(474h) Predicting Pair Correlation Functions of Soft and Hard Materials Using Machine Learning

The pair correlation function (PCF) is an important feature to study the structure of materials at the atomic or molecular level. It provides information about the probability of finding two particles at a certain distance from each other, and can be used to characterize the spatial arrangement of particles in a material. Soft materials, such as polymers, colloids, and biomolecules, are characterized by their flexibility and deformability. Their PCFs typically show long-range correlations and broad peaks, reflecting their disordered and heterogeneous nature. Hard materials, such as crystals, semiconductors, and metals, are characterized by their rigidity and strong interatomic interactions. Their PCFs typically show sharp peaks and short-range correlations, reflecting their ordered and homogeneous nature. Molecular simulations have been routinely used to estimate the PCF of materials. However, they are computationally very expensive, and estimating PCFs for the entire compositional space of a material remains a substantially challenging task. Here, we propose a machine learning pipeline (ML) for rapid prediction of PCF of a material based on the information of its composition. We conduct ~ 100 molecular simulations for a given material and calculate its PCFs for varying composition. We use these PCFs and their corresponding composition information to build, test and validate the ML method. Within this ML pipeline, the grayscale images of PCF of a material are encoded to a latent space using a convolutional neural network (CNN) autoencoder. Subsequently, a random forest regressor establishes a correlation between the composition of the material and the latent space representation of its PCF. This ML model is then used for predicting PCF for many unknown compositions of the material. We have successfully demonstrated the performance of our ML model for three representative cases – binary liquid mixtures, polymer nanocomposites and oxide glasses. This method is very generic and can accelerate the design, discovery, and fundamental understanding of composition-PCF relations of a wide range of soft and hard materials.