Parameter calibartion package used in E3SM

                    Tao Zhang ([email protected]) 
                                    05/03/2019

Where the package

/global/homes/z/zhangtao/ACME/GCM_paras_tune/algorithms/downhill_simplex/model

Structure of the package

Configuration for downhill simplex algorithm

File: downhill_config

1000     #maximum iterations
8        #number of parametes
10e-8    #threshold for convergency 
/global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/low_ne30/run/tune_model.sh     #low res model startup
/global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/low_ne30/run/metrics/TWP/mcpi  #TWP score of low res
/global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/high_twpx4v1/run/tune_model.sh #high res model startup 
/global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/low_ne30/run/metrics/Global/mcpi #Global score of low res
/global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/high_twpx4v1/run/metrics/score #TWP score of high res

Initial values for downhill simplex algorithm

File: subrange

-2e-3 -1e-4    #begin  parameter range
0.001 0.007
0.001 0.007
0.02  0.1
1 5
2 8
0.5 2
350 1400    #end parameter range 
#Parameter initial values (N+1), N is the number of parameters. For the inital run, the shape is (N+1)xN.
#When known the corresponding metrics, the shape is (N+1)x(N+1), the last column is the metrics values. 
-8.81E-04   4.13E-03    6.72E-03    7.67E-02    1.12E+00    2.00E+00    8.61E-01    3.64E+02    6.41E-01
-7.78E-04   3.80E-03    6.23E-03    7.87E-02    1.11E+00    2.00E+00    8.33E-01    3.63E+02    7.04E-01
-9.25E-04   4.15E-03    6.78E-03    7.47E-02    1.12E+00    2.00E+00    8.63E-01    3.63E+02    7.20E-01
-8.75E-04   4.11E-03    6.66E-03    7.56E-02    1.12E+00    2.00E+00    8.59E-01    3.65E+02    6.40E-01
-9.75E-04   4.20E-03    6.89E-03    7.38E-02    1.11E+00    2.00E+00    8.68E-01    3.60E+02    8.22E-01
-1.18E-03   3.69E-03    6.45E-03    7.94E-02    1.10E+00    2.00E+00    8.66E-01    3.60E+02    7.19E-01
-1.00E-04   4.15E-03    6.73E-03    7.62E-02    1.12E+00    2.00E+00    8.63E-01    3.65E+02    3.11E+02
-1.00E-04   4.07E-03    6.65E-03    7.63E-02    1.12E+00    2.00E+00    8.58E-01    3.64E+02    3.12E+02
-8.78E-04   4.12E-03    6.69E-03    7.62E-02    1.12E+00    2.00E+00    8.60E-01    3.65E+02    9.20E+00

Downhill simplex algorithm

File: downhill_simplex.c

For the inital run, reading the parameter initial value.

read_init(){
    ...
    for(i = 0; i < NDIM + 1; i++){
        for(j = 0; j < NDIM; j++){
            if(j != NDIM){
                 fscanf(fp_config, "%lf", &INIT_Paras[i][j]);
                 printf("%e ", INIT_Paras[i][j]);
              }
              else{
                  fscanf(fp_config, "%lf", &INIT_Metrics[i]);
              }
          }
          printf("\n");
      }
      
      for(i = 0; i < NDIM + 1; i++){
          INIT_Metrics[i] = get_metrics(INIT_Paras[i]);
          printf("%e \n",INIT_Metrics[i]);
      }
      ...
  }

