Latest Stable Release GPUMD-v5.9.1
GPUMD-v5.9.1 Highlights
New Features, Fixes and Improvements
nep executable
- Speed up qNEP training. @brucefan1983 (Zheyong Fan) #1747
- Reduce GPU memory usage during training. @brucefan1983 (Zheyong Fan) #1782
-
Write all
*_train.outfiles in the structure order oftrain.xyz. @erhart1 (Paul Erhart) #1803 - Add an energy-difference loss for linear combinations of named structures. @erhart1 (Paul Erhart) #1799
-
Add column headers and output metadata to
loss.out. @erhart1 (Paul Erhart) #1789 -
Add a
seedkeyword tonep.into control SNES random number generation. @erhart1 (Paul Erhart) #1797 - Extend input–model consistency checks to all model types during prediction and restart. @erhart1 (Paul Erhart) #1801
gpumd executable
- Add optional support for the Wisevolve Potential Engine through its C API. @BBBuZHIDAO (Hekai Bu) #1791
- Speed up PIMD simulations through improved GPU parallelization. @BBBuZHIDAO (Hekai Bu) #1777 #1778 #1779
- Speed up qNEP simulations and reduce memory usage when Born effective charges (BECs) are not needed. @brucefan1983 (Zheyong Fan) #1721
- Add an initial native enhanced-sampling framework. @brucefan1983 (Zheyong Fan) #1750
- Support configurable PPPM mesh spacing and mesh updates for changing simulation cells. @brucefan1983 (Zheyong Fan) #1804
- Fix stale box data affecting end-of-step calculations in certain NPT and PIMD ensembles. @erhart1 (Paul Erhart) #1806
-
Fix a bug in
ti_liquid. @brucefan1983 (Zheyong Fan) #1741 - Include additional code refactoring and performance improvements. @brucefan1983 (Zheyong Fan)
New gpumd_replica executable
Notes
- The Wisevolve Potential Engine is available at https://github.com/Wisevolve/Wisevolve-Potential-Engine
-
loss.outnow includes header lines beginning with#. Custom readers should skip these lines. - License update LGPL-3.0-or-later #1808
Quick Start
Deploy GPUMD in your environment in just three steps.
Prepare Env
Load the CUDA compilation environment.
Unzip Source
Unzip the downloaded ZIP package.
Compile
Enter src directory and run make.
Official Manual
Includes complete input/output formats and simulation examples.
Version Archive
Need an older version? Check the version history here.
v5.5
Improved the NNAP potential interface and extended the DP potential interface with DPA1-DPA4 support.
Download Source (ZIP)v5.4
Make non-full-batch training faster, Add more 4-body descriptors.
Download Source (ZIP)v5.3
Extended 4-body/5-body descriptors, added compute_ic command, added partial charge option for qNEP field force.
Download Source (ZIP)v5.1
Added compute_chunk for block averages, quantum thermal bath ensembles...
Download Source (ZIP)v4.8
Supported fine-tuning descriptors; introduced lambda_q and lambda_z loss weights; released gnep executable.
Download Source (ZIP)v4.7
Introduced a new typewise cutoff scheme, further enhancing force field definition flexibility.
Download Source (ZIP)v4.6
Officially introduced qNEP machine learning potential model, supporting high-precision simulation of charged systems.
Download Source (ZIP)v4.5
Added support for ADP (Angular Dependent Potential) force fields.
Download Source (ZIP)Citation
If your research work or product uses GPUMD, please cite the following article: