Torch-DFTD

0.5.3 · maintenance · verified Thu Apr 16

torch-dftd is a Python library providing a PyTorch implementation of the DFT-D2 and DFT-D3 dispersion correction models for quantum chemistry calculations. It leverages PyTorch's automatic differentiation capabilities to efficiently compute energies, forces, and virials. The library is currently at version 0.5.3, released on March 11, 2026. While it has seen recent updates, its GitHub repository indicates that it is "not actively maintained".

Common errors

```
ModuleNotFoundError: No module named 'ase'
```
cause The Atomic Simulation Environment (ASE) library, a core dependency for creating and manipulating atomic structures used by torch-dftd's examples and common workflows, is not installed.

fix Install ASE via pip: `pip install ase`
```
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0!
```
cause This is a common PyTorch error indicating that tensors involved in a computation are on different devices (e.g., one on CPU, another on GPU). The `TorchDFTD3Calculator` can be initialized with a device, but input data or other tensors might not be moved accordingly.

fix Ensure all relevant tensors (e.g., atomic positions if handled manually) and the `TorchDFTD3Calculator` instance are explicitly moved to the same device (e.g., `device='cuda:0'` for the calculator and `tensor.to(device)` for tensors). The quickstart handles this for the calculator.
```
AttributeError: 'Atoms' object has no attribute 'get_potential_energy'
```
cause The `get_potential_energy()` or `get_forces()` methods are called on an `ase.Atoms` object before a calculator (like `TorchDFTD3Calculator`) has been attached to it using `atoms.set_calculator(calc)`.

fix Ensure `atoms.set_calculator(calc)` is called after initializing `TorchDFTD3Calculator` and before attempting to retrieve energies or forces from the `atoms` object.

Warnings

maintenance The official GitHub repository for torch-dftd explicitly states that the project is 'not actively maintained'. While releases might still occur, active development, bug fixes for new issues, or compatibility updates for future PyTorch/Python versions may be limited.
Fix: Be aware that long-term support may be minimal. Consider forking the repository if deep customization or active maintenance is critical for your project.
gotcha The library is tested against specific Python (3.10) and CUDA (12.2) versions, and specific torch (2.0.1) and ase (3.22.1) versions. Using significantly different versions may lead to unexpected behavior or incompatibilities.
Fix: Refer to the project's `README.md` or `pyproject.toml` for the exact tested environment. It's recommended to use a Python version compatible with PyTorch < 3.11 for older PyTorch versions on macOS/Windows.

Install

pip install torch-dftd Install latest version

Imports

TorchDFTD3Calculator

from torch_dftd.torch_dftd3_calculator import TorchDFTD3Calculator

TorchDFTD2Calculator
```
from torch_dftd.torch_dftd2_calculator import TorchDFTD2Calculator
```
For DFT-D2 calculations, use TorchDFTD2Calculator instead of TorchDFTD3Calculator.

Quickstart

This quickstart demonstrates how to set up a DFT-D3 calculation for a methanol molecule using `torch-dftd` with `ase`. It initializes an `ase.Atoms` object, creates a `TorchDFTD3Calculator`, attaches it to the atoms, and then computes the potential energy and atomic forces. It dynamically selects between CUDA (GPU) and CPU for computation.

import torch
import ase.build
from torch_dftd.torch_dftd3_calculator import TorchDFTD3Calculator

# Create an ASE Atoms object for methanol
atoms = ase.build.molecule("CH3OH")

# Determine device for computation
device = "cuda:0" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")

# Initialize the DFTD3 calculator with the atoms object and device
# Common damping functions are 'zero', 'bj', 's6'
calc = TorchDFTD3Calculator(atoms=atoms, device=device, damping="bj")

# Attach the calculator to the atoms object
atoms.set_calculator(calc)

# Get potential energy and forces
energy = atoms.get_potential_energy()
forces = atoms.get_forces()

print(f"Total DFT-D3 energy: {energy:.4f} eV")
print(f"Forces (first atom):
{forces[0]}")

view raw JSON →