The NEB and min-mode following (dimer/Lanczos) saddle point finding methods use a force projection in order to direct the optimizers towards minimum energy paths and saddle points. These modifications to the force mean that the energy is no longer consistent with the force being optimized. Because of this, we can only use optimizers that are solely based upon the force (and not the energy). The quasi-Newton and quick-min (IBRION=1 and 3 respectively) optimizers that are built into VASP are both force-based, but the conjugate-gradient method (IBRION=2) is not.
Here, we present a set of optimizers that are all force-based so they can be used with the NEB and min-mode following methods. To use them, the INCAR must set IBRION=3 and POTIM=0, to disable the built in optimizers. Then, the IOPT parameter will select one of the following methods.
This version of quick-min is essential the same as what has been implemented in vasp. The conjugate-gradient method is different in that is uses a Newton’s line optimizer, and are entirely force based. The LBFGS is also different in that the NEB can be optimized globally, instead of image-by-image. The FIRE optimizer is an interesting new optimizer which has similarities to quick-min, but tends to be faster. The steepest descent method is provided primarily for testing. We recommend using CG or LBFGS when accurate forces are available. This is essential for evaluating curvatures. For high forces (far from the minimum) or inaccurate forces (close to the minimum) the quick-min or FIRE methods are recommended. These two methods do not rely on curvatures, and tend to be less aggressive, better behaved, but also less efficient than CG/LBFGS.
Here is a recent paper discussing the performance of these different optimizers with the NEB.
The min-mode following methods can only be used with one of these methods. If IOPT is not set in a dimer run, the job with die with a corresponding error message.
The following parameters are read from the INCAR file.
(IOPT = 0) Use VASP optimizers specified from IBRION (default) |
(IOPT = 1) LBFGS = Limited-memory Broyden-Fletcher-Goldfarb-Shanno |
(IOPT = 2) CG = Conjugate Gradient |
(IOPT = 3) QM = Quick-Min |
(IOPT = 4) SD = Steepest Descent |
(IOPT = 7) FIRE = Fast Inertial Relaxation Engine |
(IOPT = 8) ML-PYAMFF = Machine learning (PyAMFF) |
For these parameters, the listed values must be used:
Parameter |
Value |
Description |
|---|---|---|
IBRION |
3 |
Specify that VASP do molecular dynamics (with zero time step) |
POTIM |
0 |
Zero time step so that VASP does not move the ions |
For these parameters, the user may use desired values:
Parameter |
Recommended |
Description |
|---|---|---|
IOPT |
3 |
Quick-Min is the most beginner-friendly (default is IOPT=0) |
NSW |
100 |
Number of ionic relaxation steps (see VASP doc on NSW) |
EDIFFG |
-0.01 |
Must be negative (see VASP doc on EDIFFG |
Parameter |
Default |
Description |
|---|---|---|
MAXMOVE |
0.2 |
Maximum allowed step size for translation |
ILBFGSMEM |
20 |
Number of steps saved when building the inverse Hessian matrix |
LGLOBAL |
.TRUE. |
Optimize the NEB globally instead of image-by-image |
LAUTOSCALE |
.TRUE. |
Automatically determines INVCURV |
INVCURV |
0.01 |
Initial inverse curvature, used to construct the inverse Hessian matrix |
LLINEOPT |
.FALSE. |
Use a force based line minimizer for translation |
FDSTEP |
5E-3 |
Finite difference step size for line optimizer |
Parameter |
Default |
Description |
|---|---|---|
MAXMOVE |
0.2 |
Maximum allowed step size for translation |
FDSTEP |
5E-3 |
Finite difference step size to calculate curvature |
Parameter |
Default |
Description |
|---|---|---|
MAXMOVE |
0.2 |
Maximum allowed step size for translation |
TIMESTEP |
0.1 |
Dynamical time step |
Parameter |
Default |
Description |
|---|---|---|
MAXMOVE |
0.2 |
Maximum allowed step size for translation |
SDALPHA |
0.01 |
Ratio between force and step size |
Parameter |
Default |
Description |
|---|---|---|
MAXMOVE |
0.2 |
Maximum allowed step size for translation |
TIMESTEP |
0.1 |
Dynamical time step |
FTIMEMAX |
1.0 |
Maximum dynamical time step allowed |
FTIMEDEC |
0.5 |
Factor to decrease dt |
FTIMEINC |
1.1 |
Factor to increase dt |
FALPHA |
0.1 |
Parameter that controls velocity damping |
FNMIN |
5 |
Minimum number of iterations before adjusting alpha and dt |
The ML-PyAMFF optimizer trains a Behler-Parrinello neural network (see Physical Review Letters 2007, 98 (14) for detail) and uses it as the basis for a separate optimization process in order to reach a local minimum or a saddle point on the surrogate PES. While the model is being retrained with the updated training set for each cycle, the overall number of DFT force calls required for the optimization will eventually decrease while retaining the accuracy level below a pre-set threshold value.
Parameter |
Default |
Description |
|---|---|---|
PYAMFF_MODEL |
mlff.pyamff |
Input file name for neural network parameters |
PYAMFF_CONV |
GRADNORM |
Convergence criteria of neural network training (GRADNORM, RMSE) |
PYAMFF_ETOL |
0.001 |
Energy RMSE tolerance of neural network training |
PYAMFF_FTOL |
0.01 |
Force RMSE tolerance of neural network training |
PYAMFF_TOL |
0.001 |
Grandnorm tolerance of neural network training |
PYAMFF_FCOEFF |
1.0 |
Parameter that controls contribution of force loss to the training |
PYAMFF_MAXEPOCH |
2000 |
Maximum number of epochs of neural network training |
PYAMFF_OPT |
RPROP |
Type of optimizer for neural network training (RPROP, ADAM, LBFGS) |
PYAMFF_SWFTOL |
0.05 |
Criteria of switching optimizer type from ML-optimizer to LBFGS |
PYAMFF_MAXMOVE |
0.5 |
Maximum allowed step size in total for translation by ML-optimizer |
PYAMFF_MAXITER |
30 |
Maximum number of relaxation steps on ML potential energy surface |
The latest version (revision 195) requires a formatted input model file (set by PYAMFF_MODEL) to use ML-PYAMFF optimizer. The formatted file needs following information:
Fingerprint types (Behler-Parrinello)
Element types
Minimum distance between elements
Behler-Parrnello fingerprint parameters
Neural network parameters
An example of the file can be found in the section below. If you are already a PyAMFF user, you can use your model output file (mlff.pyamff) as it is.
Input file for neural networks: mlff.pyamff
Input files for optimizing a Cu13 cluster: Cu13.tar.gz
Completed calculation: Cu13_complete.tar.gz.
Input files for NEB calculation of a vacancy diffusion on Au(111): VacDiff_Au111.tar.gz
Completed calculation: VacDiff_Au111_complete.tar.gz.
The OUTCAR file has details of the calculation. This information is prefixed with the ML-PyAMFF tag. There are two main information:
Training information: The number of epochs to optimize models, their final loss, energy and force RMSE values
Machine-learned potential: The predicted force by the machine-learned potential and the number of steps taken on the potential.
ML-PyAMFF optimizer has been recently added to vtstcode6.3. It is being actively developed currently so we are welcome anyone to try and report bugs! Please use our discussion forum to ask questions or report bugs.