Examples for LightPFP: A Lightweight Route to Ab Initio Accuracy at Scale

This repository contains all reproduction examples accompanying the paper LightPFP: A Lightweight Route to Ab Initio Accuracy at Scale.

---

examples/maintext

Each subdirectory corresponds to one of the case studies presented in the Results section of the paper maintext.

examples/maintext
├── pretrained-students/ # Data efficiency of pretrained student models
├── Li6PS5Cl/            # Li₆PS₅Cl
├── HEA/                 # High entropy alloy
├── SiO2-etching/        # Dry etching of SiO2: application of active learning
└── MgO/                 # Melting point of MgO: application of few-shot transfer learning

---

examples/supp

Each subdirectory corresponds to one of the case studies presented in Supplementary Information II of the paper.

examples/supp
├── 1_Pt_benzene/                  # S1. Simulation of interfacial structures of Pt(111)/benzene
├── 2_miscibility/                 # S2. Miscibility of water, benzene, and heptane
├── 3_interface_thermal_resistence/# S3. Interface thermal resistance between Ni and DPO/BP
├── 4_viscosity/                   # S4. Viscosity of n-decane
├── 5_crack_propagation/           # S5. Crack propagation in graphene nanoribbon
├── 6_lubrication_friction/        # S6. Friction of Fe₂O₃ surface with lubricant and fatty acid surfactant
├── 7_polymer_ionic_liquid/        # S7. Diffusion behavior in polymer ionic liquid
├── 8_glass/                       # S8. Mechanical property of SiO₂–P₂O₅–Al₂O₃–Na₂O glass
├── 9_grain_boundary/              # S9. Heterogeneous grain boundary between FCC Cu and BCC Mo
├── 10_micelle_formation/          # S10. Micelle formation
└── 11_chemical_mechanic_polishing/# S11. Chemical mechanical polishing of Si surface

---

Note

How to open dataset files?

The datasets are stored in HDF5 format. You can open them as follows:

import h5py

path = "examples/maintext/HEA/datasets/LightPFP/HEA_crystal.h5"
with h5py.File(path, "r") as datasets:
    print(list(datasets.keys())) # ID of structure
    # -> ['000000', '000001', '000002', ...]
    print(list(datasets["000001"].keys()))
    # -> ['cell', 'forces', 'numbers', 'pbc', 'positions', 'potential_energy', 'stress']
    print(datasets["000001"]["positions"][:])
    # -> array([[0. , 0. , 0. ], ..., [5.4, 5.4, 3.6]])

---

Citation

If you use these examples or the LightPFP methodology in your work, please cite:

LightPFP: A Lightweight Route to Ab Initio Accuracy at Scale arXiv:2510.23064

---

Execution Environment

The examples were developed and tested using LightPFP on Matlantis Product. To set up the execution environment, please contact Matlantis Corporation.