Prediction of nanoscale thermal transport and adsorption of liquid containing surfactant at solid–liquid interface via deep learning

Hypothesis: Recent advances in deep learning (DL) have enabled high level of real-time prediction of thermophysical properties of materials. On the other hand, molecular dynamics (MD) have been long used as a numerical microscope to observe detailed interfacial conditions but require separate simulations that are computationally costly. Hence, it should be possible to combine MD and DL to obtain high resolution interfacial details at a low computational cost. Experiment: We proposed a novel DL encoding–decoding convolutional neural network (CNN) coupled with MD to realize the mapping from micro solid–liquid interface geometry to molecular temperature and density distribution of liquid containing surfactant. A multi-nanoscale optimization scheme was further proposed to reduce the uncertainty of DL prediction at the expense of local details to obtain more resilient predictors. Findings: The statistical results showed that the proposed CNN had high prediction accuracy and could reproduce the heat transfer and adsorption phenomena under the influence of various factors including liquid composition, wettability, and solid surface roughness, while the computational efficiency was greatly improved. Our DL method with the support of multi-nanoscale learning strategies can achieve the fast and accurate visualization and prediction of various interfacial properties of liquid and assist for interfacial material design.