Carbon-aware AI model reduces data center emissions while keeping services reliable

Industrial cloud systems could sharply reduce energy use and emissions by combining carbon-aware workload scheduling with deep reinforcement learning, researchers Maraga Alex and Sunday O. Ojo report in a new study that targets one of AI's fastest-growing sustainability problems: the rising environmental cost of large-scale computing.

The study, titled "Model for Green AI and Sustainable Computing: Energy-Efficient Architectures and Carbon-Aware Deployment in Industrial Systems," was published in Computers. It introduces the Hierarchical Clustering Deep Q-Network Carbon-Aware Placement System, or HC-DQNCAPS, a hybrid framework designed to reduce energy consumption, cut carbon emissions, maintain service reliability and improve resource use in industrial cloud and hybrid computing systems.

AI growth puts pressure on energy-hungry industrial cloud systems

The rapid adoption of AI across manufacturing, healthcare, finance, logistics and other sectors is increasing demand for large-scale computing infrastructure. AI systems require major processing capacity for training, data processing and real-time decision-making, placing heavier loads on cloud data centers, hybrid systems and edge environments.

The study argues that this growth has turned energy consumption and carbon emissions into central issues for modern computing. Data centers are already major electricity users, and the spread of AI-enabled industrial applications risks adding further pressure unless computing systems become more adaptive and environmentally aware.

Green AI shifts the focus from performance alone to performance achieved with lower energy use, lower emissions and better operating efficiency. Traditional AI development has often prioritized accuracy and speed, even when that requires large computational resources. Green AI, on the other hand, asks whether models and computing infrastructure can deliver useful results while consuming less power and producing fewer emissions.

This shift is especially important in industrial systems, where workloads are large, continuous and operationally sensitive. A factory, logistics network, hospital system or financial platform may need real-time service, strict uptime and low latency. Any sustainability strategy that reduces energy use but causes service failures would be difficult to deploy. The challenge is therefore not simply to save energy, but to save energy while meeting service-level requirements.

Existing energy-efficiency approaches include virtual machine consolidation, dynamic resource allocation and hardware-level optimization. Carbon-aware computing has also emerged as a way to shift workloads to regions or times when electricity is cleaner. But the study says many existing methods treat energy efficiency, carbon reduction and service reliability as separate problems. That fragmented approach limits their usefulness in complex industrial environments.

The proposed HC-DQNCAPS framework attempts to solve that gap by combining three elements: energy-efficient infrastructure design, carbon-aware workload scheduling and reinforcement learning-based decision-making. The system uses real-time data about server usage, network activity, workload behavior, energy consumption, carbon intensity and service-level status to decide where and when computing tasks should run.

Hybrid AI model combines clustering, carbon data and reinforcement learning

The HC-DQNCAPS framework is built around a layered architecture that collects infrastructure data, processes workload behavior, makes intelligent placement decisions and then executes actions through cloud orchestration and resource controllers.

At the infrastructure level, the system monitors CPU use, memory demand, network bandwidth, storage activity, workload arrival rates, carbon intensity, energy use and service-level status. That information gives the model a live view of both computing demand and environmental conditions.

The processing layer prepares the data for machine learning. It cleans and normalizes incoming metrics before grouping similar workloads through Hierarchical Agglomerative Clustering. This clustering step is important because industrial cloud environments can contain many different workload types, including CPU-heavy tasks, memory-intensive operations, I/O-heavy applications and real-time streaming workloads. Grouping similar workloads reduces decision complexity and allows the system to make more scalable placement decisions.

The decision-making layer is driven by a Deep Q-Network reinforcement learning agent. The agent observes the state of the system and learns which action is most effective under changing workload, energy and carbon conditions. Available actions include allocating resources, migrating workloads, consolidating servers, deferring tasks and scaling resources.

The reinforcement learning model is trained through a reward system that balances three goals: reducing energy consumption, reducing carbon emissions and avoiding service-level violations. This design is crucial because it prevents the system from optimizing one goal at the expense of another. A workload should not simply be moved to a cleaner energy region if that move causes delay, overload or service failure.

