Elemental Energies

Goal

To provide a centralized library of the most stable phases (ground states) for each element and their corresponding reference energies calculated using different Machine Learning Interatomic Potentials (MLIPs). This avoids redundant relaxations and ensure consistency in thermodynamics calculations (e.g., formation energy, stability).

Instructions

1. Retrieve Elemental Energies

To get the energies for a list of elements from a specific checkpoint:

# Env: base-agent
python .agents/skills/mat-elemental-energies/scripts/get_elemental_energies.py --elements Li Fe O --checkpoint mace-mp-medium

Library Status

As of 2026-01-30, the library is fully expanded and contains 89 elements (Ground states from H to Lr) across 36 MLIP variants, including:

• MACE: MACE-MP (S/M/L), MACE-MH-0/1 (all heads), MACE-OMAT (S/M), MACE-MATPES (PBE/R2SCAN).
• MatGL: CHGNet (MPtrj, MatPES), M3GNet (MP, MatPES), TensorNet (MatPES).
• FairChem: UMA (S/M) with all heads (omat, omol, oc20).

Constraints

• Materials Project: Structures must be queried from Materials Project to ensure they represent the true ground-state phases.
• Naming: Library files must follow the <checkpoint_name>_energies.json format.
• Units: Energies are stored as eV/atom (total potential energy of the relaxed structure divided by the number of atoms).
• Relaxation: The stored value comes from relaxing both cell and coordinates (relax_cell=True). A positions-only relaxation leaves the reference above the potential's own minimum and biases every formation energy that uses it.
• Molecular-crystal elements are sensitive: for O, N, F, Cl, H the MP ground state is a van der Waals-bound molecular crystal whose MLIP-relaxed cell differs a lot from the DFT cell, so the cell relaxation matters more here than for metals. With TensorNet-PES-MatPES-PBE-2025.2, mp-12957 (O2) gives -5.118 eV/atom fully relaxed versus -4.968 positions-only. More than one MP entry can also tie at energy_above_hull = 0 for these elements (O2: mp-12957 and mp-1524462), so pin the material_id rather than relying on a formula lookup returning a stable order.

---

Author: Bowen Deng Contact: GitHub @learningmatter-mit