Training Data (Unit-Corrected)

This directory contains molecular data with corrected units ready for DCMnet/PhysnetJax training.

Data Corrections Applied

1. Atomic Coordinates (R)

Original: Angstroms (verified)
Status: ✓ Correct
Units: Angstrom (ASE standard)

2. Energies (E)

Original: Hartree
Converted: eV (ASE standard)
Factor: ×27.211386

3. Forces (F)

Original: Hartree/Bohr
Converted: eV/Angstrom (ASE standard)
Factor: ×51.42208

4. ESP Grid Coordinates (vdw_surface)

Original: Grid index space
Fixed: Physical Angstroms
Conversion: Applied proper grid spacing (0.25 Bohr = 0.132294 Å)

Data Splits

Train: 800 samples (80%)
Valid: 100 samples (10%)
Test: 101 samples (10%)
Seed: 42 (reproducible)

Files

Energy, Forces, and Dipoles

energies_forces_dipoles_train.npz
energies_forces_dipoles_valid.npz
energies_forces_dipoles_test.npz

Each contains:

R: Atomic coordinates [Angstrom]
Z: Atomic numbers [int]
N: Number of atoms [int]
E: Energies [eV] ← CONVERTED from Hartree
F: Forces [eV/Angstrom] ← CONVERTED from Hartree/Bohr
Dxyz: Dipole moments [e·Å] ← CONVERTED from Debye

ESP Grids

grids_esp_train.npz
grids_esp_valid.npz
grids_esp_test.npz

Each contains:

R: Atomic coordinates [Angstrom]
Z: Atomic numbers [int]
N: Number of atoms [int]
esp: ESP values [Hartree/e]
vdw_surface: Grid coordinates [Angstrom] ← FIXED
vdw_grid: Same as vdw_surface (backward compatibility)
grid_dims: Original cube dimensions (if available)
grid_origin: Original cube origins [Bohr] (if available)
grid_axes: Original cube axes (if available)
Dxyz: Dipole moments [e·Å] ← CONVERTED from Debye

Units Summary (ASE Standard)

Property	Unit	Status
R (coordinates)	Angstrom	✓ Correct
E (energy)	eV	✓ Converted
F (forces)	eV/Angstrom	✓ Converted
Dxyz (dipoles)	e·Å	✓ Converted from Debye
esp (values)	Hartree/e	✓ Correct
vdw_surface	Angstrom	✓ Fixed

Usage

import numpy as np

# Load training data
train_props = np.load('energies_forces_dipoles_train.npz')

# All units are ASE-standard - ready to use!
R = train_props['R']  # Angstroms
E = train_props['E']  # eV
F = train_props['F']  # eV/Angstrom
Dxyz = train_props['Dxyz']  # e·Å (converted from Debye)

# Load grid data (if available)
train_grids = np.load('grids_esp_train.npz')
esp = train_grids['esp']  # Hartree/e
vdw_surface = train_grids['vdw_surface']  # Angstroms

Generated by: mmml.cli.fix_and_split