Impact of nanoparticle morphologies on property prediction using explainable AI.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Machine learning model performance is sensitive to data preprocessing steps like feature and sample selection.
• Explainable AI (XAI) offers methods to quantify the impact of these decisions on model outcomes.
• Accurate prediction of structure-property relationships is crucial for materials discovery.

Purpose of the Study:

• To apply residual decomposition with Shapely values to understand feature importance in predicting gold nanoparticle charge transfer properties.
• To identify specific nanoparticle shapes that are most influential for accurate property prediction.
• To assess how feature selection impacts the generalizability of models across different nanoparticle morphologies.

Main Methods:

• Utilized residual decomposition techniques.
• Applied Shapely values for feature attribution.
• Investigated machine learning models for predicting charge transfer properties of gold nanoparticles.
• Analyzed the influence of nanoparticle shape on predictive model performance.

Main Results:

• Identified specific nanoparticle shapes with significant influence on charge transfer property predictions.
• Quantified the impact of these influential shapes using Shapely values.
• Demonstrated how feature selection based on XAI can enhance model accuracy for diverse morphologies.
• Highlighted the importance of considering shape-specific contributions in structure-property modeling.

Conclusions:

• Residual decomposition combined with Shapely values is effective for interpreting machine learning models in materials science.
• Understanding feature influence, particularly nanoparticle shape, is critical for building robust predictive models.
• XAI methods enhance the reliability and interpretability of machine learning for predicting material properties.