Application of machine-learned metadata-driven model for dairy barn ventilation simulation

Abstract

As US dairy operations consolidate to meet an increasing demand for dairy products, dairy producers have begun to seek more economically efficient dairy-barn designs and ventilation systems to combat the pernicious effects of heat stress. The metabolic changes that heat stress induces not only put the animals at risk of serious illness but also diminish the herd’s productivity. Lactation, being a metabolic process, generates a great deal of body heat thereby increasing the risk of heat stress during warm weather, and that risk increases when the animals commingle in an enclosed space. For this reason, predicting heat stress and estimating its severity have taken on special importance. Recently, producers and barn builders have contended with the conditions that promote heat stress by implementing closed mechanical ventilation systems. However, if a proposed ventilation system’s performance was not accurately predicted, the system would not work effectively or efficiently. The traditional ways of assessing ventilation performance, which involved either hand-held sensors or stationery sensing monitors, tended to be time-consuming and the larger the barn the more labor-intensive. Additionally, the complex microenvironmental system that exists inside a dairy barn is further complicated by the myriad moving elements, both animals and operational machines. Given these complicating parameters, researchers have lately turned to computation fluid dynamics (CFD) when analyzing such environments because, once properly validated, CFD can accurately predict the heat and mass transfer occurring at any location inside a computer-generated model of a dairy barn. However, operating CFD correctly requires not only an extensive background knowledge and a significant amount of hands-on experience, but also high-end hardware, which is often unavailable to dairy operators and non-CFD--experts. This study proposed CFD-ML, a machine-learning (ML) based “simulator,” as a means to achieving computational predictions that are accurate enough to use as a basis for dairy-barn design. A convolutional neural network (CNN), a representative ML model, allows complex mapping between input barn geometry and the velocity and temperature fields occurring in the barn, resulting in faster prediction time and simpler usage processes. As a result, the CFD-ML model can achieve outcomes comparable to those obtained using CFD, with R2 values greater than 0.85, in far less computing time required by CFD. This finding should provide a foundation that could facilitate future CFD-ML advances in dairy science research as well as the development of a user-friendly Graphical User Interface (GUI) that bypasses the complex CFD procedures.