Predicting individual water intake in beef cattle using longitudinal data and long short-term memory models

Water intake (WI) remains an under-characterized yet essential trait in beef cattle systems, with implications for animal health, climate resilience, and resource efficiency. Existing predictive models, including those from the National Academies of Sciences, Engineering, and Medicine (NASEM), rely on static equations derived from outdated data and lack resolution at the individual animal level. In this study, we developed and evaluated a long short-term memory (LSTM) model to predict daily WI using longitudinal data on animal characteristics, dry matter intake (DMI), and engineered environmental features. Data were collected from 2,268 animals across drylot and grazing systems between 2019 and 2024, with environmental variables sourced from NOAA and NASA. Feature engineering captured biologically relevant dynamics via rolling deltas, interaction terms, and temporal encodings. The LSTM model trained with these engineered features achieved strong predictive performance (root mean square error [RMSE] = 3.85 L/d; R2 = 0.74; P < 0.001) and generalized well to unseen regional (RMSE = 4.20 L/d; R2 = 0.61; P < 0.001) and non-regional (RMSE = 4.19 L/d; R2 = 0.63; P < 0.001) drylot datasets. In contrast, models trained without engineered features failed to generalize (RMSE = 22.7 L/d; R2 = -0.96), and NASEM predictions systematically underestimated high intake values. Permutation-based feature importance analysis highlighted the value of short-term environmental stress indicators, particularly temperature-humidity index (THI) and temperature deltas. These results demonstrate that sequence-based models incorporating dynamic environmental covariates can significantly improve WI prediction in beef cattle and provide a scalable decision-support framework for water-efficient genetic selection and adaptive management.