Real-Time CFD Model Calibration of an Environmental Chamber Using Deep Reinforcement Learning

Document Type

Conference Proceeding

Publication Date

1-1-2026

Abstract

Improving indoor air quality, and thermal comfort in educational buildings can increase productivity and well-being. However, it is of paramount importance to provide energy efficiency during the building design and control strategies phases. In this regard, Computational Fluid Dynamics (CFD) models are widely used to simulate indoor environments. However, their accuracy depends on the correct tuning of internal parameters and in some scenarios each simulation can take several hours. In this study, a real-time calibration framework was developed to reduce the error between CFD simulations and experimental measurements from an environmental chamber containing a vertical green wall. The chamber was equipped with twelve wireless sensors that continuously recorded temperature and humidity. During the experiment, the temperature was set to 85 °F and the relative humidity was maintained at 40%. In addition, a thermal camera was installed in front of the green wall and captured IR images every second. Next, a custom CFD solver was developed to simulate the chamber. Subsequently, the Proximal Policy Optimization (PPO) reinforcement learning algorithm was utilized to fine-tune the CFD model by adjusting influential internal parameters, aiming to minimize the discrepancy between simulation outputs and measured chamber conditions. A reduced-order model was used during training to reduce computation time. The learning agents were trained to minimize the difference between simulated and measured environmental conditions in real time. Results showed that PPO effectively reduced simulation error and offered more stable learning. Compared to traditional control-based calibration, this learning-based approach significantly improved the alignment between CFD predictions and experimental data. The proposed method demonstrates how real-time feedback from physical systems can be used to enhance simulation accuracy and support more reliable environmental control strategies in buildings that include nature-based solutions.

Publication Source (Journal or Book title)

ASHRAE Transactions

First Page

1233

Last Page

1240

This document is currently not available here.

Share

COinS