Application of statistical and artificial intelligence models to estimate in vitro ruminal methane production in tropical forage groups

Introduction Screening tropical forages for low ruminal methane (CH4) production is essential for reducing livestock emissions and supporting global methane mitigation efforts such as the Global Methane Pledge. Although in vitro fermentation systems offer a practical alternative to in vivo measurements, large-scale forage evaluation remains resource-intensive. In this context, statistical and artificial intelligence (AI) modeling have emerged as promising tools for high-throughput screening, although their predictive performance requires further evaluation. This study analyzed the use of statistical and AI models to estimate in vitro ruminal methane production of tropical forages groups. Methods A total of 395 forage accessions representing tropical Fabaceae (n = 238) and Poaceae (n = 157) were incubated for 24 hours in an in vitro fermentation system to measure gas production (GP), methane in the GP, true dry matter digestibility (DMd) and the partition factor (PF). Chemical composition analyses determined organic matter (OM), crude protein (CP), neutral detergent fiber (NDF), acid detergent fiber (ADF), and ash. From NDF and ADF values, hemicellulose (HEM) was determined. Fermentation and chemical composition variables were used to predict in vitro methane production. Models were developed for botanical families (Fabaceae and Poaceae), functional groups (climber, herbaceous legumes, shrub-tree) and for the full dataset. Their performance was assessed using repeated 5-fold cross-validation with five repetitions. Results Model performance differed among the forage groups evaluated (RMSE = 1.61 – 5.03; R 2 = 0.76 – 0.88). Forage-group-specific multivariate models, integrating fermentation characteristics and selected chemical composition variables (eg., GP, DMd, and CP), achieved the best predictive performance when implemented using statistical and AI algorithms, particularly support vector regression with a linear kernel, followed by multiple regression. In some groups (e.g., Poaceae, climber and shrub-tree) models perform satisfactorily for analytical purposes, although further validation with independent datasets is needed. The results highlight the potential of statistical and AI-based models for estimating methane production from fermentation and compositional variables. These models provide a practical complementary tool when direct methane measurements are not feasible due to cost or analytical infrastructure limitations. Their application could expand methane assessment in forage research, particularly in resource-limited regions.