Deep Learning-Based Prediction of Interfacial Surface Tension of Pendant Droplets

Pendant drops of oxide-coated high-surface tension fluids frequently produce perturbed and distorted shapes that impede surface tension measurements. Eutectic Gallium Indium (eGaIn), a high surface tension fluid coated with a 5nm Gallium Oxide film, falls under this fluid classification, and the recent emergence of eGaIn-based applications often cannot proceed without analyzing interfacial energetics in different environments. The low viscosity, chemical reactivity, and the passive surface oxide have also contributed to the accelerated materials discovery where eGaIn is alloyed with other elements for novel catalysts, bioreactors, and sensor discovery. In this study, for the first time, we are using a regression-based Deep Learning (DL) technique to predict the surface tension of eGaIn that can be extended to analyze any pendant drop. Time and resource-consuming image analysis software with manual intervention is generally used to analyze interfacial surface tension of different fluids. Here, we address the challenges associated with physics-based interfacial energy measurements of eGaIn and introduce a deep convolutional neural network architecture that can be deployed without using Goniometers (interfacial energy measurement instruments). A set of 13081 images were acquired and further analyzed to measure the surface tension. The dataset was split into 10594 training, 1178 validation, and 1309 test samples. We trained various convolutional neural networks on this regression task to accurately predict and compare the pendant drop shape properties. Our best-performing model, ResNet-50v2, achieved a mean absolute error of 0.995 and an R2 value of 99.97%, which signifies a high correlation, and our architecture can predict 99% of the variance in outcome. eGaIn droplets can create a non-axisymmetric shape due to their oxide layer, which cannot be solved with presently available software but can easily be handled with our presented DL technique. In limited-resource settings, our approach can also be deployed with a cell phone, and this deep learning model can be further leveraged to harness the influence of the external environment.