Inverse Design of Nanoparticles Using Multi‐Target Machine Learning

Abstract

In this study a new approach to inverse design is presented that draws on the multi-functionality of nanomaterials and uses sets of properties to predict a unique nanoparticle structure. This approach involves multi-target regression and uses a precursory forward structure/property prediction to focus the model on the most important characteristics before inverting the problem and simultaneously predicting multiple structural features of a single nanoparticle. The workflow is general, as demonstrated on two nanoparticle data sets, and can rapidly predict property/structure relationships to guide further research and development without the need for additional optimization or high-throughput sampling.

Causal Paths Allowing Simultaneous Control of Multiple Nanoparticle Properties Using Multi‐Target Bayesian Inference

“…To focus on the structural characteristics of nanodiamonds that are important to all of the properties, we incorporated recent results of multi-target regression that has previously been used as a basis or nanoparticle inverse design. [34] These results (reported elsewhere) achieved remarkably good performance (with low mean square error [MSE] and mean absolute error [MAE]) that are sufficient to predict property labels simultaneously without lost of accuracy or generalizability, as shown in the Supporting Information. The original approach began with feature engineering to exclude features with correlation over 95% (see Supporting Information for correlation matrix) and a variance of less than one standard deviation.…”

“…High-dimensional numerical target labels can be simultaneously predicted with high accuracy using random forests (RFs). [34] An RF [35] was a non-parametric ensemble tree-based method that made no assumption on the structure of the model and can intrinsically handle multi-task problems as the leaf nodes can refer to any collection of relevant classes. [36][37][38] RFs constructed a large number of decision trees and combined bootstrap aggregation, also known as bagging, [39] and random feature selection methods.…”

Inverse Design of MXenes for High-Capacity Energy Storage Materials Using Multi-Target Machine Learning

“…High-dimensional numerical target labels can be simultaneously predicted with high accuracy by using treebased ensemble methods such as random forests. 32 Tree-based methods are a nonparametric explanatory approach that makes no assumption on the structure of the model but exposes interpretable feature importances, making them ideal predicting structure/property relationships. They can intrinsically handle multi-task problems as the leaf nodes can refer to any collection of relevant classes, making them ideal for inverse design.…”

“…The method is capable of outputting a set of target physicochemical characteristics that can simultaneously deliver a predefined set of properties without the need for additional optimization or an exhaustive data set that makes a priori prediction redundant. 32 In this work we apply this multi-target inverse design workflow to a 2D transition metal carbides data set generated by using electronic structure simulations to develop a property/structure relationship capable of predicting the best MXene for a given gravimetric capacity, voltage, and induced charge, simultaneously. This requires extending the methodology to include multi-target classification, as opposed to the multi-target regression models demonstrated in previous work, because the chemical formula of materials such as MXenes consists of discrete variables.…”

“…Provided sufficient data are available, machine learning (ML) provides a convenient way of rapidly predicting the electrochemical properties of MXenes (the target labels) based on the physicochemical characteristics (the structural features). , These are termed structure/property relationships and are capable of providing accurate predictions for a variety of candidate materials for energy applications. − Machine learning models have been used to show that the electrochemical potential can be predicted for new cathode materials, and the quantitative molecular structure–redox potentials relationships can be established. , In each case multiple structural characteristics, such as the numbers of oxygen atoms, lithium atoms, boron atoms, carbon atoms, and aromatic rings, are used to individually predict the voltage, capacity, or charge. , Recently a general ML approach was developed based on multi-target random forests that accurately predicts inverse property/structure relationships. The method is capable of outputting a set of target physicochemical characteristics that can simultaneously deliver a predefined set of properties without the need for additional optimization or an exhaustive data set that makes a priori prediction redundant …”