An inverse analysis to estimate the thermal properties of nanoporous aerogel composites using the particle swarm optimized deep neural network

To understand the transient heat transfer characteristics of nanoporous aerogel insulating composites, solving the inverse heat transfer problem would be crucial for identifying the temperature-dependent thermal properties of composites. In this study, with constructed a forward model to numerically investigate the heat transfer in composites, a deep neural network (DNN) model and a particle swarm optimized deep neural network (PSO-DNN) model are conducted to rapidly estimate the effective temperature-dependent thermal conductivity of the desiccated and moist composites from the temperature response measurements. With the DNN model, the retrieved thermal conductivities for desiccated composites possess low deviation to experimental measurements (<3.2%) and constantly low errors (<5.2%) from 280 K to 1080 K. The precision of the DNN solver could be enhanced by adjusting the hyperparameters of the neural networks using PSO. The retrieved thermal conductivities possess low deviation from experiments (<2.5%) and low relative errors within 1.5%. Furthermore, the robustness of the PSO-DNN solver is discussed when commercial thermocouple measurement errors are considered, within retrieving the thermal properties of desiccated and moist aerogel composites.