(202b) Machine-Learning Model for the Discovery of Dielectric Cooling Fluids

Data centres are the backbone of the world’s modern IT and internet needs. The data centre industry is expected to consume approximately 4.5% of the world’s electricity demand by 2025 [[i]], and approximately 40% of this energy is employed for uninterruptible cooling systems. Efficient direct liquid cooling is an efficient substitute for traditional air cooling; the design of these thermal management fluids has become an increasingly important area of research-oriented toward reducing energy consumption and reusing waste heat. Currently, cooling fluids under development include hydrocarbons and high molecular weight esters, whether natural or synthetic. These fluids are required to meet stringent performance goals for thermal management, signal integrity, material compatibility, and reliability for deployment in IT equipment.

Machine learning models can transform materials discovery by allowing the rapid prediction of properties and the generation of candidate materials. However, these models require large volumes of high-quality training data to accurately learn the underlying chemical space—a challenge when such data is limited. Dielectric cooling fluids are a case in point: their performance depends on various thermophysical properties, including density, heat capacity, and transport properties such as viscosity and thermal conductivity. However, data for these fluids is unavailable for much of the chemical space, which includes a diverse range of organic chemistries and is sparse across the phase space, mainly concentrated around ambient temperature.

We showcase the development of a Graph Neural Network (GNN) to predict the thermophysical properties of potential thermal management fluids. Our training dataset combines high-quality experimental data from NIST REFPROP and DIPPR 801 with lower-accuracy pseudo-data generated by molecular dynamics (MD) simulations. We use experimental data from over 1,500 compounds, curated to include only C, H, O, N and Si atoms, assessed at four different temperatures ranging from 85 to 1200 K. After evaluating the available chemical space, we “plug the holes” by generating pseudo-data from MD simulations utilizing the Transferable United-Atom force fields based on the Mie potential [[ii]]. The lower reliability of this latter data is incorporated into the loss functions by including estimates of uncertainty. The fits to density and heat capacity are robust, while those for thermal conductivity and viscosity have the highest errors, mainly attributed to the uncertainty in the data.

The trained GNN is deployed to analyze the molecular trends that contribute to higher “figures of merit” for thermal fluids and to prioritize potential candidate fluids for immersion cooling. We demonstrate that integrating lower-accuracy data can significantly improve the GNN’s performance by bridging gaps in chemical space—provided that the quality of the data is adequately considered.

References

[i] Azarifar, M., Arik M. and Chang, J. Y. Liquid cooling of data centers: A necessity facing challenges Appl Therm Eng, 247, (2024),

[ii] Fischer, M., Bauer, G. and Gross, J. Transferable Anisotropic United-Atom Mie (TAMie) Force Field: Transport Properties from Equilibrium Molecular Dynamic Simulations. Industrial & Engineering Chemistry Research 59, 8855–8869 (2020).