On generating Special Quasirandom Structures: Optimization for the DFT computational efficiency

We present our novel evolutionary algorithm for generating Special Quasirandom Structures (SQS) designed to optimize the computational efficiency of Density Functional Theory (DFT) computations. Operating on the premise that symmetry proxies non-randomness, we rigorously filter out 1.P1 candidate structures prior to evaluating correlation functions. Our extinction-based workflow includes the seeding, filtration, evaluation, extinction, and repopulation phases to produce efficient supercells with maximal local environmental distinctness. We compare our results against those generated by established software packages, on the example of the W\textsubscript{70}Cr\textsubscript{30} alloy. Although standard tools achieve (marginally) lower correlation errors, our best-performing structures require approximately five times fewer unique displacements for phonon calculations. This approach sacrifices negligible quantitative disorder accuracy to significantly reduce the computational cost of modeling thermal properties.

💡 Research Summary

The paper introduces a novel evolutionary algorithm for generating Special Quasirandom Structures (SQS) that explicitly targets the reduction of computational cost in Density Functional Theory (DFT) simulations. The central premise is that crystallographic symmetry is a proxy for non‑randomness; therefore, any candidate supercell possessing the highest symmetry (space group P1) is discarded before any expensive correlation‑function evaluation. The workflow consists of five stages: (1) seeding random supercells, (2) filtering out high‑symmetry (P1) cells, (3) evaluating a composite error metric that combines cluster‑correlation mismatches with the number of unique atomic displacements required for phonon calculations, (4) extinction of the worst‑performing structures, and (5) repopulation via mutation and crossover of the survivors. A Metropolis acceptance criterion is employed to allow occasional uphill moves, preserving population diversity.

The error metric is defined as
E = Σₙ Aₙ |ΔCₙ| · N_disp,
where ΔCₙ denotes the deviation of the n‑th cluster correlation from the ideal random alloy, Aₙ is a weight (chosen as 1/n in the presented work), and N_disp is the count of symmetry‑inequivalent atomic displacements needed for phonon calculations. By multiplying the traditional correlation error with N_disp, the algorithm preferentially selects compact supercells that still reproduce the local chemical environment but require far fewer force‑constant evaluations.

The method is benchmarked on a W₇₀Cr₃₀ alloy modeled with a 4 × 4 × 4 bcc supercell (128 atoms). Comparisons are made against structures generated by ATAT and sqsgenerator after 10⁹ Monte‑Carlo steps, which typically yield P1‑symmetry cells with minimal correlation error (E ≈ 2.01 × 10⁻⁴). The best structures from the new algorithm belong to lower‑symmetry groups (C2, Amm2) and achieve a comparable correlation error (E ≈ 2.92 × 10⁻⁴) while reducing the number of unique displacements by a factor of five. Consequently, phonon calculations—normally the most expensive part of a DFT workflow—become five times faster without a significant loss of disorder fidelity.

Additional tests of alternative weight schemes (Aₙ = n⁻ᵖ with p = 2, 4, 12 and distance‑based weights) confirm that the simple 1/n choice offers the best trade‑off between convergence speed and final structure quality. The extinction‑based selection mechanism effectively eliminates high‑symmetry candidates early, dramatically shrinking the search space and allowing the evolutionary loop to focus on configurations that maximize local environmental distinctness.

In the discussion, the authors argue that the modest increase in correlation error is acceptable for most practical applications, especially when the goal is to compute thermodynamic properties (e.g., free energies via the quasiharmonic approximation) or to explore phase stability through convex‑hull constructions. The ability to generate computationally cheap yet physically meaningful SQSs opens the door to systematic high‑throughput studies of complex alloys, high‑entropy materials, and refractory systems where DFT resources are a limiting factor.

The paper concludes that integrating symmetry‑based filtering with a displacement‑aware error function constitutes a new paradigm for SQS generation. It balances the competing demands of statistical representativeness and computational tractability, offering a practical tool for the materials‑science community engaged in ab‑initio modeling of disordered alloys. Future work may extend the approach to multi‑component high‑entropy alloys, non‑cubic lattices, and hybrid schemes that combine the evolutionary engine with machine‑learning potentials for even larger configurational spaces.