Learning Diffusion from Sparse Data: A Machine-Learning Bridge between Molecular Motion and Macroscopic Transport

Area of Science:

• Computational chemistry and materials science
• Machine learning applications in physical sciences

Background:

• Predicting molecular self-diffusion coefficients (D*) is crucial but hindered by sparse experimental data and high simulation costs.
• Existing methods struggle with data scarcity and computational expense, limiting predictions across chemical space.

Purpose of the Study:

• To develop a data-centric machine learning framework for accurate and scalable prediction of molecular diffusion coefficients (D*).
• To integrate experimental data with molecular dynamics simulations using a semisupervised distillation strategy.
• To leverage large language model embeddings for improved chemical context and generalization.

Main Methods:

• A semisupervised distillation strategy selectively incorporates simulation data based on model uncertainty.
• A Random Forest model is initially trained on experimental D* and iteratively refined.
• Pretrained large language model embeddings are used to encode chemical context.

Main Results:

• The framework expanded the dataset from 130 to over 1,000 unique molecules while maintaining label reliability.
• The model achieved a held-out test set R² of 0.87 and an overall R² of 0.92.
• Learned D* descriptors improved skin permeability modeling accuracy and correlation.

Conclusions:

• The developed framework offers a scalable solution for bridging sparse experimental data with generalizable predictions.
• The approach enables interpretable and transferable modeling of molecular transport phenomena.
• This work advances the prediction of diffusion coefficients and their application in related fields.