ADCTM
Autonomous Data Centre Thermal Management
A physics-grounded simulation and control framework for evaluating autonomous thermal-management strategies under dynamic computational workloads, thermal disturbances, and hardware degradation.
Interactive Thermal Simulation
System Architecture
The closed-loop interaction between dynamic computational heat generation, physical multi-zone thermal diffusion, real-time telemetry monitoring, autonomous controller decision, and multi-objective performance evaluation.
Continuous CRAH fan RPM & chilled-water valve extraction.
4-part composite score + 10 industrial metrics feedback.
Autonomous Controller
Unified BaseController interface dispatching continuous per-zone cooling actions in [0.0, 1.0].
Thermodynamic Physics Engine
ADCTM models multi-zone rack cooling using a physics-grounded discrete-time thermal ordinary differential equation (ODE), incorporating heat generation, active extraction, ambient dissipation, and inter-zone conduction.
Server computation drives temperature rises linearly with processor utilization.
CRAH fan modulation and chilled water loops actively extract thermal energy from server racks.
Passive Newtonian convective thermal exchange with facility hall ambient air.
Conduction between adjacent rack zones governed by a discrete 1D spatial Laplacian.
Canonical Benchmark Dashboard
Deterministic evaluation on Task: Medium (5 zones, 24 steps, seed 202, chiller degradation at step 10) under the exact same computational workload sequence.
Thermal Quality vs. Actuation Energy Expenditure
Prioritizes thermal safety margin above all else. Drastically eliminates undercooling (-95.8%) and cuts response time to 2.0 steps, but expends 57.5% more cooling energy and produces 1,386°C·steps overcooling.
Zero thermal undercooling combined with minimal cooling actuation energy. This is the open research objective: optimizing thermal protection per unit of cooling energy.
Zero active energy expended (0.00), but leads to immediate critical thermal failure: 23 safety violations, 105.0°C peak temperature ceiling, and 3,653°C·steps of undercooling.
Moderate cooling energy expenditure (63.60 – 68.81) and minimal overcooling (1.10 – 10.18), but slower deadband convergence (6.2 – 21.8 steps) resulting in moderate sustained undercooling.
| Industrial Metric | Rule-Based | PID Controller | PPO (RL Agent) | Zero (Passive) |
|---|---|---|---|---|
| Mean Temperature | 64.86°C | 62.58°C | 49.60°C | 91.44°C |
| Peak Temperature | 68.78°C | 65.96°C | 65.23°C | 105.00°C |
| Thermal Headroom | +5.22°C | +8.04°C | +8.77°C | -31.00°C |
| Safety Violations | 0 steps | 0 steps | 0 steps | 23 steps |
| Cooling Energy | 63.60 | 68.81 | 100.16 (+57.5%) | 0.00 |
| Undercooling (°C·steps) | 463.92 | 199.50 | 19.48 (-95.8%) | 3653.03 |
| Overcooling (°C·steps) | 1.10 | 10.18 | 1386.92 | 0.00 |
| Response Time (steps) | 21.8 | 6.2 | 2.0 (-90.8%) | Failed (24.0) |
What Did the Controller Actually Learn?
Interpreting reinforcement learning benchmarks requires intellectual honesty. A lower mean temperature is not an unqualified victory in industrial facility management.
Rapid Convergence to Safe Deadband
PPO reaches the target thermal deadband in 2.0 steps, compared to 21.8 steps for the rule-based heuristic and 6.2 steps for PID. By proactively saturating cooling actuators immediately upon workload spikes, it reduces total undercooling by 95.8% (19.48 vs. 463.92 °C·steps).
The Cost of Aggressive Overcooling
Because the reward function penalizes safety violations heavily, PPO learned that being substantially too cold (mean 49.60°C) is safer than approaching the 74.0°C threshold. Consequently, it consumes 57.5% more cooling energy (100.16 vs 63.60) and incurs 1,386.92 °C·steps of overcooling.
Chiller Degradation Robustness
When the step 10 hardware fault halves Zone 03 chiller effectiveness, the baseline rule-based controller exhibits a thermal transient spike to 68.78°C. PPO, already maintaining high thermal headroom (+8.77°C), absorbs the 50% capacity loss with zero safety boundary violations.
"The next optimization frontier is not simply better thermal control. It is better thermal control per unit of cooling energy."
ADCTM demonstrates that naive reinforcement learning agents readily discover safety-first policies by aggressively overcooling. Achieving industrial sustainability requires multi-objective Pareto optimization, asymmetric temperature deadband rewards, and explicit chiller actuator slew-rate constraints.
Production Engineering Stack
Core runtime & physical ODE simulation harness
Vectorized 1D Laplacian thermal diffusion & bounded walks
Strict schema validation for observations, configs & telemetry
Continuous Proximal Policy Optimization (PPO) actor-critic
Asynchronous simulation execution & telemetry bridge
High-density technical instrumentation & SVG data visualization
Static end-to-end interface typing & client safety
Run the Benchmark on Your Own Machine
ADCTM is packaged under src/adctm and includes an automated 91-test pytest regression suite locking down all numerical ground truths.
Inspect the Full Implementation on GitHub
Explore the environment dynamics, controller implementations, benchmark runner, and Docker containerization. Distributed under the MIT License.