Accurate Lattice Free Energies of Packing Polymorphs from Probabilistic Generative Models.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Calculating finite-temperature free energy differences between molecular crystal polymorphs is crucial for predicting stability but is computationally intensive.
• Existing methods like Einstein crystal require extensive sampling of intermediate states.

Purpose of the Study:

• To implement and assess machine learning-enabled targeted free energy calculations for molecular crystals.
• To compute the free energy difference between Ice XI and Ic polymorphs efficiently.
• To compare the accuracy and efficiency of different machine learning model architectures and representations.

Main Methods:

• Utilized flow-based generative models for targeted free energy calculations.
• Trained models on locally ergodic data sampled exclusively from the ensembles of interest.
• Employed an overfitting-aware weighted averaging strategy for convergence monitoring.
• Compared results with the Einstein crystal method for ground-truth validation.

Main Results:

• Machine learning methods accurately and cost-effectively compute free energy differences between disconnected metastable ensembles.
• The targeted approach avoids sampling intermediate Hamiltonians, leading to significant computational savings.
• Model performance varies with system size, temperature, and the representation of supercell degrees of freedom (Cartesian vs. quaternion-based).

Conclusions:

• Machine learning-enabled targeted free energy calculations provide a computationally efficient alternative to classical methods for polymorph stability studies.
• The choice of representation is critical for accurate and generalizable results in larger systems and at higher temperatures.
• This work represents a significant step towards efficient free energy calculations for complex molecular crystals.