Accurate and efficient protocols for high-throughput first-principles materials simulations

Technical Summary: Accurate and Efficient Protocols for High-Throughput First-Principles Materials Simulations

Problem Statement
The advent of high-throughput materials discovery relies on automated first-principles simulations, particularly Density-Functional Theory (DFT). A critical bottleneck in scaling these workflows to tens of thousands of compounds is the reliable selection of simulation parameters that balance numerical precision with computational efficiency. Specifically, the interplay between k-point sampling (discretization of the Brillouin zone) and smearing techniques (used to handle electronic occupation discontinuities at the Fermi level, especially in metals) is complex.

• Sampling Errors: In metallic systems, poor k-point sampling leads to "statistical" errors where results fluctuate randomly with mesh density, requiring prohibitively dense meshes for convergence.
• Systematic Errors: Smearing introduces a fictitious electronic temperature ($\sigma$) to smooth occupation functions, accelerating k-point convergence. However, large $\sigma$ values introduce "systematic" errors by deviating the total free energy from the physical $\sigma \to 0$ limit due to entropic contributions.
  Current workflows often use fixed parameters or do not optimize smearing, failing to control the trade-off between suppressing sampling errors and minimizing systematic errors.

Methodology
The authors developed a rigorous benchmarking framework to define Standard Solid-State Protocols (SSSP) for selecting optimized k-point and smearing parameters.

1. Dataset: 269 crystal structures (3D and 2D) were sampled from the Materials Cloud databases. To avoid symmetry-induced vanishing forces, a random atom in each unit cell was displaced by 5% of its bond length.
2. Workflow: Automated DFT calculations were performed using Quantum ESPRESSO (v7.0) with the SSSP PBEsol Efficiency 1.2 pseudopotential library.
   • Variables: The k-point sampling density was parameterized by $\lambda$ (maximum distance between adjacent k-points in reciprocal space, in $\mathring{A}^{-1}$). The smearing temperature ($\sigma$) ranged from 0.002 to 0.03 Ry using Marzari-Vanderbilt cold smearing.
   • Metrics: Total energies ($E$), forces on the displaced atom ($F$), and pseudo-stresses ($\Sigma$) were computed.
3. Error Analysis: The study decoupled errors into two distinct categories:
   • Sampling Error ($\varepsilon^k$): The deviation of a property calculated with a specific $(\lambda, \sigma)$ from the value obtained with the densest k-point mesh ($\lambda_{ref}$) at the same $\sigma$.
   • Systematic Error ($\varepsilon^s$): The deviation of a property calculated with a specific $\sigma$ (at $\lambda_{ref}$) from the reference value at the lowest possible $\sigma$ ($\sigma_{ref}$).
4. Optimization Strategy: The authors sought a "Pareto front" of parameters that minimizes computational wall time for a given average error. Crucially, they imposed a constraint that sampling errors must be suppressed to less than 5% of systematic errors ($\varepsilon^k < 0.05\varepsilon^s$). This ensures that the less controllable sampling errors do not dominate the total error budget, allowing the systematic error (controlled by $\sigma$) to be the primary variable for tuning precision vs. cost.

Key Contributions

1. Decoupled Error Analysis: The paper establishes a methodology to separately quantify and control statistical (sampling) and systematic (smearing) errors, preventing them from canceling out by chance and masking convergence issues.
2. Definition of SSSP Protocols: Three distinct protocols are defined based on the Pareto optimization, tailored for different precision/efficiency trade-offs:
   • Fast: Optimized for testing and preliminary relaxations. Uses coarse meshes and higher smearing.
   • Balanced: The recommended choice for high-throughput screening and general applications.
   • Stringent: The highest precision choice before computational costs rise sharply.
3. Material-Specific Guidelines:
   • Insulators: Benefit significantly from the Balanced protocol, as they suffer less from sampling errors and can utilize coarser meshes.
   • Metals with f-shell elements (Lanthanides/Actinides): Exhibit high systematic errors due to sharp peaks in the density of states near the Fermi level. These require the Stringent protocol (lower $\sigma$, denser meshes) to maintain accuracy.
4. Software Integration: The protocols are integrated into the AiiDA workflow framework and the Quantum ESPRESSO input generator on Materials Cloud, allowing users to automatically generate inputs based on the material class and desired protocol.

Results

• Convergence Behavior: The study confirms that increasing smearing temperature effectively suppresses sampling errors, allowing for coarser k-point meshes. However, an optimal "sweet spot" exists where the sum of sampling and systematic errors is minimized for a given computational cost.
• Lanthanide Sensitivity: Materials containing lanthanides show significantly higher systematic errors at standard smearing temperatures compared to other metals. The Stringent protocol is necessary to keep errors within acceptable thresholds for these systems.
• Protocol Performance:
  • The Balanced protocol yields average force errors of ~20.9 meV/Å for metals and is ~2.7x faster than the Stringent protocol (which is 2.7x slower than Balanced).
  • The Stringent protocol reduces average force errors to ~11.5 meV/Å for metals.
  • For insulators, the Balanced protocol is sufficient, as running the Stringent protocol would imply a computational cost 1.82x higher with only minor improvements in errors.
• Transferability: Benchmarks on derived properties (lattice constants and phonon frequencies) indicate that the errors in these "higher-order" properties correlate directly with the force errors, validating the use of forces as the primary convergence metric.

Significance and Claims
The authors claim that the SSSP protocols provide a robust, automated solution for setting DFT parameters, removing the need for manual, case-by-case convergence testing. By abstracting the complexity of k-point and smearing selection, these protocols:

• Enable reproducible and consistent high-throughput workflows.
• Prevent the waste of computational resources on over-converged calculations or the generation of unreliable data from under-converged ones.
• Are designed to be code-agnostic in principle (though implemented for Quantum ESPRESSO), with the numerical values serving as a first approximation for other plane-wave codes (e.g., VASP, Abinit) and potentially all-electron codes that utilize smearing.
• Represent a foundational step toward interoperable software ecosystems where workflows can be treated as modular building blocks for materials design.

The paper emphasizes that while these protocols are optimized for self-consistent DFT calculations of total energies, forces, and stresses, additional convergence tests (e.g., for q-grids in phonon calculations) may be required for specific linear-response properties.