XtalOpt Version 14: Variable-Composition Crystal Structure Search for Functional Materials Through Pareto Optimization

TL;DR

XtalOpt version 14 introduces Pareto-based multi-objective optimization and native variable-composition search to advance crystal structure prediction across compositional spaces. The workflow integrates a VC-MO framework, diverse genetic operations including multi-cut crossover and VC-specific mutations, and reference-energy-guided hull calculations to robustly identify metastable phases. Key contributions include NSGA-II-based Pareto parent selection, RDF-based similarity checks, and seamless integration of universal interatomic potentials for fast local relaxations via UIP interfaces. The updated GUI, seed-structure handling, and molecular-unit construction support high-throughput exploration of functional materials, enabling efficient discovery of targeted crystal structures with controlled size and composition. These enhancements collectively enable efficient exploration of complex composition spaces, improve the reliability of hull-based energetics, and provide practical tooling for accelerating materials discovery with advanced relaxation engines.

Abstract

Version 14 of XtalOpt, an evolutionary multi-objective global optimization algorithm for crystal structure prediction, is now available for download from its official website https://xtalopt.github.io, and the Computer Physics Communications Library. The new version of the code is designed to perform a ground state search for crystal structures with variable compositions by integrating a suite of ab initio methods alongside classical and machine-learning potentials for structural relaxation. The multi-objective search framework has been enhanced through the introduction of Pareto optimization, enabling efficient discovery of functional materials. Herein, we describe the newly implemented methodologies, provide detailed instructions for their use, and present an overview of additional improvements included in the latest version of the code.

Figures (5)

• Figure 1: The workflow in XtalOpt version 14, which offers a variable-composition multi-objective (VC-MO) search. Post relaxation, each structure's energies are used to obtain its distance above the convex hull, which is combined with the values of the user-specified objectives to determine the structure's suitability for parenting new offspring. Genetic operations are then applied to the selected parents to create new structures with compositions determined by the search type.
• Figure 2: Schematic illustration of the new genetic operations implemented in XtalOpt version 14, applicable only to the VC search. (a) Permutomic produces an offspring by randomly adding or removing an atom to the parent structure, followed by a small random lattice distortion. (b) Permucomp is designed to diversify the population through producing a new structure by distorting the parent lattice and decorating it with a new random composition.
• Figure 3: The secondary mutation "random supercell expansion" is applicable to the offspring produced by any of the XtalOpt genetic operations, and produces a random supercell in which an atom is randomly selected and displaced. The default probability of this mutation is 0%.
• Figure 4: The graphical user interface (GUI) of XtalOpt version 14 has been updated to include the new features: (a) the "Structure Limits" tab where the chemical system, search type, reference energies, maximum number of atoms, various unit cell limits and cell initialization schemes for the run can be set; (b) the "Search Settings" tab entries can be used to specify the search duration and run concurrency, action taken when a local optimization fails, initial seed structures, genetic operation specifications, and thresholds for symmetry and similarity detections; (c) the "Multiobjective Search" tab parameters allow for adjusting the global optimization type and its relevant settings, adding objectives for a MO run, and specifying the code's actions when an external program fails to calculate a filtration objective, and (d) the "Progress" tab maintains a live overview of the produced structures with a summary of their characteristics, such as the chemical formula, enthalpy, distance above the convex hull, Pareto front, symmetry, and ancestry information.
• Figure 5: An example of the convex hull produced by the pycxl code, from the output of the XtalOpt VC run for the Ti-O system.