Autotrophic respiration (Ra) and heterotrophic respiration (Rh) are critical components of the terrestrial carbon cycle, yet they remain inadequately quantified due to challenges in separating these processes in field measurements. The lack of continuous, ecosystem-level data for Ra and Rh hinders our understanding of the ecosystem carbon cycle and leads to significant disagreements in carbon budget simulations. Traditional methods often partition ecosystem respiration (Reco) from net ecosystem exchange (NEE) using eddy covariance (EC) flux tower data, but rarely distinguish Ra and Rh with high confidence. This study aims to address these challenges by synthesizing available datasets and extending the knowledge-guided machine learning framework, KGML-Carbon, to accurately partition Ra and Rh at the field-to-regional level. We incorporate continuous soil respiration data from auto chambers, such as the COSORE database, and AmeriFlux FLUXNET, to train and validate the framework robustness. Additionally, our existing KGML methods rely on fine-tune-based transfer learning, which can lead to overfitting and loss of valuable prior knowledge from pretraining. To address this, we utilize advanced transfer learning techniques, specifically Task-aware Modulation using Representation Learning (TAM-RL), which effectively transfers model knowledge while preserving critical prior information. Our expanded KGML-Carbon framework integrates cutting-edge scientific insights into biogeochemical processes with advanced machine learning, leveraging multi-task learning to synthesize diverse data sources. This study develops a new methodology for partitioning Ra and Rh, providing detailed, continuous data that contribute to more accurate carbon budget estimations. By using our advanced KGML-Carbon framework, we aim to enhance the understanding of the spatial and temporal dynamics, key drivers, and responses of Ra and Rh to climate change.