Abstract: To address the thermal error problem caused by thermal deformation in CNC machine tools under high-speed and heavy-load operations, a thermal error compensation method integrating Modelica-based multi-domain modeling with long short-term memory (LSTM) network is proposed. A high-fidelity digital twin model coupling mechanical, electrical, and thermodynamic multi-physics fields is established using Modelica, while the LSTM network provides data-driven compensation for nonlinear dynamic errors uncovered by the physical mechanism model. Experiments were conducted on a DMG MORI DMU 50 five-axis machining center to validate model performance through data collection of temperature, vibration, and thermal errors under preheating, step loading, and disturbance conditions. Results demonstrate that compared with the pure Modelica mechanism model, the Modelica-LSTM model achieves 51.2% reduction in root mean square error and 72% decrease in error fluctuation amplitude after compensation, showing significant prediction accuracy improvement under high-temperature and dynamic conditions. This method provides an effective physics-data collaborative solution for thermal error compensation in high-precision machine tools.