Cold manure, hot data: Machine learning prediction of methane production and emissions from psychrophilic dairy manure digestion

• Multi-farm psychrophilic AD dataset generated for Quebec dairy manure. • Region-specific SMY range (0.07–0.136 L CH 4 g −1 VS) defined for design. • Interpretable ML predicts SMY from physicochemical variables. • LOFO validation confirms transferability across independent farms. Cold-climate anaerobic digestion (AD) of dairy manure is constrained by limited region-specific methane yield data and predictive tools under psychrophilic conditions. This study integrates multi-farm experimental data with interpretable machine learning (ML) to establish methane benchmarks and develop predictive models for Quebec dairy systems at 20 °C. Manure from six commercial farms was evaluated through 38-day batch digestion and > 90-day storage experiments to characterize methane production and emissions from both raw manure and digestate. Specific methane yields remained within a relatively narrow range (0.07–0.136 L CH 4 g −1 VS) despite variability in substrate characteristics, defining a stable and realistic psychrophilic design band. Supervised ML models trained exclusively on physicochemical parameters accurately predicted methane production across farms (R 2 = 0.809–0.925) and demonstrated strong cross-farm generalization using a leave-one-farm-out validation strategy. Interpretability analysis identified acetic acid and soluble COD as dominant predictors, consistent with acetoclastic methanogenesis and substrate availability. Random Forest demonstrated stable and reliable performance for digestate emission modeling in farms with observed data. Overall, this integrated experimental–ML framework provides region-specific methane benchmarks and validated predictive tools for improving psychrophilic digester design, emission-factor development, and greenhouse gas accounting in cold-climate dairy systems.