Neural network-based reconstruction of steady-state temperature systems with unknown material composition.

Area of Science:

• Thermal Engineering
• Computational Science
• Artificial Intelligence

Background:

• Effective thermal management is vital for device performance and longevity.
• Accurate temperature distribution determination is challenged by incomplete observational data.
• Existing methods may require material property information, limiting applicability.

Purpose of the Study:

• To develop a neural network-based approach for reconstructing complete temperature fields from partial observations.
• To overcome the limitations of incomplete temperature data in thermal systems.
• To validate the model's effectiveness without needing spatial material property data.

Main Methods:

• A physics-informed fully convolutional auto-encoder was employed.
• A preceding spatial propagator network module was integrated.
• The model was validated using virtual sensing in simulated and real-world thermal systems.

Main Results:

• The neural network approach successfully reconstructed full temperature fields from sparse data, including challenging grid-edge sampling (3.12% sampling ratio).
• Achieved a relative average error of 1.1% in temperature field reconstruction.
• Outperformed the Kriging method in predicting maximum system temperature and demonstrated greater robustness with limited measurements.

Conclusions:

• The proposed neural network method offers an effective solution for temperature field reconstruction from incomplete data.
• This approach enhances thermal management by enabling accurate temperature prediction without material property information.
• The method shows significant promise for virtual sensing applications in diverse thermal systems.