Simulation for the HVACR Industry: How to Leverage Machine Learning

The HVACR Industry is Evolving

As we step into 2025, major trends in systems simulation are emerging. The heating, ventilation, air conditioning, and refrigeration (HVACR) industry is seeing accelerated growth in the adoption of simulation throughout the design and development process. This industry is looking to further modernize, appeal to consumers, and demand energy-efficient, sustainable and smart solutions. HVACR original equipment manufacturers (OEMs) are addressing these requests by implementing engineering processes that are supported by machine learning, Internet-of-Things (IoT), and AI advancements. 

How Systems Simulation Modeling & Machine Learning can Assist

The HVACR industry is being driven by new regulations that have been enacted to combat climate change as well as to increase efficiency of current and future vapor compression systems. These trends include: 

1. A big emphasis on using low global warming potential (GWP) refrigerants 
2. Heat pumps becoming more popular within the industry

To achieve practical viability and adoption of heat pumps and low GWP refrigerants in HVAC systems within the next decade, thousands of prototypes and experiments need to be performed. Fast, accurate and robust modeling and simulation can also aid in this endeavor and accelerate this process. There also needs to be extensive collaboration between industry, academia, and policy makers to comprehensively address these goals. An important tool available to industry, academia and policy makers is the huge amount of data already available from different sources. This data can be leveraged using machine learning tools to aid and enhance modeling as well as providing new physical insight into HVAC systems.

Applying Machine Learning to Thermal Model Simulations 

We at Gamma Technologies presented a paper at the Purdue Herrick Conferences on machine learning under the model speedup umbrella titled, ‘Application of Feedforward Neural Networks to Simulate Battery Electric Vehicle Air Conditioning Systems.’

In this paper a representational model of the thermal systems of a battery electric vehicle (BEV) was built in GT-SUITE and transient drive cycle simulations were carried out to compare the speed and accuracy of a machine learning based metamodel compared to a physics-based solution. The air-conditioning circuit in the EV thermal model was replaced with a feed-forward neural network trained against physical data. The battery and cabin temperatures were tracked and compared during a heat-up (ambient temperature -10 C) and cooldown (ambient temperature 30 C) cycle.  

We observed that the machine learning metamodel does a good job in capturing the battery and cabin temperatures but there is some mismatch between the evaporator and condenser heat transfer rates during the heat-up simulation. However, the speed of the metamodel is around 35% faster than the physics-based model with an RT factor of 0.17 compared to the already faster than real time physics-based solution which has an RT factor of 0.37.  

With this work we were able to show that we can successfully integrate machine learning based metamodels into GT system models using in-built tools and use them as alternatives to physics-based solutions to address various simulation needs.  

Figure: EV thermal system model with physics-based solution (red) and feedforward neural net (green)

Link to Presentation

Learn More About our HVACR Simulation Solutions 

See how GT-SUITE’s simulations solutions impact the HVACR industry here. 

GT-SUITE doesn’t take hours or days to run complex simulations. HVACR engineers can run simulations in a matter of seconds or minutes, which allows for a more iterative design cycle in addition to the ability to test more product possibilities. 

Firms such as Trane have successfully unleashed the power of GT-SUITE and simulation on HVACR scroll compressor designs. Learn how a simple sensitivity analysis enabled Trane engineers to solve a challenging performance shortfall problem that resulted in an estimated savings of between $50,000 and $40 million in product development costs.