Robust machine learning and ensemble learning approach to predict variation in experimental data for multiple...

Area of Science:

• Chemistry
• Materials Science
• Data Science

Background:

• Machine learning (ML) is crucial for predicting chemical compound properties (y) from experimental conditions (x).
• Current ML models use average measurements, failing to capture data variability.
• Existing methods for variability prediction can be inaccurate with measurement errors.

Purpose of the Study:

• To develop a robust ML method for predicting both the mean and variability of chemical properties.
• To improve prediction accuracy in datasets with multiple measurements and outliers.
• To address limitations of conventional ML approaches in handling experimental data variability.

Main Methods:

• Constructing multiple sub-datasets by selecting different y values for each sample.
• Training multiple ML models on these sub-datasets.
• Selecting models with the lowest mean absolute error for final predictions.
• Validating the method on a film thickness and haze dataset.

Main Results:

• The proposed robust method significantly outperformed conventional approaches in predicting both mean and variation.
• The method demonstrated superior accuracy compared to techniques that pre-emptively remove anomalies.
• Effective prediction was achieved even with datasets containing multiple y values and abnormal data points.

Conclusions:

• The novel robust ML method enhances prediction accuracy for chemical properties and their variability.
• This approach is effective for datasets with multiple measurements and outliers, without needing to remove samples.
• The findings offer a more reliable way to model experimental data in chemistry and materials science.