Bridgmanite Thermoelastic Eval

bridgmanite-thermoelastic-eval

Thermoelastic properties of bridgmanite using Deep Potential Molecular Dynamics — Wan et al. (2023) (arXiv:2307.07127, 2023)

What this evaluates

Evaluates the accuracy of deep-learning molecular dynamics potentials in predicting the structural, thermodynamic, and elastic properties of bridgmanite (MgSiO3-perovskite) under high-pressure and high-temperature conditions relevant to Earth's lower mantle.

Datasets

• DFT reference dataset — total ?; splits: train (-1), val (-1)
• Experimental benchmarks — total ?; splits: test (-1)
• PREM seismic model — total ?; splits: reference (-1)

Metrics

• RMSE (primary) — range: other
  • Root-mean-square error computed between deep potential predictions and DFT reference values for potential energy (meV/atom) and atomic forces (eV/Å).
• EOS & Elastic Moduli Accuracy — range: other
  • Quantitative comparison of predicted equation of state (Birch-Murnaghan fit), elastic coefficients (c_ij), bulk/shear moduli, and wave velocities against experimental measurements and PREM seismic data.

Input / output format

Input: Atomic coordinates, simulation cell parameters, and target pressure-temperature conditions for MgSiO3 bridgmanite.

Output: Predicted potential energies, atomic forces, radial distribution functions, phonon dispersions, equation of state parameters, elastic stiffness tensor (c_ij), and derived macroscopic moduli/velocities.

Scoring recipe

def evaluate(dp_model, config, gold):
    # 1. Compute RMSE for energy and forces
    rmse_e = np.sqrt(np.mean((dp_model.predict_energy(config) - gold['energy'])**2))
    rmse_f = np.sqrt(np.mean((dp_model.predict_force(config) - gold['force'])**2))
    # 2. Structural & Phonon validation
    g_r = compute_radial_distribution(dp_model.run_md(config))
    phonons = compute_phonon_dispersion(dp_model.run_md(config))
    # 3. Thermodynamic & Elastic properties
    eos_params = fit_birch_murnaghan(dp_model.compute_PV(config))
    elastic_tensor = compute_elastic_tensor(dp_model.compute_stress_strain(config))
    K_S, G = voigt_reuss_hill(elastic_tensor)
    V_p = np.sqrt((K_S + 4*G/3) / dp_model.density)
    V_s = np.sqrt(G / dp_model.density)
    return {'rmse_energy': rmse_e, 'rmse_force': rmse_f, 'eos': eos_params, 'moduli': (K_S, G), 'velocities': (V_p, V_s)}

Common pitfalls

• Classical MD neglects zero-point motion (ZPM), causing systematic underestimation of pressure and volume at low temperatures (e.g., 300 K).
• Isothermal elastic coefficients are reported, but adiabatic corrections may be required for direct comparison with seismic data at high temperatures.
• PBE and PBEsol functionals systematically overestimate volume and deviate ~12.5% from experimental elastic data, making them unsuitable for this system.

Evidence (verbatim from paper)

We quantified the root-mean-square error (RMSE) for the DP's predictions on the validation sets. Fig. 1 shows the outcomes of this analysis for the DP-LDA calculation. The RMSE for potential energies amounted to approximately 0.82meV / atom, while the RMSE for atomic force prediction errors was around 0.07eV / A.

Citation

@misc{wan2023thermoelastic,
  title={Thermoelastic properties of bridgmanite using Deep Potential Molecular Dynamics},
  author={Wan et al. (2023)},
  year={2023},
  note={arXiv:2307.07127}
}

• arXiv: 2307.07127