Version Information

2024/3/25 Version 1.9.2

  • SLHMC: support for deformation in only 2D and 1D directions of cell shapes

  • SLHMC: enhanced error log output for Quantum ESPRESSO calculations

  • added setting for maximum standard deviation of energy and charge (maxEDiv, maxQDiv).

  • added feature to ignore characters following element names in training data.

  • fixed issue with recognition of element name Pm in training data.

2023/11/15 Version 1.9.1

  • Fixed bug which was causing invalid content of sannp.gdist when Super Epoch method is employed

2023/8/3 Version 1.9

  • Changed ReaxFF calculation process to be performed with double precision

    • Errors that was occurring in some systems are avoided.

    • ReaxFF calculations in single-precision version sannp are internally processed in the same way as double-precision version sannp_d.

  • Support training data that has elements not defined in ReaxFF parameter definition file

  • Added a method to define one force field using the average value of multiple neural network models (→models)

    • Extrapolating precision of force field is improved and make the simulations of molten state / liquid system / amorphous etc. easier.

  • Added Super Epoch method (→superEpochs)

    • Split the training data into subsets and serially perform the training process using each subset.

    • Reduce the memory consumption while suppressing the overtraining.

  • Changed default values of some of the parameters in sannp.prop/sannp.behler(values suitable for inorganic crystal are adopted)

  • SLHMC: changed default values of some of the parameters in slhmc.prop

  • Added Windows 11 to supported OS

2023/3/8 Version 1.8.1

  • resolved issue that MPI library included in NeuralMD1.8 does not work properly on CentOS7

    • In corresponding case, an error about libhwloc.so is output when executing a calculation.

    • Windows ver. 1.8 has no problem and update is not required.

    • Since NanoLabo Tool2.7 has the same problem, download and install NanoLabo Tool2.7.1.

2023/3/6 Version 1.8

  • Δ-NNP method using ReaxFF

  • Included Python script (traj_to_train.py) to convert ASE trajectory file into training data

  • Other minor corrections, etc.

2022/9/30 Version 1.7

  • Released “Pro edition”

  • [Pro] Added GPU support version (included as new executable file)

  • Added function to specify whether to apply bias term of the last layer when performing test calculation (--test-with-bias, --test-without-bias)

  • SLHMC: Optimized default values of auto tuning range of NNP-MD steps (nnpStepUp, nnpStepDown)

  • SLHMC: Added setting to perform NNP-MD with NPH

  • SLHMC: Added function to output CRASH log when terminated with error

  • Excluded CentOS6 from supported OS

  • Unified CentOS version and AlmaLinux version (unified installer)

  • Other adjustments, etc.

2022/5/20 Version 1.6

  • Added the function to run self-learning hybrid Monte Carlo methods (included as a new executable)

  • Modified the method of subtracting Coulomb energy within the cutoff radius from the training data when charge calculation is enabled

  • Other minor corrections, etc.

2022/2/4 Version 1.5

  • Added the function to set individual temperatures for each pair of elemental substitutions in the Metropolis method

  • Added the function to perform in-situ testing during training

  • Added the function to extract test data by splitting the training data

  • Improved the algorithm for calculating the optimization of the classical force field in Δ-NNP

  • Added the option to wait without terminating in case of license error (maximum number of concurrent executions) (Linux only)

  • Added the function to set the bias term of the final layer to 0 (leveling of atomic energy) when LAMMPS is executed

  • Added the information on the number of adoptions/total number of structures to the log output of the Metropolis method

  • Fixed the bug in sannp.behler template output that output was different from specified when different cutoffs were specified for G2 and G3

  • Changed MPI library/compiler of Linux executable to Open MPI/oneAPI (previously Intel MPI/Intel Compiler)

  • Other adjustments, etc.

2021/7/29 Version 1.4

  • Added the function to generate structures by Monte Carlo calculations using the Metropolis method

  • Added Δ-NNP, a combination of classical force field and NNP

  • Added twisted tanh function and GELU function to neural network activation functions

  • Improved the initial estimation method of atomic energy when using HDNNP

  • Added the function to check for incorrect values/outliers in the training data

2021/2/22 Version 1.3

  • Supported for MPI parallelism (Linux only)

  • Improved accuracy of power by exact second-order differentiation of energy

  • Added the setting to periodically output neural network to file during training

  • Improvement of libBFGS output message

  • Other adjustments, etc.

2020/12/18 Version 1.2

  • Improved the performance of L-BFGS (use external library: libLBFGS)

  • Add the function to generate batch script for Windows

2020/09/11 Version 1.1

  • Added Chebyshev symmetry functions

  • Added G3 function to Behler symmetry functions

  • Added weighted symmetry functions

  • Added the function using cos as cutoff function for symmetric functions

  • Added L-BFGS method to the neural network training algorithm.

  • Supported the installer execution in the environment without GUI.

  • Other adjustments, etc.

2020/03/05 Version 1.0

  • Initial release