Sugarcane is a cornerstone of Mexico’s rural economy, supporting over 2 million jobs and contributing 12% of agricultural GDP (Gross Domestic Product). Despite ranking as the world’s sixth-largest producer, the industry faces declining yields due to climate variability, structural challenges, and limited technology adoption among smallholders. To address critical forecasting gaps, we developed a predictive model using public mill data from CONADESUCA (47 mills) and compared machine learning approaches: Long Short-Term Memory (LSTM), Random Forest, and Dense Neural Networks. LSTM demonstrated superior performance in capturing production trends, with key mills achieving RMSE values as low as 0.153 and \(R^2\) up to 0.981. While climate variables provided marginal improvements, results suggest operational factors may dominate yield variability. The study highlights both the potential of AI-driven forecasting for mill planning and persistent data limitations—particularly the lack of field-level records from small producers.