Towards machine learning based prediction of magnetic nanoparticle properties

Abstract

Magnetic nanoparticles (MNP) must fulfill specific requirements according to their dedicated application, e.g. as contrast agents for magnetic particle imaging (MPI) or as heating agents in magnetic fluid hyperthermia (MFH). The synthesis of MNP with specific properties is, however, a complex process, which is not fully understood. Based on standardized parameter studies performed with automated MNP manufacturing techniques, Support Vector Regression (SVR) was used to predict the MNP properties and combine them with the relevant synthesis parameters.

An SVR model with 24 input parameters, i.e. all parameters of the synthesis setup, was used to predict a target property, here the MNP size. For training, the data was split 80?/?20 into training and test set. An extensive grid search for the SVR parameters C, ?, ? and the kernel was conducted using Leave-One-Out cross validation.

The predictive model shows accurate predictions of the desired target property (here the MNP size) with a variance of ±30 nm when predicting the test set. In addition, the predictive model is interpretable and allows to analyze correlations between individual input parameters and the target property. For specific diameters, the MPI signal was determined and will be included in the predictive model in the future.

To conclude, the proposed SVR method yields promising predictions of key MNP characteristics. This method is versatile and can be applied to any synthesis procedure and parameter spectrum making the prediction of the MNP properties possible for specified production process parameters and paving the way for autonomous tailored MNP synthesis.