When known the corresponding metrics

read_init(){
    ...
    for(i = 0; i < NDIM + 1; i++){
        for(j = 0; j < NDIM+1; j++){
            if(j != NDIM){
                 fscanf(fp_config, "%lf", &INIT_Paras[i][j]);
                 printf("%e ", INIT_Paras[i][j]);
              }
              else{
                  fscanf(fp_config, "%lf", &INIT_Metrics[i]);
              }
          }
          printf("\n");
      }
      
      for(i = 0; i < NDIM + 1; i++){
          //INIT_Metrics[i] = get_metrics(INIT_Paras[i]);
          printf("%e \n",INIT_Metrics[i]);
      }
      ...
  }

updating parameter values in atm_in

File: tune_model.sh

#!/bin/bash
# parameter list
paras="zmconv_dmpdz:zmconv_c0_ocn:zmconv_c0_lnd:cldfrc_dp1:clubb_c1:clubb_c8:clubb_c14:ice_sed_ai"
# updating parameter values, which are generated by downhill simplex algorithm
sampling_file="sampling_data"
# the directory of the sampling file
sample_dir="/global/homes/z/zhangtao/ACME/GCM_paras_tune/algorithms/downhill_simplex/model/"
# the directory of the "run_e3sm.sh" file, which is to start the E3SM. 
run_dir="/global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/low_ne30/run/"
# the metrics directory
metrics_path="$run_dir/metrics"

#undating the parameter values in atm_in 
cd  $run_dir
paras_num=`echo $paras | awk -F ':' '{print NF}'`
for i in `seq 1 $paras_num`
do
    para=`echo $paras |cut -d : -f $i`
    var_line=`sed -n "1p" $sample_dir/$sampling_file`
    para_val="$para=`echo $var_line |cut -d ' ' -f $i`"
    echo $para_val
    sed -i "/\<$para\>/c \\  $para_val" $run_dir/atm_in
done

# run the E3SM 
./run_e3sm.sh
wait

# convert the output into the formate by lat x lon 
echo "e3sm has finished running!"
./remap_2d.sh low_ne30.cam.h1.2009-07*nc
./remap_2d.sh low_ne30.cam.h0*

# compute the metrics score 
cd $metrics_path/TWP
./metrics_tune.csh > metrics.log
cd $metrics_path/Global
./metrics_tune.csh > metrics.log

runing E3SM model

File: run_e3sm.sh

# set OpenMP environment variables
export 'OMP_STACKSIZE'='128M'
export 'OMP_PROC_BIND'='spread'
export 'OMP_NUM_THREADS'='4'
export 'OMP_PLACES'='threads'
# run E3SM model
srun  --label  -n 2560 -c 4  --cpu_bind=cores  /global/cscratch1/sd/zhangtao/acme_scratch/cori-knl/low_ne30/bld/e3sm.exe

getting matrics

File: metrics_tune.csh, calc_metrics_tune.ncl The metrics_tune.csh

computes the climatology mean of model simulation outputs
calls the "calc_metrics_tune.ncl", computes the performance score compared with obs
- The standard deviation of variable x
- The correlation coefficient between variable x and variable y - Loss function E(m): where the model simulation and observation are represented by "mod" and "obs", respectively. $R_0$ is the maximum value of the correlation coefficient (Here $R_0=1$). k is used to control the weight of spatial correlation (Here $k=4$). The lower of E(m), the better of simulation performance. - Merge multi-variable into a one-objective where $N^F$ is the number of variables, $E_m^F$ is the tuning model simulation, $E_r^F$ is the default model simulation. If $\chi < 1$, tuning simulation is better than the default simulation. The smaller of this value, the better of the tuning performance.
Record the parameter values and the corrsponding metrics into the MySQL database.

mysql uq_e3sm  -u  uq_e3sm_admin -h nerscdb04.nersc.gov -p3Ii3i3i2fdd_2s25j333jjdd < metrics.sql

Accessing the MySQL Database

mysql uq_e3sm  -u  uq_e3sm_admin -h nerscdb04.nersc.gov -p3Ii3i3i2fdd_2s25j333jjdd < metrics.sql

select id,c0_ocn,c0_lnd,PRECT,CLDHGH,CLDLOW,CLDMED,LWCF,SWCF,MCPI,NOTE,date from e3sm_tune_MOO;

Compile the downhill simplex algorithm

File: downhill_simplex.c

gcc -o downhill_simplex downhill_simplex.c -lm  -g

Submit the calibration package

File: submit.sh

sbatch submit.sh

tzhang-ccs / gcm_paras_tune Goto Github PK

gcm_paras_tune's Introduction