Generative Pipeline for the Discovery of Single-Phase Immersion Coolants for Data Centres

Immersion cooling offers a route to managing the rising heat densities of data centres, but identifying high-performance dielectric liquids remains an open problem. Candidates must optimise several coupled thermophysical properties while meeting dielectric, safety, and compatibility constraints. We present a pipeline that trains a graph-to-SMILES VAE semi-supervised on DIPPR 801 and searches its latent space with an iterative KDE sampling scheme, exploiting the local structure of the Figure of Merit (FOM). One returned candidate was characterised experimentally with our industrial partner: its measured FOM of 233 exceeds the performance of commercial coolants. Two further candidates are under investigation.