Agentic AI-Enabled Framework for Thermal Comfort and Building Energy Assessment in Tropical Urban Neighborhoods

This paper proposes an agentic AI framework that integrates large language models with lightweight physics-based models to autonomously evaluate and optimize thermal comfort and building energy performance in tropical urban environments through rapid, closed-loop reasoning and action.

The Gist

Imagine you are the mayor of a very hot, crowded city (like Singapore). You want to make the streets cooler for people walking around and save money on air conditioning for the buildings. But you have a problem: the city is huge, the weather is complicated, and running the computer simulations to figure out the best solutions usually takes a team of experts weeks of work.

This paper introduces a new "Smart City Assistant" (an Agentic AI framework) that acts like a super-smart project manager who can talk to you in plain English, do the heavy math, and give you a clear plan in minutes.

Here is how it works, broken down with simple analogies:

1. The Problem: The "Too Many Tools" Mess

Usually, to check if a street is too hot, you need one expert to draw the 3D map, another to calculate the wind, a third to figure out the sun's path, and a fourth to guess how much electricity the buildings will use. It's like trying to build a house where the plumber, electrician, and carpenter all speak different languages and refuse to talk to each other. It's slow, expensive, and prone to mistakes.

2. The Solution: The "AI Project Manager"

The authors built a system where a Large Language Model (LLM) acts as the "Project Manager."

• You speak: You type a simple request like, "Show me where the hottest spots are in this neighborhood and tell me how to cool them down."
• The AI listens: Instead of just chatting, the AI understands you are asking for a physics simulation. It knows exactly which tools to grab.
• The AI acts: It automatically sets up the 3D map, pulls in the weather data, runs the wind and sun calculations, and checks the energy usage—all without you touching a single complex setting.

3. How It Keeps Things Accurate: The "Strict Chef"

You might worry, "Can a chatbot really do complex physics?"
The answer is: No, but it doesn't have to.
Think of the AI as a Head Chef and the physics models as the Specialized Cooks.

• The Head Chef (AI) decides what to cook (the strategy) and when to cook it.
• The Specialized Cooks (Lightweight Physics Models) actually do the chopping and frying. They are fast, simple, and follow strict rules.
• Crucially: The Head Chef isn't allowed to guess the ingredients. If the recipe calls for "wind speed," the AI must get that number from a trusted weather station (like a reliable supplier), not make it up. This ensures the results are scientifically real, not just a hallucination.

4. The Magic Trick: Finding the "Hidden Trap"

The most exciting part of the paper is what happened when they tested a common idea: "Let's paint everything white to reflect the sun!"

• The Expectation: The AI suggested painting roofs and walls white (high "albedo").
• The Result for Buildings: Great! The buildings stayed cooler, and air conditioning costs dropped by about 10%.
• The Result for People: Disaster! The white ground reflected so much sunlight that it actually made the air feel hotter for pedestrians. It's like standing under a mirror in the sun; the mirror reflects the heat right back at you.
• The AI's Insight: The AI caught this "Albedo Penalty." It realized that while white roofs are good for the building, white sidewalks in a sunny plaza are bad for people.
• The Fix: The AI changed its advice: "Paint the roofs white to save energy, but keep the sidewalks darker or add shade trees to protect the walkers."

5. Why This Matters

This system is like having a crystal ball for city planners.

• Speed: It turns weeks of work into minutes.
• Clarity: It doesn't just give you a spreadsheet of numbers; it gives you a report that says, "Here is the hot spot, here is why it's hot, and here is exactly what to do."
• Safety: It checks its own work. If a strategy helps the building but hurts the people, it flags the conflict and suggests a better compromise.

In a Nutshell

This paper describes a new way to design cities. Instead of hiring a team of experts to run slow, complex simulations, you can now talk to an AI "City Architect." You tell it your goal, and it automatically runs the physics, finds the hidden traps (like the white sidewalk problem), and gives you a clear, actionable plan to make your city cooler and more energy-efficient. It's the difference between trying to fix a car engine with a hammer versus having a robot mechanic that diagnoses the problem and fixes it for you.

Technical Summary

1. Problem Statement

Urban planners in tropical cities like Singapore face significant challenges due to the Urban Heat Island (UHI) effect and high building cooling demands. Key practical questions include identifying pedestrian heat-stress hotspots, determining which building clusters drive high cooling loads, and evaluating the trade-offs of surface strategies (e.g., cool paints, green façades).

Current solutions, such as the "Cooling Singapore 2.0" initiative, rely on integrated climate and building energy models. However, running these studies at the neighborhood scale is laborious and time-consuming because:

• Workflows span multiple disparate tools.
• Model preparation requires extensive manual parameterization.
• There is a lack of seamless integration between natural language design intent and rigorous physics-based simulation outputs.

