ML LOGFILE
Every VASP run with activated machine learning (INCAR contains ML_LMLFF = .TRUE.) will generate a file called ML_LOGFILE. In this log file a summary of settings and the development of quantities related to machine learning are presented in a compact, yet human-readable and post-processing friendly way. It complements the usual ab initio log output in the OUTCAR and OSZICAR files for machine learning VASP runs.
File layout
The machine learning log file is split into multiple sections, visually separated like this:
* SECTION TITLE **************************************************************************************************************************** ... content ... ********************************************************************************************************************************************
The actual composition of log sections may depend on the machine learning mode of operation (see ML_ISTART). Usually, in the beginning there will be a couple of sections describing the estimated memory consumption, machine learning settings and preexisting data. Then follows the main loop, which is split into a header and the actual loop body containing data describing the learning progress (ML_ISTART = 0, 1) or prediction (ML_ISTART = 2). Finally, there may be sections about actual memory consumption and timing statistics. The following chapters describe the contents of the log file sections in more detail:
Memory consumption estimation
This is usually the first section of the ML_LOGFILE and contains an estimation of memory requirements based on VASP files read on startup. In the simplest case (ML_ISTART = 0) it depends on the settings in the INCAR and POSCAR file. For example, the expected memory consumption may vary with the number of elements present in the POSCAR file. Various INCAR tags also influence the memory demand, e.g. ML_MB or ML_MRB2. A continuation or prediction run, i.e. ML_ISTART = 1, 2, may also take settings from the files ML_AB or ML_FF into account.
* MEMORY INFORMATION *********************************************************************************************************************** Estimated memory consumption for ML force field generation (MB): Persistent allocations for force field : 516.9 | |-- CMAT for basis : 20.3 |-- FMAT for basis : 458.5 |-- DESC for basis : 2.6 |-- DESC product matrix : 2.3 Persistent allocations for ab initio data : 8.1 | |-- Ab initio data : 7.8 |-- Ab initio data (new) : 0.3 Temporary allocations for sparsification : 460.9 | |-- SVD matrices : 28.0 Other temporary allocations : 15.5 | |-- Descriptors : 4.7 |-- Regression : 6.5 |-- Prediction : 4.2 Total memory consumption : 1001.4 ********************************************************************************************************************************************
While the individual items in the above listing are of rather technical nature the most important number is given in the last line: Total memory consumption approximates the peak memory usage during this VASP run. However, since not all memory is always allocated at the same time the actual consumption may vary over time.
The following part summarizes how each part might scale in memory with the number of CPUs:
scaLAPACK shared memory CMAT for basis x x FMAT for basis x DESC for basis x x DESC product matrix x Ab initio data Ab initio data (new) SVD matrices x Descriptors Regression x Prediction
| Mind: This is only an estimate, the actual memory requirement may be even higher. Moreover, this is only the usage for the machine learning part of VASP which in a training run adds up to the memory of the ab initio part. |
Machine learning setup
This section gives an overview of the most important INCAR tags concerning machine learning settings. The tags are grouped by topics and the tabular layout provides a short description, the current value and a "state indicator" (see the actual section header for an explanation).
* MACHINE LEARNING SETTINGS **************************************************************************************************************** This section lists the available machine-learning related settings with a short description, their selected values and the INCAR tags. The column between the value and the INCAR tag may contain a "state indicator" highlighting the origin of the value. Here is a list of possible indicators: * : (empty) Tag was not provided in the INCAR file, a default value was chosen automatically. * (I) : Value was provided in the INCAR file. * (i) : Value was provided in the INCAR file, deprecated tag. * (!) : A value found in the INCAR file was overwritten by the contents of the ML_FF file. * (?) : The value for this tag was never set (please report this to the VASP developers). Tag values with associated units are given here in Angstrom/eV, if not specified otherwise. Please refer to the VASP online manual for a detailed description of available INCAR tags. General settings -------------------------------------------------------------------------------------------------------------------------------------------- Machine learning operation mode : 0 (I) ML_ISTART Descriptor settings -------------------------------------------------------------------------------------------------------------------------------------------- Radial descriptors: ------------------- Cutoff radius of radial descriptors : 5.00000E+00 ML_RCUT1 Gaussian width for broadening the atomic distribution for radial descriptors : 5.00000E-01 ML_SION1 Number of radial basis functions for atomic distribution for radial descriptors : 8 ML_MRB1 Angular descriptors: -------------------- Cutoff radius of angular descriptors : 5.00000E+00 ML_RCUT2 Gaussian width for broadening the atomic distribution for angular descriptors : 5.00000E-01 ML_SION2 Number of radial basis functions for atomic distribution for angular descriptors : 8 ML_MRB2 Maximum angular momentum quantum number of spherical harmonics used to expand atomic distributions : 4 ML_LMAX2 ...
