<p>Geometric precision in thin-walled parts critically relies on effectively managing their processing-induced deformation, wherein accurate deformation prediction constitutes the essential foundation for implementing subsequent control strategies. Existing predictive approaches, predominantly founded on numerical simulations or analytical models, often encounter limitations due to necessary simplifications of the actual machining process, inherent assumptions within solution procedures, and the persistent challenges associated with measuring and reliably forecasting residual stress distributions within the part. This paper introduces a method for predicting deformation through representation learning of the energy field derived from clamping force. The approach begins by examining energy variations in machining, where the deformation energy resulting from an unbalanced residual stress field is captured via changes in clamping force and in the geometric state of the part before and after deformation. Subsequently, considering the distinct stress equilibrium conditions of the part under clamped and released states, the deformation potential energy is modeled as a process of accumulation and release. Furthermore, the accumulation and release of deformation potential energy are treated as two distinct physical processes driven by the unbalanced residual stress field. A subspace learning method is employed to map both processes into a common subspace, thereby establishing a relationship model between energy accumulation and release. By integrating in-process clamping force and part geometry data, this work predicts part deformation in the released state. Experimental validation confirms the method's effectiveness for deformation prediction of machined parts under free-state conditions.</p>