2 Lecture II: From empirical to mechanistic to AI models, making livestock production more efficient.

Abstract With the arrival of the 4th Agriculture Revolution (Agriculture 4.0) and its envisioned fusion of animal production systems with emerging digital technologies and automation, new opportunities have arisen to increase the efficiency and precision with which we feed animals. The barriers of the past have been a lack of automation and efficient means of data capture given the multitude of variables influencing performance metrics and efficiency within a commercial system, as well as appropriate analytical tools with which to drive decision support and system optimization in such a complex environment. The objective of this talk is to review new areas digital tools are being applied to improve efficiency and ponder the complexity of model required to solve problems and implement solutions. As two case-studies, recent research has focused on (1) the prediction of pellet quality, an important Key Performance Indicator in commercial feed manufacturing, and on (2) the prediction of feed intake of dairy cows. In commercial feed manufacturing, pellet quality has substantial impacts on both mill and downstream animal efficiency. Pellet quality is the result of a multitude of factors including ingredient selection and nutrient level, feed manufacturing conditions, as well as the outdoor conditions, leading to difficulty predicting PDI in a commercial setting. This research group has implemented, for the first time, machine learning (ML) based predictions of pellet quality and mill efficiency based on the automated capture of manufacturing conditions, environmental conditions and formulation data across multiple mills. However, of interest is the similarity in model evaluation metrics (e.g. concordance correlation coefficient, CCC) between ML and simpler empirical multivariable approaches. Limits in predictive ability may have more to do with the relative importance of collected driving variables, as opposed to their volume and algorithm choice, and some key drivers of PDI variation are still not routinely measured, let alone recorded (e.g. particle size) commercially. Feed intake is also particularly challenging to predict, given the multitude of factors that impact it and large variations within day, between days, and between animals. The application of ML approaches to predict DMI from cow-level data lead to similar CCC values as common empirical equations when the parameters were refitted in a similar way to the ML model building process (5-fold cross validation, 20% of data held back for independent evaluation). Hybridization of modelling approaches, via providing the empirical models as additional features for the ML models to learn from, improved predictions slightly. Hybridization with more complex models may yield improvements to field predictions. As we explore the Agriculture 4.0 frontier, efficiency gains will require not only the identification of the correct tool for the problem, but also the digitized state and capacity of the end user to implement the solution.