Efficient Molecular Crystal Structure Prediction and Stability Assessment with AIMNet2 Neural Network Potentials.

Area of Science:

• Materials Chemistry
• Computational Materials Science
• Crystallography

Background:

• Accurate crystal structure prediction (CSP) is crucial for materials discovery but computationally expensive.
• Traditional methods rely on first-principles calculations, which become prohibitive for large systems requiring millions of energy evaluations.
• The Cambridge Crystallographic Data Centre's (CCDC) blind tests highlight the need for more efficient CSP methodologies.

Purpose of the Study:

• To develop and validate a computationally efficient approach for crystal structure prediction.
• To demonstrate the effectiveness of machine-learned interatomic potentials (MLIPs) trained on molecular clusters for CSP.
• To assess the performance of the AIMNet2 MLIPs in accurately characterizing the CSP landscape and ranking crystal stability.

Main Methods:

• Training target-specific AIMNet2 machine-learned interatomic potentials (MLIPs) on density functional theory (DFT) calculations of molecular clusters (n-mers).
• Utilizing gas-phase dispersion-corrected DFT reference data for training the MLIPs.
• Applying the trained MLIPs to assess the relative stability of candidate crystal structures in CSP workflows.

Main Results:

• MLIPs trained on n-mer data successfully extended to crystalline environments, accurately characterizing the CSP landscape.
• The methodology correctly ranked candidate crystal structures by relative stability, demonstrating its efficacy.
• AIMNet2 potentials showed strong performance across diverse chemical systems relevant to pharmaceuticals, optoelectronics, and agrochemicals.

Conclusions:

• Target-specific MLIPs offer a significant acceleration of CSP workflows.
• This approach effectively captures the physics of thermodynamic crystal stability using only molecular cluster data, avoiding costly periodic calculations.
• AIMNet2 MLIPs present a promising and efficient alternative to full DFT calculations for routine CSP tasks.