Physics-Informed Neural Networks for Predicting Liquid Dairy Manure Temperature During Storage

Temperature is a critical factor that impacts the microbial, physical, and chemical processes associated with aerial emissions from stored manure. The aerial emissions can pollute the environment and degrade the quality of manure as a fertilizer. Further, temperature is commonly used as an input parameter for on-farm decision support tools and nutrient accounting models to assess sustainability. Unfortunately, there is no standard method to estimate manure temperature during storage. This study introduces a physics-informed neural networks (PINN) framework as a robust and alternative method to predict manure temperature during storage. The performance of the PINN is compared to a finite-elements heat transfer model, a data-driven neural network, and a simulation-study trained neural network. The study data was collected from three on-farm manure storages: a clay pit receiving scraped manure, a concrete tank accepting scraped manure, and a concrete tank receiving flushed manure. The four models were trained using 67% of the data collected from the concrete tank storage receiving scraped manure. Each model was tested on three different unobserved datasets (unobserved 33% data from the concrete tank receiving scraped manure, data from the clay pit, and concrete tank receiving flushed manure). Our study found that the PINN model performed better and was more accurate in generalizing manure temperature prediction compared to purely data-driven neural networks and purely physics-based models. Thus, if used in decision support tools and nutrient cycling models, the PINN model predicted temperatures could improve their accuracy in assessing sustainable manure management practices