The model's mathematical objective combines energy use, carbon emissions and penalties for service-level breaches into one weighted cost function. This gives operators flexibility. A system with strict service requirements can assign more weight to SLA compliance, while a system with stronger climate targets can place greater weight on carbon reduction.

The framework also includes hard constraints for CPU, memory and bandwidth availability. This means workload placement decisions must respect physical infrastructure limits. The goal is not only theoretical optimization, but a practical placement system that can operate within real resource limits.

The researchers tested the model in a simulated multi-cloud industrial computing environment designed to represent geographically distributed data centers with different hardware, energy profiles, network conditions and workload patterns. The simulation included dynamic virtual machine provisioning, workload migration, real-time carbon monitoring, SLA-aware scheduling and variable workload demand.

The HC-DQNCAPS model was compared against several baseline and advanced methods, including First-Come-First-Serve scheduling, Energy-Aware VM Allocation, Carbon-Unaware Reinforcement Learning, Proximal Policy Optimization, Double Deep Q-Network scheduling and Multi-Agent Deep Reinforcement Learning. This comparison allowed the researchers to test whether combining clustering, carbon awareness and reinforcement learning delivered measurable gains over both traditional and advanced scheduling systems.

Results show lower emissions, fewer SLA breaches and higher resource use

The results show that HC-DQNCAPS outperformed the comparison methods across the study's main performance measures. The model reduced energy consumption by about 30% to 35% compared with traditional scheduling approaches. It also cut carbon emissions by about 25% to 30% while keeping service-level violations below 5%.

The strongest energy savings came from combining workload clustering, adaptive virtual machine migration, carbon-aware placement and dynamic resource allocation. Traditional First-Come-First-Serve scheduling used the most energy because it processed workloads sequentially without considering resource efficiency or system-wide optimization. Energy-Aware VM Allocation improved energy use through server consolidation, but lacked the adaptive learning and carbon-aware capabilities of the proposed model.

Carbon emissions followed a similar pattern. First-Come-First-Serve produced the highest emissions because workload placement did not account for environmental conditions. Energy-aware and reinforcement learning methods improved performance, but HC-DQNCAPS performed best because it considered both carbon intensity and energy efficiency when deciding workload placement.

The model also achieved better service reliability. By including service-level penalties directly in the reinforcement learning reward function, HC-DQNCAPS reduced SLA breaches without abandoning sustainability goals. This finding is significant because one common concern about carbon-aware scheduling is that shifting or deferring workloads could degrade service quality. The study's results suggest that intelligent scheduling can reduce emissions while maintaining operational reliability.

Resource utilization also improved. HC-DQNCAPS increased resource use by about 20% compared with baseline methods. The clustering component helped by grouping workloads with similar resource needs, allowing the system to allocate computing capacity more efficiently and reduce idle infrastructure. Better utilization means fewer wasted resources, lower power draw and stronger operational efficiency.

The study reports that the reinforcement learning agent reached stable convergence after about 450 training episodes, suggesting that the model learned effective workload placement policies within the simulation. Statistical testing using ANOVA and Wilcoxon signed-rank tests confirmed that the results were significant at the 95% confidence level.

The researchers argue that the framework's modular design makes it suitable for a range of industrial computing settings, including cloud data centers, hybrid cloud operations and distributed edge computing environments. Its ability to respond to changing workload demand, carbon intensity and infrastructure status could make it useful for organizations trying to meet both operational and sustainability targets.

The study is limited by its use of a simulated environment. Real-world industrial cloud systems involve noisy sensors, heterogeneous hardware, unpredictable network delays, market constraints and complex operational policies that may affect performance. According to the authors, future research should test HC-DQNCAPS on production-scale industrial infrastructure and expand the model to include renewable energy forecasting, carbon intensity forecasting, edge computing and Internet of Things environments.

The next phase could also involve multi-agent reinforcement learning and federated reinforcement learning, allowing distributed cloud-edge systems to optimize workloads without centralizing all decision-making. Carbon pricing and energy market signals could further improve scheduling decisions by linking environmental performance with economic incentives.