Prediction of the phase transition temperatures of functional nanostructured liquid crystals: a machine learning method based on small data for the design of self-assembled materials

Abstract

Here we demonstrate the prediction of the isotropization temperatures of nanostructured ionic liquid crystals (ILCs) by a machine learning method. ILCs, which self-assemble into dynamic and well-ordered nanostructures, have been extensively studied because they can be used as various functional materials including electrolytes, water treatment membranes, and stimuli-responsive materials. In general, machine learning is not easily applied to predict dynamic functions of such self-assembled materials because of the complex structure–function relationship and small data size. Our approach is to use the machine learning method combined with our experimental experiences and chemical insights as researchers. The training data of order–disorder transition (isotropic) temperatures of 116 wedge-shaped ILCs reported by our groups or other groups have been analyzed by this method, leading to the successful construction of the prediction model with a root mean squared error value of 23.7 °C. The model is straightforward, interpretable and reasonable from the viewpoint of chemistry. Moreover, the constructed model predicted the transition temperatures of 45 wedge-shaped ILCs in the test data. The model has sufficient accuracy to predict the transition temperatures of certain kinds of ILCs, which helps the molecular design of ILCs. The transition temperature of isotropization, which is the transition from an ordered liquid-crystalline state to a random isotropic state, is the key factor for ILCs to achieve excellent properties. It is important to predict the transition temperatures before starting experiments, but this has been challenging because complex factors determine the transition temperatures. The model constructed in the present study may accelerate the development of functional ILCs.