Thermal-flow field prediction in energy-equipment components using a physics-guided conditional diffusion framework: validation on compressor mufflers

Real-time and accurate prediction of internal thermal-flow fields in energy-equipment components is crucial for understanding flow and heat-transfer mechanisms, supporting digital twin modeling, and enabling performance optimization. However, irregular geometric boundaries, frequently varying operating conditions, and strongly coupled multiphysics processes make high-fidelity numerical simulations insufficient for rapid response, while conventional machine learning methods often exhibit limited generalization and reduced physical consistency. To address these challenges, this study proposes a physics-guided conditional diffusion framework for rapid thermal-flow field prediction under varying operating conditions. The framework generates irregular-boundary masks using morphological closing to explicitly identify effective fluid regions, adopts a cosine noise schedule and a FiLM-UNet denoising network to learn the conditional probability distribution of complex thermal-flow fields, and incorporates operating conditions, geometric masks, and Fourier-feature-encoded multi-harmonic phase information as combined conditional inputs. Furthermore, self-attention and classifier-free guidance are integrated to enhance long-range spatial dependency modeling and cross-condition generalization. Composite physics losses are embedded in the training process to improve the physical consistency of predictions. The framework is validated using the transient field of a reciprocating compressor suction muffler and the steady field of a discharge muffler. Under extrapolation conditions, compared with the conventional conditional diffusion model, the proposed framework reduces MSE by 64.81% and 95.53%, respectively, and reduces the vorticity MSE and spectral MSE by 75.85% and 19.44%, respectively. These results demonstrate that the framework provides an effective approach for rapid field-level perception and digital twin modeling of complex energy-equipment components.