Li–Sn Phase Discovery

Neural-network potentials + large-scale structure search to accelerate DFT-grade materials discovery

Overview

This project combines a neural-network interatomic potential (NNP) with evolutionary structure prediction to explore the Li–Sn alloy system at a scale that would be impractical with DFT alone. The core idea is a two-tier workflow: use the NNP to perform high-throughput relaxations and ranking over a massive candidate set, then validate only the most promising structures with targeted DFT re-optimization and phonon calculations.

In practice, this meant building a robust training dataset of DFT energies and forces, training and iterating on an NNP that stays stable inside geometry-optimization loops, and deploying it for large-scale screening (on the order of a million relaxations) with systematic escalation to DFT for final decisions.

Approach

Data generation & curation

  • Generated DFT reference data (energies + forces) across Li–Sn compositions and diverse structural motifs (cells up to ~32 atoms).
  • Managed a large training corpus (over 1 TB of energy–force data accumulated through iterative expansion).
  • Used elemental Li and Sn models as a base and trained a dedicated binary model on curated binary structures.

Model & training

  • Model type: Behler–Parrinello–style NNP (atomic environment descriptors → atomic energies → total energy).
  • Binary architecture: 145–10–10–1 (~1880 parameters).
  • Descriptors: symmetry functions with a 7.5 Å cutoff.
  • Training target: energies and forces (forces strongly improve relaxation behavior and local geometry fidelity).
  • Held-out accuracy (reported): ~10 meV/atom energy RMSE and ~49 meV/Å force RMSE.

Using the model in the loop (screening + verification)

A surrogate model is most useful when it behaves well during relaxations and preserves ranking quality near the stability boundary. The workflow was built around that constraint:

  • Run evolutionary searches at fixed compositions to generate candidate pools.
  • Perform the bulk of local relaxations with the NNP (high throughput).
  • Re-optimize a filtered subset with DFT (high precision).
  • Audit near-hull candidates carefully, where small errors can change convex-hull membership.

Scale

  • NNP relaxations: ~1.16 million local relaxations (≈952k CPU hours).
  • DFT re-optimizations: 4450 selected structures.
  • Estimated DFT-only equivalent: ~100 million CPU hours.

This split made it feasible to explore broad composition/structure space while keeping final stability claims anchored in DFT.

Finite-temperature stability (phonon workflow)

To go beyond 0 K enthalpies, we incorporated vibrational contributions using a staged strategy:

  1. Compute NNP vibrational free energies for low-energy candidate sets (within ~20 meV/atom).
  2. Build a finite-T hull at 600 K using G = H(DFT) + ΔF_vib(NNP).
  3. Promote only the most competitive finalists (within ~10 meV/atom of tie-lines) to DFT phonons.
  4. Evaluate stability across 0–800 K (10 K steps), removing dynamically unstable phases.

Scale:

  • NNP phonons: 2300 structures
  • DFT phonons: 233 structures

Key Results

Ambient pressure (0 GPa)

  • Identified overlooked low-energy phases that refine the convex hull, including a new stable hR48–Li₃Sn candidate.
  • Predicted a competitive mS40–LiSn₄ polymorph that becomes stable at elevated temperature in the Gibbs analysis.
  • Constructed and validated a stable Li-rich BCC-derived phase hR75–Li₁₉Sn₆ extremely close to the tie-line.

High pressure (20 GPa)

At 20 GPa, the search yielded revised Li-rich stability with new/updated ground states:

  • aP26–Li₁₁Sn₂ as the most Li-rich ground state (with a near-degenerate polymorph).
  • Improved mS48–Li₅Sn configuration.
  • Newly stable mS44–Li₉Sn₂.

Tools & Stack

  • NN training & iteration: In-house developed NN software MAISE for training loops. MAISE-NET for automated dataset generation and active learning and hyperparameter tuning.
  • Hardware: HPC clusters with Intel Xeon CPUs for NN training and CPU nodes for DFT calculations. All using slurm workload manager.
  • DFT: VASP
  • Global optimization: evolutionary structure prediction driven by NNP relaxations
  • Phonons: NNP screening + targeted DFT phonons

Publication

Kharabadze, S., Thorn, A., Koulakova, E. A., & Kolmogorov, A. N. (2022). Prediction of stable Li-Sn compounds: boosting ab initio searches with neural network potentials. npj Computational Materials, 8, 136.