter ranges drawn from prior literature, vendor technical documentation, and publicly available industry reports. Variables were defined and processed in IBM SPSS Statistics version 25, with all metrics treated as continuous scale variables representing thermodynamic and fluid-dynamic behavior. The liquid-to-air dataset includes CPU utilization, IT load, inlet and outlet temperatures, airflow, mass flow rate, fan speed, and pressure drop. The liquid-to-liquid dataset covers secondary flow rate, primary and secondary inlet and outlet temperatures, temperature differences, heat transfer rate, thermal efficiency, pump power, and pressure drop.

Two neural network families were then trained against these metrics. The first, a Multilayer Perceptron, was tested with one and two hidden layers, with the number of units increased in steps of two across configurations ranging from 2 to 30 neurons. Hyperbolic tangent activation was used in hidden layers with an identity function at the output, batch training, and a scale-conjugate gradient optimizer. The second family, a Radial Basis Function network, was evaluated with varying minimum and maximum unit counts per hidden layer. A recursive feedback mechanism allowed the team to systematically vary network depth, hidden unit allocation, and learning configurations, then measure how each change reshaped the normalized importance of every input variable.

The sensitivity analysis produced a clear hierarchy. For the single-hidden-layer MLP applied to liquid-to-air metrics, temperature difference, IT load, and airflow consistently showed strong relative influence, reaching peak normalized importance values of 100 percent in multiple scenarios. CPU utilization peaked at 12 and 30 hidden units, while IT load peaked at 18 and 26 units, indicating that feature attribution is genuinely sensitive to architecture. Pressure drop and fan speed contributed moderately but in a configuration-dependent way. Configurations with an intermediate neuron count of roughly 16 to 24 units offered the best balance between predictive performance and model complexity, which the authors propose as a practical design guideline.

The liquid-to-liquid analysis told a parallel story with different protagonists. Inlet temperature on the primary side and outlet temperature on the secondary side repeatedly reached 100 percent normalized importance across multiple configurations. As neuron counts increased beyond 18 units, secondary-side temperature difference, heat transfer rate, and thermal efficiency grew more influential. Interestingly, the authors explain that these three indicators never quite reach full normalized importance because their values are scaled relative to the most influential variable in each configuration, and because they are strongly physically interdependent, so their importance is distributed among them rather than concentrated in one. Hydraulic parameters such as pump power and pressure drop showed moderate but non-negligible contributions, hinting at coupled thermo-hydraulic effects.

Adding a second hidden layer shifted the rankings again. In the two-hidden-layer MLP for liquid-to-air cooling, CPU utilization, IT load, and temperature difference emerged as dominant predictors, frequently hitting 100 percent relative contribution, while inlet and outlet temperatures showed moderate, architecture-sensitive influence. For liquid-to-liquid cooling with two hidden layers, primary inlet temperature and secondary temperature difference consistently achieved the highest normalized importance, and heat transfer rate and thermal efficiency gained relevance in architectures with 14 or more units. The Radial Basis Function results were strikingly stable by comparison: importance values hovered around 0.110 to 0.112 across nearly all liquid-to-air variables, with normalized importance mostly above 97 percent, while pressure drop alone varied more noticeably.

The authors are careful to frame the scope of these findings. The figures and datasets are explicitly synthetic trend-oriented simulations rather than direct measurements from a physical experimental platform, generated to enable controlled evaluation of thermo-hydraulic behavior under scenarios that would be difficult to reproduce consistently in production data centers. The team notes that acquiring large-scale real-world thermal datasets from operational AI-oriented facilities remains technically challenging due to infrastructure restrictions, operational sensitivity, cost, and limited public availability. Within that framing, the synthetic deployment analysis suggests widespread cooling distribution unit adoption, with a 92 percent implementation rate in hybrid air-and-liquid environments across the sample of 100, facility-level form factors predominating at 45 percent, and liquid-to-liquid heat exchange accounting for 70 percent of solutions versus 30 percent for liquid-to-air.

The connection to mobile and IoT infrastructure runs throughout the work. The authors argue that 5G and 6G networks and large-scale IoT deployments generate highly dynamic, dense computational workloads across both centralized and edge data centers, producing thermal stress and power variability. The consistent dominance of thermodynamic variables such as temperature difference, airflow, and mass flow suggests, in their interpretation, that efficient heat dissipation is what maintains reliable operation under fluctuating demand, while the strong influence of CPU utilization and IT load links network traffic directly to thermal behavior. They propose that such neural network models could eventually inform thermal-aware resource management in edge computing environments.

The study also sketches an ambitious roadmap. Future extensions could integrate transformer-based models with self-attention mechanisms to capture long-range dependencies among workload fluctuations, rack-level thermal variation, and cooling demand, or convolutional networks to analyze thermal maps and infrared rack imagery for hotspot detection. Diffusion-based generative models, the authors suggest, could produce high-dimensional synthetic thermo-hydraulic datasets preserving realistic operational distributions, valuable where real telemetry is restricted. Looking further ahead, they envision hybrid architectures combining these approaches into autonomous thermal-management ecosystems capable of dynamically regulating coolant flow, airflow distribution, pump speed, and thermal load balancing.

Subject of Research: Neural network sensitivity analysis of liquid cooling performance metrics in high-density data centers

Article Title: Sensitivity analysis of liquid-to-air and liquid-to-liquid cooling metrics under multilayer perceptron hyperparameter tuning for mobile and IoT-driven data centers

Article References: Cortés, A. (2026). Sensitivity analysis of liquid-to-air and liquid-to-liquid cooling metrics under multilayer perceptron hyperparameter tuning for mobile and IoT-driven data centers. Discover Networks, 2(1), Article 33. https://doi.org/10.1007/s44354-026-00035-0

Image Credits: AI Generated

DOI: 10.1007/s44354-026-00035-0

Keywords: data centers, liquid cooling, neural networks, multilayer perceptron, radial basis function, sensitivity analysis, thermal management, IoT, edge computing, synthetic data, rack density, hyperparameter tuning