Active meta-learning for predicting and selecting perovskite crystallization experiments.

Area of Science:

• Materials Science
• Chemistry
• Artificial Intelligence

Background:

• Autonomous experimentation systems optimize experimental design using algorithms and historical data.
• Machine learning in chemistry faces challenges due to limited, expensive, and time-consuming data acquisition.
• Active learning and meta-learning offer solutions for efficient learning with scarce data.

Purpose of the Study:

• To apply and evaluate meta-learning and active learning strategies for halide perovskite growth prediction.
• To determine optimal methods for incorporating historical data into machine learning models for crystal prediction.
• To assess the performance of the Probabilistic LATent model for Incorporating Priors and Uncertainty in few-Shot learning (PLATIPUS) in experimental settings.

Main Methods:

• Utilized the model-agnostic meta-learning (MAML) model and the PLATIPUS approach.
• Trained models on a dataset of 1870 reactions involving 19 organoammonium lead iodide systems.
• Compared PLATIPUS against k-nearest neighbor and decision tree active learning algorithms, and a random baseline, using four new chemical systems.

Main Results:

• Identified optimal strategies for integrating historical data into active and meta-learning models.
• PLATIPUS demonstrated superior prediction accuracy for reaction outcomes leading to crystals.
• PLATIPUS outperformed other active learning algorithms and random selection within a budget of 20 experiments.

Conclusions:

• PLATIPUS, an active learning extension of MAML, is highly effective for materials discovery in data-limited scenarios.
• Meta-learning approaches, particularly PLATIPUS, offer significant advantages over traditional active learning for predicting chemical synthesis outcomes.
• This study validates the utility of advanced machine learning for accelerating the discovery of functional materials like halide perovskites.