Accurate Chemically-Sensitive-Space Factors for Enhanced High-Pressure Crystal Structure Prediction: A Leap beyond Machine Learning

We developed a hierarchical iterative random sampling statistics (HIRSS) method for accurate and efficient high-pressure structure prediction. The HIRSS can effectively explore chemically sensitive space and identify crystal structures that satisfy certain chemical constraints, namely stoichiometry, coordination numbers, bond lengths, and angles. Other factors, such as crystal densities, symmetries of space groups, lattice systems, and neighbouring relations of chemical functional groups, are used to define the boundaries of the chemical space and guide the search. Thus, accurate factors of the chemically sensitive space can be used as a robust fingerprint of random sampling statistics to extract crystal structures with three-dimensional atomic coordinates. Our results indicate that predicted crystal structures using chemically sensitive space factors through HIRSS have significantly reduced sampling errors, thereby achieving accurate and efficient high-pressure structure prediction for a small number of samples. For complex dense silanes SH₄ at high pressure, HIRSS yielded the two lowest enthalpy structures from only 20 crystal structure samples. The predicted structures matched results from the machine learning method using data-derived potentials trained from more than 6,500 structure samples. Additionally, HIRSS structure predictions revealed a new structure with comparable enthalpies that was not discovered by the previously mentioned machine learning method.