Conditioned Diffusion Framework for Temperature Field Generation via Multi-Source Physical Parameters Fusion and Sensor Layout-Decoupled Modeling

Global temperature sensing is essential for state monitoring, fault diagnosis, and over-temperature alerts in aerospace, electronics, and energy systems. However, existing discriminative reconstruction methods rely on static and rigid input specifications, limiting flexibility in accommodating multi-source physical parameters and tightly coupling models to predefined temperature sensor layouts. Therefore, a conditional generative framework (PRISM-Diffusion) is proposed that integrates diverse physical parameters while decoupling model training from sensor configurations, covering diverse temperature sensing scenarios. A unified representation of temperature-related physical parameters (e.g., heat sources, boundary conditions) is constructed and adaptively fused via conditional encoding and multi-head cross-attention. Classifier-free guidance further enhances the directional influence of high-level semantic physical parameters during generation. During training, a joint conditioning strategy with “full condition/partially-missing/unconditional” for temperature-related physical parameter dropout is employed, allowing the model to adaptively accept diverse physical parameter inputs in practical scenarios. Notably, the training only focuses on noise estimation conditioned on physical parameters, without any dependency on temperature sensor layouts. For temperature sensor information, measurements from varying sensor layouts are incorporated into the reverse sampling chain of trained model via observation-induced graph construction and stochastic prior alignment, enabling guided generation without task-specific retraining. Within a unified probability framework, multiple generation paradigms are enabled: (1) forward simulation driven solely by physical parameters; (2) inverse reconstruction from sparse measurements; and (3) joint generation combining sparse observations with multiple physical parameters. In 2-dimensional complex heat-source simulations, PRISM-Diffusion generates representative thermal fields with RMSE/RMSE_h < 0.2 K and MaxAE/MaxAE_h < 0.51 K in joint reconstruction using complete physical parameters and 36 sparse sensors, while maintaining noise resilience. Overall, this framework avoids the fixed input-to-field mapping inherent in conventional methods and establishes a flexible probabilistic paradigm for thermal state inference.