Prediction of thermal conductivity of pure liquids and mixtures using neural network

A new predictive tool exploiting neural network for evaluating thermal conductivity of liquids and mixtures at ambient or saturated pressures, is proposed. It covers a wide range on molecular species Including hydrocarbons, alcohols, water as well as inorganics, with values extending over the range of 40 to 700 mW/m/K. A three-layer forward network has been trained using experimental data to provide a preliminary set of weights which are progressively refined. This strategy has been adopted so to make it possible to automatically update the weights as new information becomes available. The predictions are significantly better than any correlation or physico-chemically based models.