Bias free multiobjective active learning for materials design and discovery

Shefang Wang,Giriprasad Melpatti Jothiappan,Brian Yoo,Berend Smit,Kevin Maik Jablonka

doi:10.1038/s41467-021-22437-0

Shefang Wang, Giriprasad Melpatti Jothiappan + Show 3 more

Open Access

https://doi.org/10.1038/s41467-021-22437-0

Copy DOI

Abstract

The design rules for materials are clear for applications with a single objective. For most applications, however, there are often multiple, sometimes competing objectives where there is no single best material and the design rules change to finding the set of Pareto optimal materials. In this work, we leverage an active learning algorithm that directly uses the Pareto dominance relation to compute the set of Pareto optimal materials with desirable accuracy. We apply our algorithm to de novo polymer design with a prohibitively large search space. Using molecular simulations, we compute key descriptors for dispersant applications and drastically reduce the number of materials that need to be evaluated to reconstruct the Pareto front with a desired confidence. This work showcases how simulation and machine learning techniques can be coupled to discover materials within a design space that would be intractable using conventional screening approaches.

Highlights

The design rules for materials are clear for applications with a single objective
No single optimum is generally preferred over all the others; the most valuable information any search in the design space can give is the set of all possible materials for which none of the performance indicators can be improved without degrading some of the other indicators
12 9 -ΔGads analysis on the last two features already highlights a key difference between the theory and our model: our model provides us with insights into what happens when we change the composition

Summary

Introduction

The design rules for materials are clear for applications with a single objective. For most applications, there are often multiple, sometimes competing objectives where there is no single best material and the design rules change to finding the set of Pareto optimal materials. No single optimum is generally preferred over all the others; the most valuable information any search in the design space can give is the set of all possible materials for which none of the performance indicators can be improved without degrading some of the other indicators In statistical terms, these materials are referred to as the set of all Pareto-optimal solutions (i.e., the Pareto front). We initialize a model with a small sample of our design space and iteratively add labels, i.e., measurements or simulation results, to the training set where the model needs them most This allows us to efficiently build a model that is able to solve the question of what materials are Pareto optimal and which ones we should discard for further investigation. This means we can only say if a material is Pareto dominating or not, but we cannot directly compare them; the introduction of a total order is nothing more than a (subjective) formula on how to compare apples and pears[16]

Objectives

Methods

Results

Conclusion