Physics-informed neural networks for predicting liquid dairy manure temperature during storage

Rana A. Genedy, Matthias Chung, Jactone A. Ogejo

Abstract: This study presents a physics-informed neural networks (PINN) modeling approach we developed as a practical application of neural computing to predict manure temperature during storage at a dairy farm all year round. Manure temperature is a factor that impacts the microbial and chemical processes associated with releasing aerial pollutants. Therefore, knowing the dynamics and quantifying the manure temperature value during storage is critical for designing methods to mitigate its potential contribution to polluting the environment. Also, manure temperature is a pertinent input parameter for on-farm decision support tools and nutrient accounting models used to assess sustainable manure management practices. However, there is no standard method to estimate it. Currently, decision support tools use surrogates derived from various ambient air temperature averages instead of the manure temperature, which underestimates the contaminants lost to the atmosphere during the manure storage. In this study, we developed a PINN model as an alternative for predicting the stored manure temperature. We compared its performance to three other models (finite-elements heat transfer, classical data-driven neural network, and simulation-based neural network). We collected field data from three dairy farms with different manure management practices and storage structures to train, validate, and test the models. The manure temperatures predicted by the PINN model were the closest in magnitude to the measured temperatures at the three manure storages on the three dairy farms. In addition, the PINN model predictions were less biased and more data-efficient than the other models. The predictive ability of the models was comparable during validation (R2 > 0.90); however, the PINN model had superior generalization accuracy. The R2 for the PINN model during the testing phase exceeded 0.70. In contrast, R2 ranged from − 0.03 and 0.66 for the finite-elements heat transfer, classical data-driven neural network, and simulation-based neural network models. These results suggest that using the PINN model to estimate manure temperature in decision support tools and nutrient cycling models would provide more realistic outcomes for assessing sustainable manure management practices. The outcomes of this study contribute to the field of precision agriculture, specifically designing suitable on-farm strategies to minimize nutrient loss and greenhouse gas emissions during the manure storage periods and improve the accuracy of metrics used to assess sustainable manure management practices.