Generalizable Long Short-Term Memory Models for Beef Cattle DMI Under Grazing1

Lay Summary Beef producers need practical ways to monitor how much each animal eats—both in feedlots and on pasture—because intake drives growth, efficiency, and cost. Measuring intake directly in large groups is expensive and hard, especially on pasture. In this work, we use passive technologies that farms already deploy—walk-on scales, waterers with meters, and simple animal ID tags—combined with open weather data to predict each animal’s dry matter intake (DMI) day by day. We train a sequence model (an LSTM) that “reads” the last seven days of an animal’s weight change, water use, and weather and then predicts the next day’s intake. We evaluate a single full model and a series of lightweight “iterative” updates that progressively add grazing data and adjust only a small prediction layer (“head”), with an optional linear scale-and-shift (“affine calibration”) after prediction. Across all evaluations, the LSTM predicted DMI with lower error than widely used equations and remained accurate in grazing. In the iterative tests, small bias-only adjustments to the head were stable across years, and a simple linear calibration corrected the final iteration’s scale drift on the unseen grazing split. Because the inputs are deployable (no cameras, no proprietary sensors) and predictions are accurate at the individual level, this approach can help producers monitor animals in the systems where they actually live—grazing—supporting selection and management decisions that reduce feed use and environmental footprint.