Existing ab initio data
This section will appear in continuation runs (e.g. ML_ISTART = 1) and summarizes the ab initio data found in the ML_AB file.
* AVAILABLE AB INITIO DATA ***************************************************************************************************************** Number of stored (maximum) ab initio structures: 114 ( 1500) * System 1 : 114 , name: "Si cubic diamond 2x2x2 super cell" * System 2 : 0 , name: "Si cubic diamond 2x2x2 super cell" Maximum number of atoms per element: * Element Si : 64 ********************************************************************************************************************************************
Main loop
The central part of the ML_LOGFILE is the main loop: depending on the machine learning mode ML_ISTART it contains data collected over all the time steps (or other iterative schemes) along the VASP run. The main loop layout is carefully designed to minimize the file size while at the same being self-descriptive. To achieve this, it is split into two parts: the description blocks in the main loop header explain the available data and present its arrangement in lines and columns. Then, the main loop body contains the actual data (mostly raw numbers) in the previously defined layout. The separation of data and its description avoids unnecessary repetition and simplifies post-processing.
Header
The main loop header consists of multiple blocks each introducing one of the log lines appearing later in the loop body. In the example below the log line STATUS is described: there will be 8 columns (counting also the word STATUS) in the given order. The meaning of each column is also briefly explained here. Next, the log line STDAB is outlined with its 5 columns, and so on...
* MAIN LOOP ******************************************************************************************************************************** # STATUS ############################################################### # STATUS This line describes the overall status of each step. # STATUS # STATUS nstep ..... MD time step or input structure counter # STATUS state ..... One-word description of step action # STATUS - "accurate" (1) : Errors are low, force field is used # STATUS - "threshold" (2) : Errors exceeded threshold, structure is sampled from ab initio # STATUS - "learning" (3) : Stored configurations are used for training force field # STATUS - "critical" (4) : Errors are high, ab initio sampling and learning is enforced # STATUS - "predict" (5) : Force field is used in prediction mode only, no error checking # STATUS is ........ Integer representation of above one-word description (integer in parenthesis) # STATUS doabin .... Perform ab initio calculation (T/F) # STATUS iff ....... Force field available (T/F, False after startup hints to possible convergence problems) # STATUS nsample ... Number of steps since last reference structure collection (sample = T) # STATUS ngenff .... Number of steps since last force field generation (genff = T) # STATUS ############################################################### # STATUS nstep state is doabin iff nsample ngenff # STATUS 2 3 4 5 6 7 8 # STATUS ############################################################### # STDAB #################################################################### # STDAB This line contains the standard deviation of the collected ab initio reference data. # STDAB # STDAB nstep ........ MD time step or input structure counter # STDAB std_energy ... Standard deviation in energy (eV atom^-1) # STDAB std_force .... Standard deviation in forces (eV Angst^-1) # STDAB std_stress ... Standard deviation in stress (kB) # STDAB #################################################################### # STDAB nstep std_energy std_force std_stress # STDAB 2 3 4 5 # STDAB #################################################################### ...
| Tip: The second column is always the current (time) step. |
Body
Right after the header the main loop body presents the time series of collected information from the VASP run. The chunks of data belonging to the same time step are fenced in dashed lines. The keywords, e.g. STDAB, which were described in the header start each line. Usually the STATUS line is located at the beginning of each time step chunk, summarizing the actions which took place during this step. In the example below the STATUS line indicates "learning", i.e. the machine learning force field was retrained.
... -------------------------------------------------------------------------------- STATUS 82 learning 3 T T 0 72 LCONF 82 Si 1222 1228 SPRSC 82 129 129 Si 1228 1224 REGR 82 1 1 1.27238822E+00 5.73175466E-02 7.83203623E-12 REGR 82 1 2 1.28510216E+00 5.73084508E-02 7.75332075E-12 REGRF 82 1 3 1.29486873E+00 5.73015362E-02 7.69391276E-12 2.23430718E+16 5.75166077E+09 STDAB 82 1.28851006E-01 1.02791005E+00 1.07081172E+01 ERR 82 1.21269596E-02 2.35740491E-01 4.40365370E+00 CFE 82 2.71935242E-01 2.20681769E-01 7.30391193E-01 LASTE 82 1.63070075E-02 2.66475855E-01 7.17595981E+00 BEE 82 4.72039040E-05 1.03291046E-01 3.02999592E-02 9.56824349E-02 6.23077315E-01 4.66683801E-01 THRHIST 82 1 8.45535075E-02 THRHIST 82 2 8.99995395E-02 THRHIST 82 3 9.42765991E-02 THRHIST 82 4 9.37027237E-02 THRHIST 82 5 9.78682111E-02 THRHIST 82 6 1.02991465E-01 THRHIST 82 7 1.04972577E-01 THRHIST 82 8 1.02574658E-01 THRHIST 82 9 9.68150073E-02 THRHIST 82 10 8.90700596E-02 THRUPD 82 9.54674570E-02 9.56824349E-02 6.60216623E-02 1.06906899E-02 BEEF 82 4.58511233E-05 9.95065359E-02 2.94732909E-02 9.56824349E-02 6.03276708E-01 4.51396163E-01 -------------------------------------------------------------------------------- ...
