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

This paper introduces XtalOpt Version 14, an enhanced evolutionary multi-objective algorithm that utilizes Pareto optimization and integrates various computational potentials to efficiently predict ground-state crystal structures with variable compositions for functional materials.

The Gist

Imagine you are a master chef trying to discover the perfect new recipe for a dish. You want to find the absolute best version of a meal that is not only delicious (stable) but also has specific qualities, like being low in calories or high in protein (functional properties).

This paper introduces XtalOpt Version 14, a sophisticated computer program that acts like a super-chef's assistant. Its job is to automatically invent, test, and refine millions of potential crystal "recipes" (structures made of atoms) to find the best ones.

Here is how the new version works, explained through simple analogies:

1. The Big Upgrade: Cooking with Variable Ingredients

In the past, this program was like a chef who could only cook a specific dish with a fixed amount of ingredients (e.g., exactly 2 eggs and 1 cup of flour). If you wanted to see what happened with 3 eggs, you had to start a whole new search.

Version 14 is different. It can now cook with variable ingredients. It can mix and match different amounts of elements (like swapping 2 eggs for 3, or 1 cup of flour for 2) to see which combination creates the best dish. It doesn't just look for the perfect "2-egg" cake; it explores the entire pantry to find the best cake, regardless of the exact ratio of ingredients.

2. The "Pareto" Strategy: Finding the Best Compromises

When looking for a new material, you often have conflicting goals. You might want a material that is both super hard and very light. Usually, making something harder makes it heavier.

The new version uses a strategy called Pareto Optimization. Imagine you are shopping for a car. You want it to be fast, cheap, and safe.

• The Old Way: You tried to combine these into a single "score" (e.g., Speed + Cost + Safety = 100 points). This often forced you to pick a "middle-of-the-road" car that wasn't great at anything.
• The New Way (Pareto): The program finds a list of "best-in-class" cars where you can't improve one feature without hurting another. It gives you a menu of top-tier options: "Here is the fastest car," "Here is the cheapest car," and "Here is the safest car." This helps scientists see all the best possible trade-offs without forcing a single, arbitrary choice.

3. The Genetic Kitchen: Mixing and Matching Recipes

The program uses an "evolutionary" approach, similar to how nature evolves species. It starts with a population of random crystal structures and tries to breed the best ones.

• Crossover (Mixing): It takes two parent structures and cuts them up to mix them, like splicing two DNA strands. The new version can now cut the parents in multiple places (like cutting a loaf of bread into many slices and swapping them) to create more diverse offspring.
• New Mutations (The "Permutomic" and "Permucomp" chefs):
  • Permutomic: This is like a chef who randomly adds or removes a single ingredient (an atom) to see if the taste improves.
  • Permucomp: This is a chef who completely changes the recipe's ingredient list (composition) to try something totally new.
  • Note: These new "chefs" only work when the program is allowed to change the ingredient ratios (Variable Composition).

4. Using "AI Taste Buds" (Machine Learning)

Traditionally, testing if a crystal structure is stable required running extremely slow, heavy-duty physics simulations (like using a giant, slow oven to bake every single cake).

XtalOpt 14 now comes with a special interface script that lets the program use Machine Learning Potentials. Think of this as giving the chef "AI taste buds." Instead of baking every cake in a real oven, the AI can instantly predict if a cake will taste good based on its ingredients. This allows the program to test thousands of recipes in the time it used to take to test just a few, making the search for new materials much faster.

5. Keeping the Kitchen Tidy (Similarity Checks)

In a massive search, the program might accidentally create the same recipe twice, or two recipes that are almost identical (like a cake that is just rotated slightly).

The new version has a better similarity check. Instead of just looking at the ingredient list, it looks at the "shape" of the cake. If two structures are too similar (like twins), the program marks them so it doesn't waste time testing the same thing twice. It uses a mathematical "fingerprint" (called a Radial Distribution Function) to tell if two structures are truly different.

6. The "Convex Hull" Map

To know if a recipe is a "winner," the program checks its energy against a map called the Convex Hull.

• Imagine a map where the lowest points represent the most stable, perfect crystals.
• The program calculates how far a new structure is from this "lowest point." If it's very close to the bottom, it's a stable, promising material. If it's high up on a hill, it's unstable and likely to fall apart.

Summary

XtalOpt Version 14 is a powerful, open-source tool that helps scientists discover new materials. It is faster and smarter than before because it:

1. Can mix and match different ingredient ratios (Variable Composition).
2. Finds the best trade-offs between different goals (Pareto Optimization).
3. Uses AI to speed up the testing process (Machine Learning Potentials).
4. Has better tools to avoid repeating the same work (Similarity Checks).

It is designed to help researchers efficiently find the "perfect recipes" for the next generation of functional materials, from better batteries to stronger metals.

