To effectively minimize nutrient wastage, it's essential to reduce nutrient consumption without hampering animal performance. Precision feeding embodies the current state-of-the-art methodology to accomplish this. Here, nutrient requirements are calculated daily for each individual, and feed composition is adapted to match these requirements. Although precision feeding demonstrates a significant potential to reduce nutrient losses, we foresee that the impact in the field will be hampered by its limited capacity to flexibly adapt to the wide variation in commercial conditions. Many of the input features and parameters (e.g. maximum protein deposition rate, amino acid digestibility coefficients) are fixed values that do not represent the individual and time-dependent variability in animal responses. Second, precision feeding relies on continuous updates from the research community to keep up with the fast-developing production circumstances, where it is impossible to recreate all possible scenarios that impact nutrient efficiency. 

This project will develop a machine learning framework that digests sensor data and metadata, into a robust estimate of the input features and hyperparameters that serve the precision feeding calculations for pigs. Whenever new farm data is available, the framework will be retrained to learn from its own mistakes by comparing the predicted animal performance with the measured performance. By including multiple data streams (climate data, stocking density, feed matrix, etc.), as well as alternative representations of the data (time domain and frequency domain), we significantly augment the capability to capture the complexity of real-life conditions. This will enable training for tailored nutrient requirement calculations that drive the farm-specific feeding programs, as well as identifying farm-specific parameters that need better management, paving the path for continuous improvements in nutrient efficiency in pigs.