User Tools

Site Tools


howto:cdft

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
howto:cdft [2018/01/10 07:39]
nholmber [CDFT in summary]
howto:cdft [2018/07/24 07:38] (current)
nholmber Update to CP2K 6.1
Line 80: Line 80:
  
 Settings for the CDFT SCF loop are controlled by the input section [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT]]. An example of a typical CDFT input is given below. These parameter selections should be suitable for most systems. Settings for the CDFT SCF loop are controlled by the input section [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT]]. An example of a typical CDFT input is given below. These parameter selections should be suitable for most systems.
 +
 +<note important>​The CDFT input structure was altered slightly in CP2K version 6.1. The changes are indicated below. Input files for subsequent example calculations are provided for both versions 5.1 and 6.1, though using the latest available version is always recommended. </​note>​
  
 <code cp2k> <code cp2k>
Line 100: Line 102:
       ! Optimizer step size       ! Optimizer step size
       STEP_SIZE -1.0       STEP_SIZE -1.0
-      ! Line search settings +      ​! Note that the section CDFT_OPT exists in CP2K version >= 6.1 
-      MAX_LS 5 +      ! Remove section for CP2K version 5.1 (keywords are unchanged) 
-      CONTINUE_LS +      &​CDFT_OPT ON 
-      FACTOR_LS 0.5 +        ​! Line search settings 
-      ! Finite difference settings for calculation of Jacobian matrix +        MAX_LS 5 
-      JACOBIAN_STEP 1.0E-2 +        CONTINUE_LS 
-      JACOBIAN_FREQ 1 1 +        FACTOR_LS 0.5 
-      JACOBIAN_TYPE FD1 +        ! Finite difference settings for calculation of Jacobian matrix 
-      JACOBIAN_RESTART FALSE+        JACOBIAN_STEP 1.0E-2 
 +        JACOBIAN_FREQ 1 1 
 +        JACOBIAN_TYPE FD1 
 +        JACOBIAN_RESTART FALSE 
 +      &END CDFT_OPT
     &END OUTER_SCF     &END OUTER_SCF
   &END CDFT   &END CDFT
Line 114: Line 120:
 </​code>​ </​code>​
  
-The structure of this input section is quite straightforward. The keyword [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF#​EPS_SCF|EPS_SCF]] defines the CDFT constraint convergence threshold $\varepsilon$ and [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF#​OPTIMIZER|OPTIMIZER]] selects the CDFT optimizer. Using Newton or quasi-Newton optimizers (Broyden methods) is recommended for most applications. These optimizers accept additional control settings that define how the Jacobian matrix is calculated (keywords JACOBIAN_*) and how to optimize the step size $\alpha$ (keywords *_LS). MD simulations with a single constraint might benefit from using the bisect optimizer, which avoids building the Jacobian matrix, in case a considerable amount of the total time per MD step is spent in building the Jacobian. Notice, however, that the frequency of Jacobian rebuilds [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF#​JACOBIAN_FREQ|JACOBIAN_FREQ]] can be controlled on a per MD step and per CDFT SCF step basis. The Broyden optimizers require less frequent rebuilds of the Jacobian matrix because the matrix is [[https://​en.wikipedia.org/​wiki/​Broyden%27s_method|rank-one updated]] every iteration, although the stability of the method with respect to the rebuild frequency needs to be carefully studied.+The structure of this input section is quite straightforward. The keyword [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF#​EPS_SCF|EPS_SCF]] defines the CDFT constraint convergence threshold $\varepsilon$ and [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF#​OPTIMIZER|OPTIMIZER]] selects the CDFT optimizer. Using Newton or quasi-Newton optimizers (Broyden methods) is recommended for most applications. These optimizers accept additional control settings that define how the Jacobian matrix is calculated (keywords JACOBIAN_*) and how to optimize the step size $\alpha$ (keywords *_LS). These keywords are available in the [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF/​CDFT_OPT|CDFT_OPT]] section. MD simulations with a single constraint might benefit from using the bisect optimizer, which avoids building the Jacobian matrix, in case a considerable amount of the total time per MD step is spent in building the Jacobian. Notice, however, that the frequency of Jacobian rebuilds [[inp>​FORCE_EVAL/​DFT/​QS/​CDFT/​OUTER_SCF/CDFT_OPT#​JACOBIAN_FREQ|JACOBIAN_FREQ]] can be controlled on a per MD step and per CDFT SCF step basis. The Broyden optimizers require less frequent rebuilds of the Jacobian matrix because the matrix is [[https://​en.wikipedia.org/​wiki/​Broyden%27s_method|rank-one updated]] every iteration, although the stability of the method with respect to the rebuild frequency needs to be carefully studied.
  
 Above, for instance, the Jacobian is explicitly calculated every CDFT SCF iteration and MD step by perturbing each constraint Lagragian using a first order forward difference stencil with a step size of $10^{-2}$. The Newton step size is optimized with backtracking line search using the update formula $\alpha_n = 0.5*\alpha_{n-1}$ for a maximum of 5 steps as long as the CDFT constraint error decreases. Above, for instance, the Jacobian is explicitly calculated every CDFT SCF iteration and MD step by perturbing each constraint Lagragian using a first order forward difference stencil with a step size of $10^{-2}$. The Newton step size is optimized with backtracking line search using the update formula $\alpha_n = 0.5*\alpha_{n-1}$ for a maximum of 5 steps as long as the CDFT constraint error decreases.
howto/cdft.txt · Last modified: 2018/07/24 07:38 by nholmber