Abstract:The resource utilization of livestock and poultry manure represents a critical strategy for mitigating agricultural non-point source pollution and achieving carbon neutrality. However, mainstream technologies, specifically aerobic composting and anaerobic digestion, are significantly constrained by the "black box" nature of multiphase medium coupling, non-linear kinetics, and microbial community succession. Traditional mechanistic models, most notably the Anaerobic Digestion Model No. 1 (ADM1), struggle to accommodate the high heterogeneity of feedstocks due to challenges in parameter calibration and structural rigidity. Consequently, these processes face persistent engineering bottlenecks, including low organic matter conversion efficiency, process instability, and uncontrollable biosecurity risks. To address these challenges, this study systematically reviews recent advances in the application of machine learning technologies to livestock manure resource utilization. We classify and analyze the application logic of three primary algorithmic categories: (1) tree-based models such as Random Forest and eXtreme Gradient Boosting (XGBoost); (2) deep learning architectures including Artificial Neural Networks (ANN) and Convolutional Neural Networks (CNN); and (3) intelligent optimization techniques, exemplified by Genetic Algorithms (GA). Their applications are evaluated in the modeling of multi-dimensional process parameters, interpretation of microbial community mechanisms, and contactless intelligent sensing. Furthermore, we examine the integration of these algorithms with traditional biological theories to circumvent the limitations of single-model approaches. Results demonstrate that machine learning algorithms outperform traditional mechanistic models in handling highly noisy and non-linear datasets. In process prediction, tree-based models such as Categorical Boosting (CatBoost) and XGBoost, when optimized by GA, achieve high predictive accuracy for key physicochemical indicators, including the carbon-to-nitrogen ratio and seed germination index. For mechanistic interpretation, Random Forest shows a strong capacity for feature selection, identifying core functional genera such as Stenotrophomonas and Bacillus involved in lignocellulose degradation, and revealing that mobile genetic elements are the principal biological drivers of horizontal gene transfer of antibiotic resistance genes. In dynamic simulation, ANN and Nonlinear AutoRegressive models with eXogenous inputs (NARX) effectively capture the temporal fluctuations of biogas production at an industrial scale, significantly reducing prediction errors and surpassing the performance of Response Surface Methodology. In intelligent sensing, the incorporation of attention mechanisms such as Squeeze-and-Excitation Networks (SENet) and Efficient Channel Attention (ECA) into CNN architectures markedly enhances the accuracy of compost maturity identification under complex field conditions. Ultimately, machine learning enables a paradigm shift from empirical management to intelligent decision-making in livestock manure treatment by overcoming "black box" limitations through data-driven, non-linear mapping and autonomous feature learning. However, challenges remain concerning model interpretability and physical consistency. Future research should aim to develop "grey box" models that deeply integrate physicochemical mechanisms with data-driven algorithms, ensuring compliance with mass and energy conservation laws in data-scarce environments. Additionally, constructing multimodal predictive systems that incorporate multi-omics data is critical for simultaneously enhancing resource conversion efficiency and enabling precise control over biological safety risks. Close-