Direwolf is a quantum chemistry program that we are currently developing.
The current version mainly focuses on fast ωB97M-V calculations, with the goal of building high-quality training sets and synthetic data at lower cost. This is intended to address a very practical problem in drug design: many targets and chemical spaces lack sufficient experimental data, and it is impossible to fill all of these gaps through experiments alone.
VV10 is optimized, and not slower than wB97X-D too much
Software Features
The current version mainly includes:
Fast calculations with ωB97M-V, COSX/RI-J, and 3c methods
Native support for RESP charges
Very low memory usage, making it suitable for cloud deployment and large-scale batch calculations
Direwolf currently places particular emphasis on computational efficiency and deployment cost. We hope it can naturally serve as a quantum chemistry backend for AI/ML data-generation workflows.
Input and Method Combinations
Components such as J and K can be freely combined. For example:
It also supports Gaussian- and ORCA-style input files, making it easier to migrate existing workflows.
Core–Harness Architecture
Direwolf uses a Core–Harness (Driver) architecture.
The Core is only responsible for completing a single-step quantum chemistry calculation, while the Harness handles higher-level workflows such as:
Geometry Optimization
QM/MM
Periodic calculations
Other complex computational workflows
The Core can also be integrated with external software such as ASE, so not every function needs to be implemented internally.
Future Development
We plan to continue adding:
Analytic Hessian for ωB97M-V
Casida LR-TDDFT
Additional Harnesses for OPT, frequency, TS, and BOMD
More benchmarks will also be released gradually.
Advanced Customized Version
In addition to the open-source version, we are also developing several modules that are more suitable for research groups and project teams.
COSMO-RS
Combining residual learning with internal data for more accurate prediction of drug-related physicochemical properties, including:
pKa, solubility in different solvents, and other solution-related properties.
QuantaCache
A quantum chemistry data warehouse designed for research groups and project teams.
Calculation results can be stored directly. Density matrices are saved after Wigner-D rotation and can later be used directly to train semi-empirical methods, MLIPs, or other machine-learning models.
QuantaCache
From Engine to Direwolf
The development of Direwolf can be traced back to Engine, a quantum chemistry program that I started developing in 2017.
This time, we are opening up part of its capabilities. We hope Direwolf can be not only a quantum chemistry program, but also part of the data infrastructure for AI drug discovery and scientific machine learning.
Everyone is welcome to use it, and issues are also welcome.