Analytic nuclear gradients for machine-learned orbital-free DFT
This repository provides four SCIAI-DFT models for density optimization and geometry optimization with machine-learned orbital-free density functional theory (ML-OFDFT).
| Directory | Model | Description |
|---|---|---|
Graphormer/ |
Graphormer | Graphormer base model |
Graphormer-NGL/ |
Graphormer NGL | Graphormer fine-tuned with the nuclear gradient loss |
eSEN/ |
eSEN | eSEN base model |
eSEN-NGL/ |
eSEN NGL | eSEN fine-tuned with the nuclear gradient loss |
Each directory contains safetensors weights, a self-contained model.yaml
inference configuration, and the original hparams.yaml and
hparams_resolved.yaml files. SHA256SUMS records the checksums of all four
weight files.
Usage
Install the code from
sciai-lab/structures25-nuclear-gradients,
then load a checkpoint with the project helper:
from pathlib import Path
from huggingface_hub import snapshot_download
from mldft.utils.instantiators import instantiate_model
repo = Path(snapshot_download("sciai-lab/structures25-nuclear-gradients"))
model = instantiate_model(repo / "eSEN-NGL" / "eSEN-NGL.safetensors", device="cpu")
The resolved hyperparameter files provide the model-specific data transforms and density-optimization settings needed by the project evaluation scripts.
Scope and limitations
These models predict the learned kinetic-plus-exchange-correlation contribution from atom-centered density coefficients and molecular geometry. They are intended for ML-OFDFT calculations on neutral organic molecules in the chemical regime represented by the training data. They are not general-purpose interatomic potentials. Strongly distorted, reactive, dissociative, or connectivity-changing structures remain challenging.