Apppublic

madhuishere-123/datacenter-cooling-env

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

Data Center Cooling Environment

OpenEnv-compatible RL environment for AI-driven data center thermal management Meta x Scaler OpenEnv Hackathon 2026

Overview

This environment simulates a realistic multi-zone data center cooling system where an AI agent must balance thermal safety, energy efficiency, carbon awareness, and equipment health while interpreting natural language incident reports that require planning and reasoning.

Why This Requires an LLM

A PID controller can react to current sensor readings. But this environment generates natural language incident reports that demand reasoning:

—Load spike predictions: pre-cool zones before ML training jobs deploy
—Weather forecasts: gradually ramp cooling before heat waves
—Equipment degradation warnings: shift cooling strategies before failures
—Carbon grid updates: time aggressive cooling to green energy windows
—Cascade risk alerts: prioritize overheating zones to prevent chain failures

These incidents actually affect the simulation - scheduled events trigger real load spikes, weather shifts, and equipment degradation.

Action Space

—fan_speeds: list of floats 0.0-1.0 per zone
—chiller_setpoints: list of floats 15.0-25.0 per zone

Observation Space

Per zone: temperature, serverload, fanspeed, chillersetpoint, equipmenthealth, status. Global: outdoortemp, pue, gridcarbon_intensity, incidents (natural language), timestep.

Tasks

Task	Zones	Outdoor	Steps	Challenge
task_easy	1	25C	100	Stable conditions
task_medium	3	32C	150	Variable load, sensor noise
task_hard	5	40C	200	Heat wave, cascading failures
task_critical	4	42C	150	Zone 0 at 30C, 60% power budget

Reward Function

—Temperature safety (30%) - zones within 18-27C
—PUE efficiency (25%) - lower power usage
—Carbon awareness (18%) - reduce fans when carbon high
—Equipment health (10%)
—Action stability (10%) - smooth transitions
—Temperature diversity (7%) - target 22C

Baseline Scores

Task	Score
task_easy	0.876
task_medium	0.822
task_hard	0.730
task_critical	0.754
Mean	0.795

Model: meta-llama/Llama-3.3-70B-Instruct

API Endpoints

—GET /health - Health check
—POST /reset - Reset with task_id and seed
—POST /step - Take action
—GET /state - Current state
—GET /tasks - List tasks

Setup

ash export API_BASE_URL="https://router.huggingface.co/v1" export HF_TOKEN="your-token" export MODEL_NAME="meta-llama/Llama-3.3-70B-Instruct" export ENV_BASE_URL="https://madhuishere-123-datacenter-cooling-env.hf.space" python inference.py

Tech Stack

Python 3.11, FastAPI, Uvicorn, Pydantic v2, NumPy, OpenAI SDK, Docker

Authors

Madhumita SM and Anirudh Kumar R - SASTRA Deemed University