README.md

---
license: cc-by-4.0
tags:
- materials-science
- MACE
- machine-learning-interatomic-potential
- DFT
- VASP
- hydrogen-evolution-reaction
- electrocatalysis
- monolayer-amorphous-carbon
- 2D-materials
library_name: mace
---

# MAC-HER: Fine-tuned MACE MLIP for Hydrogen Adsorption ΔG Prediction on Monolayer Amorphous Carbon

Supplementary data and model for:
**"Harnessing Structural Disorder: Unraveling Hydrogen Evolution in Monolayer Amorphous Carbon via
First-Principles Simulations and Machine-Learned Potentials"**
Sreehari M S, Ashutosh Krishna Amaram, Raghavan Ranganathan — Department of Materials Engineering,
IIT Gandhinagar. [arXiv:2605.07670](https://arxiv.org/abs/2605.07670)

## Model Description
A MACE machine-learning interatomic potential, naively fine-tuned from the MACE-MATPES-PBE0
foundation model, used to predict the Gibbs free energy of hydrogen adsorption (ΔG_H) across the
surface of monolayer amorphous carbon (MAC).

- **Base checkpoint:** MACE-MATPES-PBE0 (medium, 9,063,204 parameters, 89 elements)
- **Fine-tuning strategy:** naive fine-tuning (output head only, no replay), 655,534 parameters updated
- **Radial cutoff:** 6 Å
- **Framework:** mace-torch, PyTorch, ASE

## Training Data
MAC structures (~200 atoms) were generated via LAMMPS melt-quench (ReaxFF) at multiple quench
rates (10 K/ps, 100 K/ps), equilibration temperatures (300 K, 400 K, 500 K), and vacancy concentrations
(3%, 5%, 6%), giving 24 structural protocols. DFT single-point and relaxation calculations (VASP,
PAW-PBE, DFT-D3, 400 eV cutoff) on these configurations and their H-adsorbed counterparts produced
the training set.

- **Total DFT frames:** 1,572 (MAC and MAC–H)
- **Split:** 1,415 train / 157 validation (~10%)
- **Test set:** 175 (MAC and MAC-H), mostly independent static and relaxation-trajectory frames, filtered for non-redundancy
  (no consecutive frames, ≥0.08 eV energy separation)

Files provided in `data/` (train/val/test) and `dft_deltaG_benchmark/` (VASP input/output for the benchmark
adsorption sites).

## Model Performance
| Metric | Value |
|---|---|
| Energy RMSE (test) | 1.65 meV/atom |
| Force RMSE (test) | 29.15 meV/Å |
| ΔG_H MAE vs. DFT (independent benchmark) | 0.161 eV |
| Activity classification accuracy (\|ΔG_H\| < 0.1 eV threshold) | 95% |

## Intended Use
Predicting site-resolved ΔG_H across MAC surfaces (undoped, various vacancy concentrations/quench
rates) as a scalable alternative to exhaustive DFT sampling, and as the basis for extracting local
structural descriptors (coordination number, curvature, ring statistics, hexagonal order, ripple height)
that govern HER activity in monolayer amorphous carbon. Not validated for other amorphous carbon polymorphs,
doped variants, or non-H adsorbates without re-fine-tuning.

## How to Use
This model is loaded via `mace_mp` and used with ASE's structural optimizers. Relaxation follows a
hybrid FIRE→LBFGS scheme: FIRE handles the initial rugged energy landscape down to fmax = 0.1 eV/Å,
then LBFGS refines to the final convergence criterion of fmax = 0.01 eV/Å — this combination was
benchmarked in the paper as the most efficient for MAC's amorphous, noisy energy surface.

```python
from mace.calculators import mace_mp
from ase.io import read, write
from ase.optimize import FIRE, LBFGS
import numpy as np

# Load structure
atoms = read('dft_deltaG_benchmark/4%_vacancy_100kps_QR/site_001/POSCAR', format='vasp')
atoms.set_pbc(True)

# Load fine-tuned MACE model
atoms.calc = mace_mp(
    model="model/6_MACE_MAC_HER.model",
    default_dtype="float64",
    device="cuda",  # or "cpu"
)

print("Initial energy:", atoms.get_potential_energy(), "eV")
print("Initial max force:", np.max(np.abs(atoms.get_forces())), "eV/Å")

# Stage 1: FIRE — coarse relaxation of the rugged landscape
fire = FIRE(atoms, logfile=None, dt=0.05, maxstep=0.05, dtmax=0.5, Nmin=10, finc=1.05, fdec=0.5)
fire.run(fmax=0.1)

# Stage 2: LBFGS — fine convergence
lbfgs = LBFGS(atoms, logfile=None, maxstep=0.05, memory=100)
lbfgs.run(fmax=0.01)

print("Final energy:", atoms.get_potential_energy(), "eV")
write('CONTCAR-relaxed', atoms, format='vasp', direct=True)
```

ΔG_H for a given site is then obtained from separate relaxations of the pristine MAC surface and the
H-adsorbed configuration, following the standard adsorption free-energy expression (see paper Eq. 2–3).

## Files in This Repository
- `model/` — fine-tuned MACE checkpoint
- `dft_deltaG_benchmark/` — VASP input/output files for benchmark adsorption sites (isolated local environments and
  full-surface relaxations)
- `data/` — train/val/test splits (1,572 DFT frames)
- `scripts/` — feature extraction (local environment descriptors: coordination number, curvature,
  ring statistics, hexagonal order, ripple height), MACE structure relaxation using ASE module and python notebook for feature analysis.

## Citation
@misc{s2026harnessingstructuraldisorderunraveling,
      title={Harnessing Structural Disorder: Unraveling Hydrogen Evolution in Monolayer Amorphous Carbon via First-Principles Simulations and Machine-Learned Potentials}, 
      author={Sreehari M S and Ashutosh Krishna Amaram and Raghavan Ranganathan},
      year={2026},
      eprint={2605.07670},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2605.07670}, 
}