Carbon-Aware Machine Learning for Energy-Efficient Quantum Data Centers: Joint Optimization of Workload Scheduling and Cooling

Abstract

Quantum workloads are increasingly delivered through cloud platforms, yet emerging quantum data centers face a distinctive energy challenge: facility power is often dominated by cryogenic cooling systems and auxiliary infrastructure rather than by compute alone. This study proposes a carbon-aware co-optimization framework for green quantum data centers that jointly integrates deadline-aware workload scheduling with dynamic cooling setpoint control. In contrast to prior studies that typically consider scheduling and thermal management independently largely in classical data-center environments the proposed framework evaluates their combined effects on facility-level energy consumption, operational carbon emissions, and fairness across tenants within an end-to-end simulation tailored to quantum infrastructure, where cooling setpoints strongly influence cryogenic efficiency. A simplified quantum data-center model with temperature-dependent cooling efficiency is developed, together with heterogeneous quantum job arrivals characterized by runtimes, priorities, and deadlines. Time-varying carbon-intensity and ambient-temperature signals are incorporated to emulate renewable-driven grid dynamics. Carbon-aware scheduling supports both flexible and strict deferral policies that shift eligible jobs toward lower-carbon periods while respecting deadline constraints. In parallel, a lightweight model predictive control strategy enumerates feasible cooling setpoint trajectories to minimize predicted facility energy use subject to temperature bounds and ramp-rate limits. Performance is evaluated across multiple random seeds, with uncertainty quantified using confidence intervals and statistical significance assessed via bootstrap testing. Simulation results show that MPC-based cooling control reduces total facility energy consumption by approximately 9% relative to a fixed-setpoint baseline. Carbon-aware job deferrals provide additional emissions reductions ranging from several percent to double-digit values depending on deferral aggressiveness, with explicitly quantified trade-offs in waiting time and SLA violations. Fairness impacts are assessed using Jain’s index, the Gini coefficient, and per-class waiting times. A publicly reproducible implementation is provided to support validation and future extensions to higher-fidelity quantum facility and workload models.