The co-pyrolysis technology of agricultural and forestry waste and polluted soil has simultaneously achieved biomass resource utilization and soil remediation. However, the heterogeneity of raw materials, non-linear processes, and long ecological response cycles pose serious challenges to its precise regulation. This article systematically reviews the new paradigm of machine learning driven closed-loop design and its multi-scale improvement mechanism for soil health. Firstly, the key applications of machine learning in co-pyrolysis systems were elucidated: by integrating multimodal sensing data with deep neural networks, random forests, and other algorithms, a proxy model was constructed between raw material attributes, process parameters, product performance, and repair effects. Based on this, an intelligent control loop integrating perception, prediction, optimization, and feedback was formed. Secondly, the hierarchical mechanism of closed-loop system driving soil functional regeneration was revealed layer by layer from four scales: molecular/nano interface (quantum confinement catalysis and covalent bond orientation fixation), micro/millimeter structure (biomimetic topology reconstruction and dynamic pore regulation), in situ field (nutrient cycling activation and pollutant bioavailability reduction), and ecosystem (carbon sink function enhancement and microbial network reconstruction). Furthermore, the core bottlenecks such as path uncertainty caused by raw material fluctuations, quantum tunneling barriers within nanopores, physical limits of multifunctional performance of biochar, thermodynamic contradictions of system energy self-sustaining, unpredictability of long-term ecological response, and adaptation conflicts of current standard systems were summarized. Interdisciplinary integration directions such as quantum chemistry computing, ultrafast laser regulation, biomimetic reactors, synthetic biology biochar symbiotic systems, and blockchain trusted carbon management were also discussed. This review provides a systematic theoretical framework for the deep intersection of machine learning and thermochemical remediation technology, and also points out the technical path for achieving precise, intelligent, and sustainable remediation of polluted soil.