Improvement of Diffusion Coefficient Prediction by Active Learning.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Chemical Engineering

Background:

• Predicting diffusion coefficients in mixtures is crucial but experimentally challenging due to data scarcity.
• Machine learning (ML) models offer potential but require extensive training data, which is costly to acquire.
• Active learning (AL) strategies can optimize experimental design for targeted data acquisition.

Purpose of the Study:

• To investigate AL strategies for planning diffusion coefficient measurements.
• To improve ML-based prediction of diffusion coefficients at infinite dilution (D_ij^∞).
• To evaluate the impact of AL-guided data on matrix completion methods (MCMs).

Main Methods:

• Systematic testing of AL strategies on synthetic data for D_ij^∞.
• Utilizing uncertainty sampling as an effective AL strategy.
• Performing pulsed-field gradient (PFG) nuclear magnetic resonance (NMR) spectroscopy for new D_ij^∞ measurements.
• Retraining hybrid MCMs with newly acquired experimental data.

Main Results:

• Uncertainty sampling proved effective for planning D_ij^∞ measurements.
• Nineteen new D_ij^∞ data points were measured for previously uncharacterized mixtures.
• Hybrid MCMs incorporating semiempirical model (SEGWE) predictions showed substantial accuracy improvements.
• The relative mean squared error on the test set was nearly halved for one MCM.

Conclusions:

• AL strategies enable significant improvements in ML predictions of diffusion coefficients with minimal experiments.
• The effectiveness of AL depends on the specific ML model and its integration with prior information.
• Targeted experimental design is key to maximizing the value of ML in predicting physical properties.