Toward Multiphysics-Informed Machine Learning for Sustainable Data Center Operations: Intelligence Evolution with Deployable Solutions for Computing Infrastructure

The revolution in artificial intelligence (AI) has brought sustainable challenges in data center management due to the high carbon emissions and short cooling response time associated with high-power density racks. While machine learning (ML) offers promise for intelligent management, its adoption is hindered by safety and reliability concerns. To address this, we propose a multiphysics-informed machine learning (MPIML) framework that integrates physical priors into data-driven models for enhanced accuracy and safety. We introduce an integrated system architecture comprising three core engines: DCLib for versatile facility modeling, DCTwin for high-fidelity multiphysics simulation, and DCBrain for decision-making optimization. This system enables critical predictive and prescriptive applications, such as carbon-aware IT provisioning, safety-aware intelligent cooling control and battery health forecasting. An illustrative example on an industry-grade data center cooling control demonstrates that our MPIML approach reduces annual carbon emissions up to 200 kilotons compared with conventional methods while ensuring operational constraints are met. We conclude by outlining key challenges and future directions for developing autonomous and sustainable data centers.

💡 Research Summary

The paper addresses the growing sustainability challenges of modern data centers (DCs) driven by AI workloads, high power‑density racks, and the resulting carbon emissions and short cooling response times. While machine‑learning (ML) techniques have shown promise for intelligent DC management, their adoption is hampered by two fundamental issues: (1) the need for large, diverse datasets that capture rare or abnormal operating conditions, and (2) the lack of safety and reliability guarantees, which makes operators risk‑averse. To overcome these barriers, the authors propose a Multiphysics‑Informed Machine Learning (MPIML) framework that embeds physical priors from thermodynamics, fluid dynamics, psychrometrics, and electrical power theory directly into data‑driven models.

The MPIML system is built around three tightly coupled engines:

1. DCLib – a Python‑based, object‑oriented library that provides reusable classes for all major DC components (building envelope, cooling equipment such as CRAH units, chillers and cooling towers, power infrastructure including UPS and transformers, IT hardware, and workload descriptors). DCLib enables users to construct a complete digital representation of a facility from design through operation and automatically generates configuration files for downstream simulation.

2. DCTwin – a high‑fidelity, differentiable digital‑twin platform built on PyTorch and Nvidia PhysicsNeMo. It consumes DCLib’s configuration and runs multiphysics simulations that couple Navier‑Stokes equations, energy‑balance relations, psychrometric calculations, and power‑flow models. DCTwin integrates open‑source solvers (OpenFOAM for CFD, EnergyPlus for building energy, Mesmo for electrical distribution, CloudSim for IT scheduling) to produce both synthetic training data and rigorous validation benchmarks. By combining physics‑based models with data‑driven residual networks, DCTwin reduces the computational burden of full CFD runs through model reduction and transfer learning, while preserving physical consistency.

3. DCBrain – the decision‑making layer that supports model‑free reinforcement learning (RL), model‑based planning, and rule‑based optimization. Physical constraints (e.g., fan power ∝ speed³, conservation of energy, humidity limits) are incorporated as penalty terms or Lagrange multipliers in the RL loss, guaranteeing that learned policies never violate safety‑critical limits. DCBrain also consumes real‑time grid carbon intensity signals, enabling carbon‑aware workload placement and renewable‑integration strategies.

The authors illustrate the framework with an industry‑grade cooling‑control case study. Compared with a conventional PID controller and a pure‑data‑driven DRL controller, the MPIML‑augmented DRL policy reduces annual carbon emissions by up to 200 kilotons (≈0.2 Mt CO₂eq) while satisfying temperature, humidity, and power constraints. The inclusion of physics priors prevents unrealistic temperature spikes during training and narrows the simulation‑to‑real‑world performance gap to less than 15 %.

A key conceptual contribution is the “data‑physics spectrum” (Figure 2), which maps scenarios from data‑poor/design‑phase (simulation‑only) to data‑rich/operation‑phase (dense sensor streams) and from complete physics (all governing equations known) to no physics (black‑box). The paper identifies two prevalent middle‑ground regimes: (a) unknown parameters – the governing equations are known but coefficients (e.g., fan affinity constants) must be identified from operational data; and (b) missing terms – major processes are modeled but secondary effects (solar gain, infiltration, human metabolic heat) are omitted, requiring an ML residual to capture the unmodeled dynamics. This taxonomy guides practitioners on whether to focus on parameter identification or residual learning.

Future research directions highlighted include:

• Scaling multiphysics simulations to real‑time inference using GPU‑accelerated solvers and model‑order reduction techniques.
• Extending MPIML to privacy‑preserving federated learning across multiple DC sites.
• Formal verification of safety‑critical policies derived from DCBrain, possibly via reachability analysis or theorem proving.
• Developing a “digital‑twin‑of‑a‑twin” architecture where the twin itself is continuously refined by online data, enabling lifelong learning.

In summary, the paper presents a coherent, end‑to‑end MPIML ecosystem that bridges the gap between high‑fidelity physics and flexible machine learning, delivering tangible sustainability gains (lower PUE, reduced carbon footprint) while maintaining the safety and reliability required for mission‑critical data center operations.