PS6-3. Enhancing Methane Emissions Modeling Through Feature Engineering and Machine Learning.

Abstract Enteric methane produced by beef cattle is a major greenhouse gas and represents a loss of feed energy, making accurate prediction of methane emissions an important step toward developing effective mitigation strategies and improved production performance. This study evaluated the effectiveness of predictive models for estimating methane emissions from ruminant data by comparing tree-based machine learning models with traditional regression approaches. A dataset compiled from multiple experimental studies on enteric methane emissions in beef cattle was used (895 entries), that contained variables such as dry matter intake (DMI), dietary composition such as crude protein, neutral detergent fiber (NDF), and acid detergent fiber (ADF), treatment descriptions, methane emissions expressed in g/day or g/kg DMI (methane yield), along with measurement methodologies including respiration chambers and hood/headbox systems. Seven predictive algorithms were trained and evaluated to predict methane yield: Linear Regression, Ridge Regression, Lasso Regression, Decision Tree Regressor, Random Forest Regressor, Gradient Boosting Regressor, and Extreme Gradient Boosting (XGBoost). Results showed that tree-based ensemble models consistently outperformed traditional linear models across multiple evaluation metrics. Among these, XGBoost demonstrated the strongest predictive performance, achieving a Concordance Correlation Coefficient (CCC) of 0.77 and the lowest mean absolute error (MAE) of 2.65 g/kg DMI, compared to Linear and Regularized regression models (CCC< 0.57). Model interpretability using SHAP (Shapley Additive Explanations) analysis further revealed that the ensemble models effectively identified key biological drivers of methane production, highlighting Dietary Fiber and DMI as the most influential predictors. Overall, the findings indicate that tree-based machine learning approaches, particularly XGBoost, provide a more accurate and flexible framework for predicting enteric methane yield than conventional statistical methods, as they can capture high-dimensional interactions among dietary variables while maintaining interpretability aligned with established ruminant physiology. These results also demonstrate the potential for combining recursive feature selection to aid interpretable machine learning models, that can be used in low resource systems to support the development of precise and data-driven methane mitigation strategies in beef production systems.