Technical Summary

Technical Summary: XtalOpt Version 14

Problem Statement
The discovery of novel functional materials requires efficient navigation of complex potential energy surfaces (PES) across vast chemical and structural spaces. Traditional Crystal Structure Prediction (CSP) methods often rely on fixed-composition searches or single-objective global optimization, which can limit the discovery of (meta)stable phases with variable stoichiometries or specific target properties. While machine learning (ML) interatomic potentials and universal interatomic potentials (UIPs) offer high-throughput capabilities, existing evolutionary algorithms often lack native integration for variable-composition (VC) searches and multi-objective global optimization (MOGO) that utilizes Pareto optimization. Furthermore, managing the computational cost of exploring diverse compositions while maintaining structural diversity and avoiding duplicate evaluations remains a challenge.

Methodology
XtalOpt Version 14 introduces a comprehensive update to its evolutionary multi-objective global optimization framework, specifically designed to handle variable-composition searches. The core workflow involves generating an initial population of structures based on user-defined chemical formulas, performing local structural relaxation, and iteratively applying genetic operations to evolve the population.

Key methodological components include:

• Variable-Composition (VC) Search: The algorithm now natively supports searching across a composition space. Users define chemical systems via a chemicalFormulas input, which can specify fixed compositions, a range of compositions (multi-composition), or a full variable-composition search.
• Pareto Optimization: To address multi-objective problems (e.g., minimizing energy while maximizing a specific property), the code implements the NSGA-II algorithm. This replaces or supplements the previous scalar fitness function. Parent selection is performed via binary tournament selection based on non-dominated sorting (Pareto rank) and crowding distance, ensuring a diverse set of solutions along the Pareto front.
• Enhanced Genetic Operations:
  • Multi-cut Crossover: Allows cutting parent unit cells at multiple points to generate offspring with mixed atomic arrangements.
  • Permutomic: A VC-specific operation that randomly adds or removes a single atom from a parent structure to explore stoichiometric variations.
  • Permucomp: A VC-specific operation that generates a new structure with a random composition and total atom count (up to a user-defined limit) by distorting the parent lattice and decorating it with new atoms.
  • Random Supercell Expansion: A secondary mutation that expands a structure into a supercell and displaces a random atom, useful for exploring charge density waves.
• Convex Hull and Reference Energies: The energetic fitness of structures is determined by their distance above the convex hull. The code calculates this using the Qhull library, requiring user-defined reference energies for elements and relevant subsystems to ensure accurate hull construction for MC and VC searches.
• ML and UIP Integration: The release includes a Python wrapper script (vasp_uip.py) to facilitate local structural relaxation using modern UIPs (MACE, CHGNet, Orb, MatterSim) via the Atomic Simulation Environment (ASE). This allows for rapid screening of thousands of structures, which can be followed by high-accuracy DFT calculations.
• Similarity and Duplicate Checking: The code offers two modes for identifying duplicates: the exact matching scheme (XtalComp) and a new approximate method based on Radial Distribution Functions (RDF) to detect similar structures, particularly useful for molecular systems or slight rotational differences.

Key Contributions

1. Native VC-MO Search: The implementation of a variable-composition search integrated with Pareto optimization, enabling the simultaneous exploration of composition space and multiple material properties.
2. New Genetic Operators: Introduction of permutomic and permucomp operations specifically designed to diversify the population in multi-element, variable-composition searches.
3. Pareto Front Implementation: Integration of the NSGA-II algorithm for parent selection, providing a robust alternative to scalar fitness functions for multi-objective problems.
4. UIP Interface: Provision of easy-to-use scripts to interface XtalOpt with universal interatomic potentials, significantly reducing the computational cost of structural relaxations in high-throughput searches.
5. Refined Input/Output and Controls: Updates to input flag conventions (e.g., chemicalFormulas replacing empiricalFormula), new volume constraints based on covalent radii, and enhanced output files (e.g., hull.txt, movie snapshots) for monitoring convex hull evolution.

Results and Performance
Benchmarking on the Ti-O system demonstrated that the inclusion of permutomic and permucomp operations increases the number of unique chemical compositions produced in a VC search by 30% or more compared to searches without these operators. Computational profiling indicates that while XtalOpt itself consumes a fraction of the total CPU time (e.g., ~35% for 100 structures, decreasing as the number of structures increases due to the dominance of external local optimization tasks), the parent selection process (specifically non-dominated sorting) is the most computationally demanding internal task, consuming approximately 7% of the code's CPU time. The use of fast UIPs for relaxation renders the overhead of XtalOpt's internal tasks negligible compared to the total runtime.

Significance
The authors state that Version 14 of XtalOpt paves the way for efficient ground-state searches for crystalline phases of materials with variable compositions and desired target properties. By integrating Pareto optimization and variable-composition capabilities with modern ML potentials, the code facilitates the high-throughput exploration of chemical systems to predict novel (meta)stable phases. The work emphasizes that these developments, combined with bug fixes and workflow improvements, allow researchers to systematically search for functional materials across complex composition spaces that were previously computationally prohibitive or methodologically difficult to navigate. The paper notes that the software is fully open-source and available via the official website and the Computer Physics Communications Library.