Since the release of v1.0 in 2017, GPUMD has undergone multiple version updates, but the update from v3.9.5 to v4.0 is massive. This update has received contributions and help from numerous developers and users, and has also been supported by the Science and Technology Innovation 2030 Major Project. GPUMD 4.0 was released on April 29, 2025, and the corresponding paper was published online in the domestic journal MGE Advances on August 3, 2025. Bohai University is the first author unit of this paper, with joint communication from multiple top universities and research institutions.
Overview of New Features in GPUMD-v4.0
The new features of GPUMD-v4.0 cover potentials, integration algorithms, property calculations, hardware support, and other key aspects, significantly enhancing the software's simulation capabilities and scope of application.
Major Improvements in Potentials
- Added support for Deep Potential, greatly enriching GPUMD's machine learning potentials (Main contributor: Ke Xu).
- Added two types of hybrid potentials (ILP+SW and ILP+NEP), which have extremely high accuracy for simulating multilayer 2D materials (homogeneous or heterogeneous) (Main contributor: Hekai Bu).
- Added general-purpose EAM/Alloy potentials (Main contributor: Yongchao Wu).
-
Made many core improvements to the NEP model: extended the upper limit of the parameter
l_max, supported outputting NEP descriptors, supported setting structure energy weights, published the NEP89 large model, and supported fine-tuning of large models (Main contributor: Zheyong Fan); also added the function of training atomic dipole moments and polarization tensors (Main contributor: Yifan Li).
Enhancements in Integration Algorithms and Property Calculations
Integration Algorithms: Added energy minimization with variable simulation boxes (Main contributor: Yongchao Wu), and the function to calculate liquid free energy using thermodynamic integration methods (Main contributor: Tobias Hainer).
Property Calculations: Added functions to calculate angular distribution functions (Main contributor: Yongchao Wu) and angle-dependent radial distribution functions (Main contributor: Bin Xu); added a flexible command for outputting trajectories (Main contributor: Zheyong Fan); and introduced a new command to calculate the extrapolation degree of NEP models (Main contributor: Shuning Pan).
Comprehensive Hardware Platform Support
Achieved unified support for both NVIDIA and AMD graphics cards, and successfully adapted to support the domestic DCU architecture (Main contributor: Haikuan Dong), providing great convenience for deployment on different supercomputing platforms.
"In addition to the additions and improvements of the main features mentioned above, there are also many developers and users who have helped improve and enrich the software ecosystem of the GPUMD package. We would like to thank them all here."