PNcsp+: A Periodic Number-Based Crystal Structure Prediction Method Enhanced by Machine Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Crystallography

Background:

• Crystal structure prediction (CSP) is crucial for materials discovery but faces challenges in efficiency and interpretability due to vast configurational spaces and expensive optimizations.
• Existing template-based and machine learning (ML) methods often require large datasets, complex similarity metrics, or opaque pipelines.

Purpose of the Study:

• To introduce PNcsp+, an enhanced and chemically interpretable CSP framework.
• To leverage the Mendeleev Periodic Number (PN) as a transparent descriptor for elemental similarity.
• To improve the efficiency and accuracy of crystal structure prediction.

Main Methods:

• Developed PNcsp+ with an expanded prototype library and improved data management.
• Integrated ML-assisted prototype scoring using neural network models (MACE, M3GNet, ALIGNN-FF).
• Utilized the Periodic Number (PN) for chemically interpretable elemental similarity assessments.

Main Results:

• Achieved state-of-the-art performance on the CSPBench dataset, surpassing alternative methods.
• Reached 86.1% space group accuracy and 85.0% structure matching accuracy in Top-5 predictions without relaxations.
• Demonstrated autonomous emergence of molecular components in hybrid systems guided by PN-derived similarities.

Conclusions:

• PNcsp+ offers an efficient, scalable, and interpretable CSP framework by combining periodic trends with targeted ML evaluation.
• The approach accelerates discovery in both inorganic and hybrid chemical spaces.
• Highlights the power of fundamental chemical principles in modern computational materials science.