Data Fusion of Deep Learned Molecular Embeddings for Property Prediction

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Deep learning models offer accurate material property predictions but struggle with sparse data.
• Existing multitask learning methods are limited by data completeness and property correlations.
• Sparse datasets hinder the development of reliable predictive models in materials science.

Purpose of the Study:

• To develop an improved multitask learning framework for material property prediction using sparse data.
• To enhance the accuracy and efficiency of predictive models when training data is limited.
• To create a method that leverages property-specific representations from pretrained single-task models.

Main Methods:

• Fusing deep-learned embeddings from independent, pretrained single-task models.
• Developing a novel multitask model architecture that reuses these embeddings.
• Validating the approach on quantum chemistry benchmark datasets and sparse experimental data.

Main Results:

• The fused multitask model significantly outperforms standard multitask models on sparse datasets.
• Reusing pretrained embeddings leads to richer, property-specific representations.
• The proposed method requires fewer trainable parameters for model extension.

Conclusions:

• Fused embeddings offer a robust solution for multitask learning with sparse data in materials science.
• This technique improves predictive accuracy and model efficiency, broadening applicability.
• The method provides a foundation for more effective data-driven material discovery.