Active meta-learning for predicting and selecting perovskite crystallization experiments

Abstract

Autonomous experimentation systems use algorithms and data from prior experiments to select and perform new experiments in order to meet a specified objective. In most experimental chemistry situations, there is a limited set of prior historical data available, and acquiring new data may be expensive and time consuming, which places constraints on machine learning methods. Active learning methods prioritize new experiment selection by using machine learning model uncertainty and predicted outcomes. Meta-learning methods attempt to construct models that can learn quickly with a limited set of data for a new task. In this paper, we applied the model-agnostic meta-learning (MAML) model and the Probabilistic LATent model for Incorporating Priors and Uncertainty in few-Shot learning (PLATIPUS) approach, which extends MAML to active learning, to the problem of halide perovskite growth by inverse temperature crystallization. Using a dataset of 1870 reactions conducted using 19 different organoammonium lead iodide systems, we determined the optimal strategies for incorporating historical data into active and meta-learning models to predict reaction compositions that result in crystals. We then evaluated the best three algorithms (PLATIPUS and active-learning k-nearest neighbor and decision tree algorithms) with four new chemical systems in experimental laboratory tests. With a fixed budget of 20 experiments, PLATIPUS makes superior predictions of reaction outcomes compared to other active-learning algorithms and a random baseline.

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

“…The underlying modeling problem is of the following form: Given a chemical system comprised of a metal halide, one specific organoammonium halide salt (for brevity, we refer to this as the "amine"), a solvent, and an additive (e.g., formic acid)-find the set of concentrations for each of these species, that result in the formation of a large, high-quality single crystalline product via an inverse temperature crystallization reaction. (16,21) Concentrations of the three solutes (metal halide, amine, and formic acid) define a 3dimensional space, and only compositions within the convex hull of the initial stock solutions are feasible. (22) Precision limits on the robotic liquid handler result in a discrete state set grid of approximately 20,000 feasible compositions within the composition space for each chemical system.…”

“…Eleven active learning models were assigned the initial prediction task: Bayesian Additive Regression Trees with Transfer Learning (BART), (23,24) PLATIPUS (PLT), (21,25) Bayesian Optimization using Gaussian Processes (BGP),( 26) Falcon (FAL), Falcon GPBO (G), (26,27) Falcon DNGO (D),( 28) Gryffin (GR),( 29) Gaussian Process with Transfer Acquisition Functions (TA) (30), and Falcon with Historical Data (FH), as well as active learning k-Nearest Neighbors (KNN) (31,32) and Decision tree (DT) (33) models which serve as a baseline for model performance. (See Methods for a complete description of each model.)…”

“…The general structure of the campaign, including the historical dataset, experimental details, and performance of the PLATIPUS model were previously described in Ref. (21). The campaign is split into model training and refinement, active learning, and recommendation phases.…”

“…The historical datasets were acquired using the automated ITC experimental procedure described above, and previously described in Ref. (21). Each data item describes an inverse-temperature crystallization (ITC) metal halide perovskite synthesis by including concentrations of lead iodide, formic acid, and an organoammonium cation (referred to as the amine), other reaction conditions (such as temperature), and outcomes.…”

“…Exploitation recommender results described here were previously reported in Ref. (21). DT: Decision Tree(33) serves as a baseline model with a maximum depth of 12, 5 minimum samples per split and 1 minimum sample per leaf and class weight ratio of 0.105 for failure and 0.895 for success.…”