2. Methodology

The authors propose an Agentic AI-enabled framework that integrates Large Language Models (LLMs) with lightweight physics-based models to create a closed-loop reasoning-action process. The system is built on a LangGraph-style state machine that orchestrates five distinct stages:

A. Core Architecture & Workflow

1. Intent Analysis: An LLM (GPT-5.1) interprets user queries in plain language to propose necessary inputs (geometry, weather) and tools.
2. Geometry Analysis: The system loads Stereolithography (STL) files, cleans meshes, extracts building footprints, assigns unique IDs, and generates an index map to trace simulation results back to specific buildings.
3. Solver Orchestration: An "Orchestrator Agent" merges parameters from a strict five-level priority hierarchy to ensure input governance:
   • Priority 1: User overrides.
   • Priority 2: Real-time weather services (e.g., NEA Singapore).
   • Priority 3: Climate databases (e.g., IWEC).
   • Priority 4: LLM suggestions (only if data is missing).
   • Priority 5: Configuration defaults (JSON).
   • Result: A structured, auditable JSON input file is generated.
4. Material Strategy Recommendation: The agent executes a "baseline-then-refined" loop, proposing material adjustments (albedo, emissivity) and comparing before/after metrics.
5. Integration & Reporting: The LLM synthesizes solver metrics (cooling load, PET, wind) into a human-readable report with actionable design recommendations, linking every claim to specific output files.

B. Lightweight Physics-Based Models

To balance speed and accuracy, the framework uses simplified physics models rather than full-scale, computationally expensive solvers for every iteration:

• CFD (Computational Fluid Dynamics): Uses a 2D potential flow model executed at multiple vertical slices (pseudo-3D) to estimate wind velocity fields. It applies logarithmic wind profiles and empirical adjustments for temperature/humidity based on local wind speeds.
• Radiation & Surface Energy Balance: Solves a transient energy balance equation for surface temperatures, accounting for shortwave/longwave radiation and convective heat transfer. It calculates Mean Radiant Temperature (MRT) at pedestrian levels.
• Thermal Comfort & Energy:
  • Comfort: Calculates Physiological Equivalent Temperature (PET) using the Munich Energy-balance Model for Individuals (MEMI).
  • Energy: Estimates building cooling load using a simplified 1D Conduction Transfer Function (CTF) model, aggregating heat fluxes to determine Energy Use Intensity (EUI).

3. Key Contributions

• Agentic Workflow for Urban Design: A novel framework where the LLM acts as a "workflow layer," translating semantic intent into rigorous physics simulations without replacing the solver.
• Input Governance & Traceability: A strict parameter resolution hierarchy ensures that LLMs are advisory, not authoritative, regarding physical boundary conditions. The system logs all prompts, responses, and parameter provenance for full auditability.
• Closed-Loop Intervention: The system autonomously identifies hotspots, proposes mitigation strategies (e.g., material changes), re-simulates, and refines recommendations based on feedback.
• Discovery of Non-Linear Trade-offs: The framework successfully identified the "Albedo Penalty," demonstrating that while high-albedo materials reduce building cooling loads, they can increase pedestrian thermal stress due to reflected shortwave radiation.

4. Results and Analysis

The framework was validated in a high-density tropical district (Singapore) through two scenarios:

Scenario I: Autonomous Baseline Audit

• Task: The agent interpreted a vague prompt ("late March-May inter-monsoon") to set up a coupled CFD-solar simulation.
• Outcome: Successfully identified specific thermal stress hotspots (PET > 52°C) and energy-intensive building clusters.
• Insight: The agent correctly linked high PET values to stagnant airflow and high sky view factors, and high cooling loads to unshaded façade orientations. It generated location-specific recommendations (e.g., overhead shading for specific plazas).

Scenario II: Closed-Loop Material Intervention

• Task: The agent applied high-albedo coatings to roofs, walls, and ground surfaces to test mitigation.
• Outcome:
  • Building Energy: Confirmed a ~9–11% reduction in daily cooling loads for high-rise towers.
  • Pedestrian Comfort: Detected a ~1°C increase in PET at ground level in sun-exposed plazas.
• Mechanism: The agent identified the physical cause: while surface temperatures dropped (reducing longwave emission), the increased reflectivity of the ground significantly raised the shortwave radiative load on pedestrians.
• Refined Strategy: The agent self-corrected its recommendation, advising high-albedo roofs (for energy savings) but low-reflectivity pavements in plazas (to protect pedestrian comfort).

5. Significance

• Efficiency: Drastically reduces the time and expertise required to run integrated comfort-energy studies, making neighborhood-scale analysis accessible to non-specialists.
• Scientific Rigor: Maintains scientific determinism by keeping the LLM separate from the physics engine, ensuring that boundary conditions are derived from authoritative sources.
• Cross-Disciplinary Utility: Facilitates evidence-based urban design by bridging the gap between qualitative design intent and quantitative performance metrics.
• Scalability: The backend-agnostic design allows the framework to scale from lightweight exploratory models to high-fidelity solvers or AI surrogates in the future.

This study demonstrates that Agentic AI can effectively orchestrate complex urban simulation workflows, enabling rapid, traceable, and physically grounded decision-making for climate-resilient urban planning.