Improving Data and Prediction Quality of High-Throughput Perovskite Synthesis with Model Fusion

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• High-throughput experimentation generates large datasets for materials discovery.
• Combining diverse predictive models can enhance accuracy over individual models.
• Metal halide perovskites are promising materials with complex synthesis requirements.

Purpose of the Study:

• To apply Combinatorial Fusion Analysis (CFA) to improve the prediction of metal halide perovskite synthesis.
• To assess the performance of fused models compared to individual machine learning models.
• To utilize fusion models for data quality control in high-throughput experimentation.

Main Methods:

• Developed four individual machine learning models: support vector machines, random forests, weighted logistic classifier, and gradient boosted trees.
• Explored 66 combinations of these models using CFA.
• Evaluated model performance based on precision in predicting crystal formation.

Main Results:

• The majority of fused models demonstrated improved prediction precision compared to individual models.
• The best fusion models achieved a 3.9 percentage point increase in precision over the best individual models.
• Fusion models successfully identified mislabeled data and highlighted synthesis sensitivity to temperature variations.

Conclusions:

• CFA is an effective method for combining diverse machine learning models to enhance prediction accuracy in materials synthesis.
• Model fusion can serve as a quality control mechanism for high-throughput experimental data.
• This approach revealed an unconsidered factor (temperature variation) influencing perovskite crystallization.