Note however, that the number and composition of lines in each time step is not fixed and depends on the actual procedures which were carried out. As an example, the much shorter block below only contains a STATUS and BEEF line because the machine learning force field was only used for prediction (indicated by "accurate"), so only the Bayesian error estimate was computed (BEEF line):
... -------------------------------------------------------------------------------- STATUS 63 accurate 1 F T 3 53 BEEF 63 4.67236540E-05 1.09788403E-01 2.90204790E-02 9.56824349E-02 6.29349214E-01 4.74949548E-01 -------------------------------------------------------------------------------- ...
Timing information
This last section provides timings of different machine learning program parts (ab initio code parts are not considered). There are separate columns for system clock (wall time) and CPU time (summing all threads of a process).
* TIMING INFORMATION *********************************************************************************************************************** Program part system clock (sec) cpu time (sec) ---------------------------------------------------|--------------------|------------------- Setup (file I/O, parameters,...) | 0.242 | 0.240 Descriptor and design matrix | 10.540 | 10.536 Sparsification of configurations | 9.183 | 9.177 Regression | 14.778 | 14.770 Prediction | 32.461 | 32.450 ---------------------------------------------------|--------------------|------------------- TOTAL | 67.204 | 67.173 ********************************************************************************************************************************************
Post-processing usage
Although the main loop design looks complicated at first glance it serves an important purpose: straightforward post-processing. The time series of a specific quantity can be easily constructed from the ML_LOGFILE, just by "searching" for the corresponding keyword. For example, the evolution of the prediction errors is generated by extracting all lines starting with the keyword ERR. In Linux, this can be done via the command line tool grep:
grep ERR ML_LOGFILE
An alternative to grep in Windows is the Select-String tool in the Powershell:
Select-String -CaseSensitive -Pattern "ERR" ML_LOGFILE | select-object -ExpandProperty Line
This will combine the contents of the main loop header and body to the following result:
# ERR ###################################################################### # ERR This line contains the RMSEs of the predictions with respect to ab initio results for the training data. # ERR # ERR nstep ......... MD time step or input structure counter # ERR rmse_energy ... RMSE of energies (eV atom^-1) # ERR rmse_force .... RMSE of forces (eV Angst^-1) # ERR rmse_stress ... RMSE of stress (kB) # ERR ###################################################################### # ERR nstep rmse_energy rmse_force rmse_stress # ERR 2 3 4 5 # ERR ###################################################################### ERR 2 8.77652825E-05 1.00592308E-02 2.68800480E-02 ERR 3 3.01865279E-05 1.06283576E-02 5.81209819E-02 ERR 4 1.52820686E-04 1.31384993E-02 1.10439716E-01 ERR 5 1.62739008E-04 1.74252575E-02 1.40488725E-01 ERR 6 2.97462508E-04 2.32615279E-02 1.79092561E-01 ERR 7 2.10891509E-04 2.79123925E-02 1.94566420E-01 ERR 8 3.26150852E-04 3.15081244E-02 1.76637577E-01 ERR 9 7.03479132E-04 3.42249550E-02 1.66830771E-01 ERR 10 2.41808229E-04 3.54422133E-02 1.80246157E-01 ERR 11 2.46299647E-04 3.70102675E-02 2.01262013E-01 ERR 12 3.57654922E-04 3.93143970E-02 2.20533745E-01 ERR 14 1.95974374E-04 4.31813231E-02 2.44026531E-01 ERR 15 4.94080997E-04 4.73774930E-02 2.74308998E-01 ERR 16 9.62150633E-04 5.07005683E-02 3.17482301E-01 ERR 18 1.31336233E-03 5.39222716E-02 3.25526268E-01 ERR 21 1.07020831E-03 5.67663475E-02 3.04995023E-01 ERR 24 9.88977484E-04 6.37987961E-02 3.83686143E-01 ERR 26 9.63361971E-04 6.81972633E-02 4.92021943E-01 ERR 29 1.81730719E-03 7.47758864E-02 6.38563225E-01
The output can be redirected to a file or piped into other post-processing tools. For example, the prediction error data can be directly plotted in gnuplot: first, redirect the output to a file:
grep ERR ML_LOGFILE > err.dat
Then, start gnuplot and type:
p 'err.dat' u 2:4 w l
to show the force error along time steps.