Workshops Archive

Fast and Accurate ML Surrogates from Simulation: An HPC Liquid-Cooling Models Case Study


Workshop: Sustainable Supercomputing

Authors: Yishak Tadele Nigatu (University of Trento); Wesley Brewer, Jonathan McConnell, and Valentine Anantharaj (Oak Ridge National Laboratory); and Sandro Fiore (University of Trento)

Abstract: ExaDigiT is an open-source state-of-the-art digital twin framework for liquid-cooled HPC facilities. This framework models cooling systems using Functional Mock-up Unit (FMU)-based thermo-fluid simulators, which are computationally expensive and impractical for large-scale and real-time applications. This paper develops deep learning surrogates from a baseline LSTM through Deep Operator Network (DeepONet) variants to a Deep Multi-scale \& Multi-physics Network (DeepM\&Mnet) architecture. The paper builds on comprehensive data analysis that revealed structural properties of the cooling system: primary-secondary loop dynamic decoupling, spatial independence among Cooling Distribution Units (CDUs), and variable-dependent thermal time constants. Because most of these outputs are strongly persistent, we score against a persistence baseline in addition to $R^2$. Evaluated across 2,827 outputs (11 $\times$ 257 CDUs), the model beats persistence on 1,541 of the 1,799 outputs that carry learnable structure (85.7\%); the remaining 1,028 are shown to be persistence-locked, in particular, one group at a measured sensor noise floor. The mean $R^2$ is 0.926 (median 0.999), and at the 60\,s evaluation window, one forward for all 257 CDUs costs 4.1\,ms against 79.3\,s of FMU wall clock, a 19{,}514$\times$ speedup, turning a 10{,}000-member uncertainty ensemble run from 220 hours into 41 seconds. Architectures agree on aggregate $R^2$ to within 0.0002.


Back to Sustainable Supercomputing Archive Listing

Back to Full Workshop Archive Listing