Serendipity based recommender system for perovskites material discovery: balancing exploration and exploitation across multiple models

Abstract

Machine learning is a useful tool for accelerating materials discovery, however it is a challenge to develop accurate methods that successfully transfer between domains while also broadening the scope of reaction conditions considered. In this paper, we consider how active- and transfer-learning methods can be used as building blocks for predicting reaction outcomes of metal halide perovskite synthesis. We then introduce a serendipity-based recommendation system that guides these methods to balance novelty and accuracy. The model-agnostic recommendation system is tested across active- and transfer-learning algorithms, using laboratory experiments for training and testing and a time-separated hold out that includes four different chemical systems. The serendipity recommendation system achieves high accuracy while increasing the scope of the synthesis conditions explored.

In Pursuit of the Exceptional: Research Directions for Machine Learning in Chemical and Materials Science

“…58 Algorithmic performance can also depend on the initial data set (the "cold start" problem), and available data sets often exhibit sampling biases. 59 This problem can be partially mitigated by adding additional constraints to maximize the explored input space 54 or by incorporating human expertise in the loop. 60 While previous research articles have benchmarked computational methods and metrics for this task, 61,62 and a recent perspective discussed types of machine-learning guided iterative experimentation toward this goal, 15 a more critical view of the field is that regardless of the accuracy produced by these methods, they will not generate the materials necessary to enable paradigm shifts.…”

“…This is synergistic with our previous recommendation to avoid premature optimization. Sloppiness can be active (e.g., adding randomness to materials experiment plans 59 or using an additional cost function to experiment generation that maximizes experiment diversity 54 ) or passive (e.g., taking advantage of uncontrolled changes in laboratory temperature and humidity as natural experiments 171 ). Variations in parameter values ("microsloppiness") are more easily achieved, but less likely to lead to large improvements; variations in reagent identity or steps ("macro-sloppiness") typically must be deliberately programmed.…”

“…Active learning methods were used to sample uncertain new experimental conditions, and the algorithm was then employed to select the optimal set of input parameters to rapidly achieve the desired growth rate. Other illustrative ML-enhanced materials optimization examples include nanocrystal growth and optical properties in a microfluidic system, mechanical properties of 3d-printed structures, crystal growth conditions, − halide alloy stability, , and superconductivity . See refs and for more comprehensive reviews.…”

In Pursuit of the Exceptional: Research Directions for Machine Learning in Chemical and Materials Science

“…Active learning methods were used to sample uncertain new experimental conditions, and the algorithm was then employed to select the optimal set of input parameters to rapidly achieve a desired growth rate. Other illustrative ML-enhanced materials optimization examples include nanocrystal growth and optical properties in a microfluidic system, 50 mechanical properties of 3d-printed structures, 51 crystal growth conditions, [52][53][54] and halide alloy stability 20,55 , and superconductivity. 56 See Refs.…”

The rise of automated curiosity-driven discoveries in chemistry

“…In a recent study, Schrier 26 and colleagues proposed a novel approach combining ML with a serendipity-based recommendation system. Using active learning strategies, i.e.…”