[Paper] Phythesis: Physics-Guided Evolutionary Scene Synthesis for Energy-Efficient Data Center Design via LLMs

Source: arXiv - 2512.10611v1

Overview

The paper introduces Phythesis, a framework that combines large language models (LLMs) with physics‑guided evolutionary optimization to automatically generate three‑dimensional, simulation‑ready (SimReady) layouts for data‑center (DC) design. By embedding physical constraints directly into the generative loop, Phythesis produces energy‑efficient configurations that outperform pure‑LLM approaches, addressing a critical bottleneck in scaling modern DC infrastructure.

Key Contributions

• Bi‑level optimization architecture that alternates between LLM‑driven layout generation and physics‑informed parameter tuning.
• Self‑criticizing LLM loop: the language model evaluates its own proposals, iteratively refining topology to satisfy spatial and operational constraints.
• Physics‑guided evolutionary search for asset selection (e.g., racks, cooling units) and parameter optimization (e.g., airflow rates, power distribution).
• Quantitative gains: 57.3 % higher successful generation rate and an 11.5 % improvement in Power Usage Effectiveness (PUE) versus a baseline LLM‑only generator.
• Scalable pipeline demonstrated on three different generation scales (small, medium, large), showing robustness across varying DC sizes.

Methodology

1. LLM‑Driven Layout Generation

   • The LLM receives a high‑level specification (e.g., total floor area, cooling budget, rack density).
   • It produces a textual description of a 3‑D scene, which is parsed into a geometric representation (positions, orientations of racks, aisles, CRAC units, etc.).
   • A self‑critique step prompts the LLM to evaluate the draft against a checklist of physical rules (clearance, weight limits, airflow paths) and suggest corrections.

2. Physics‑Guided Evolutionary Optimization

   • The parsed layout becomes the seed for an evolutionary algorithm (EA).
   • The EA mutates asset parameters (e.g., fan speeds, power distribution settings) and swaps component types while evaluating each candidate with a fast physics simulator that computes thermal distribution, airflow, and power consumption.
   • Fitness combines PUE, constraint satisfaction, and a similarity score to the LLM’s original intent, steering the EA toward physically viable, energy‑efficient solutions.

3. Iterative Bi‑Level Loop

   • After the EA converges, the refined layout is fed back to the LLM, which may adjust high‑level topology (e.g., re‑arrange aisles) based on the EA’s feedback.
   • The loop repeats until convergence criteria (stable PUE, no constraint violations) are met, yielding a final SimReady model ready for detailed CFD or CFD‑thermal simulation.

Results & Findings

*Success = layout passes all hard physical constraints and is exportable to industry‑standard simulators.

The experiments across three scales (≈ 1 kW, 10 kW, 100 kW rack deployments) show that Phythesis consistently reduces cooling overhead while preserving required compute density. The modest increase in runtime is offset by the elimination of manual redesign cycles.

Practical Implications

• Accelerated DC Planning: Architects can feed high‑level requirements into Phythesis and receive a validated, simulation‑ready layout within minutes, cutting weeks off the traditional design cycle.
• Energy Cost Savings: An 11.5 % PUE improvement translates directly into lower electricity bills and carbon footprint—critical for hyperscale operators under sustainability mandates.
• Plug‑and‑Play Integration: The output is compatible with existing CFD, BIM, and DCIM tools, allowing seamless hand‑off to detailed engineering teams.
• Rapid “What‑If” Exploration: Developers can iterate over different cooling technologies (liquid vs. air), rack densities, or floor‑plan constraints without hand‑crafting each scenario.
• Extensible to Other Facilities: The bi‑level approach can be adapted for warehouses, labs, or edge‑computing pods where physics‑driven layout matters.

Limitations & Future Work

• Simulation Fidelity vs. Speed: The physics engine used in the EA is a simplified thermal model; higher‑fidelity CFD would increase runtime and may expose new constraint violations.
• LLM Hallucination: Although self‑critique reduces nonsense, the LLM can still propose infeasible component types or dimensions that require manual curation.
• Scalability to Ultra‑Large DCs: Experiments capped at ~100 kW rack clusters; scaling to multi‑MW campuses may need hierarchical decomposition.
• Future Directions: Incorporating reinforcement learning to replace the EA, integrating real‑world sensor data for closed‑loop optimization, and extending the framework to co‑optimize electrical and mechanical infrastructure jointly.

Authors

• Minghao LI
• Ruihang Wang
• Rui Tan
• Yonggang Wen

Paper Information

• arXiv ID: 2512.10611v1
• Categories: cs.AI, cs.NE
• Published: December 11, 2025
• PDF: Download PDF