[1] | 1 | ! |
---|
| 2 | ! $Header$ |
---|
| 3 | ! |
---|
| 4 | SUBROUTINE conccm (dtime,paprs,pplay,t,q,conv_q, |
---|
| 5 | s d_t, d_q, rain, snow, kbascm, ktopcm) |
---|
| 6 | c |
---|
| 7 | USE dimphy |
---|
| 8 | IMPLICIT none |
---|
| 9 | c====================================================================== |
---|
| 10 | c Auteur(s): Z.X. Li (LMD/CNRS) date: le 14 mars 1996 |
---|
| 11 | c Objet: Schema simple (avec flux de masse) pour la convection |
---|
| 12 | c (schema standard du modele NCAR CCM2) |
---|
| 13 | c====================================================================== |
---|
| 14 | cym#include "dimensions.h" |
---|
| 15 | cym#include "dimphy.h" |
---|
| 16 | #include "YOMCST.h" |
---|
| 17 | #include "YOETHF.h" |
---|
| 18 | c |
---|
| 19 | c Entree: |
---|
| 20 | REAL dtime ! pas d'integration |
---|
| 21 | REAL paprs(klon,klev+1) ! pression inter-couche (Pa) |
---|
| 22 | REAL pplay(klon,klev) ! pression au milieu de couche (Pa) |
---|
| 23 | REAL t(klon,klev) ! temperature (K) |
---|
| 24 | REAL q(klon,klev) ! humidite specifique (g/g) |
---|
| 25 | REAL conv_q(klon,klev) ! taux de convergence humidite (g/g/s) |
---|
| 26 | c Sortie: |
---|
| 27 | REAL d_t(klon,klev) ! incrementation temperature |
---|
| 28 | REAL d_q(klon,klev) ! incrementation vapeur |
---|
| 29 | REAL rain(klon) ! pluie (mm/s) |
---|
| 30 | REAL snow(klon) ! neige (mm/s) |
---|
| 31 | INTEGER kbascm(klon) ! niveau du bas de convection |
---|
| 32 | INTEGER ktopcm(klon) ! niveau du haut de convection |
---|
| 33 | c |
---|
| 34 | REAL pt(klon,klev) |
---|
| 35 | REAL pq(klon,klev) |
---|
| 36 | REAL pres(klon,klev) |
---|
| 37 | REAL dp(klon,klev) |
---|
| 38 | REAL zgeom(klon,klev) |
---|
| 39 | REAL cmfprs(klon) |
---|
| 40 | REAL cmfprt(klon) |
---|
| 41 | INTEGER ntop(klon) |
---|
| 42 | INTEGER nbas(klon) |
---|
| 43 | INTEGER i, k |
---|
| 44 | REAL zlvdcp, zlsdcp, zdelta, zz, za, zb |
---|
| 45 | c |
---|
| 46 | LOGICAL usekuo ! utiliser convection profonde (schema Kuo) |
---|
| 47 | PARAMETER (usekuo=.TRUE.) |
---|
| 48 | c |
---|
| 49 | REAL d_t_bis(klon,klev) |
---|
| 50 | REAL d_q_bis(klon,klev) |
---|
| 51 | REAL rain_bis(klon) |
---|
| 52 | REAL snow_bis(klon) |
---|
| 53 | INTEGER ibas_bis(klon) |
---|
| 54 | INTEGER itop_bis(klon) |
---|
| 55 | REAL d_ql_bis(klon,klev) |
---|
| 56 | REAL rneb_bis(klon,klev) |
---|
| 57 | c |
---|
| 58 | c initialiser les variables de sortie (pour securite) |
---|
| 59 | DO i = 1, klon |
---|
| 60 | rain(i) = 0.0 |
---|
| 61 | snow(i) = 0.0 |
---|
| 62 | kbascm(i) = 0 |
---|
| 63 | ktopcm(i) = 0 |
---|
| 64 | ENDDO |
---|
| 65 | DO k = 1, klev |
---|
| 66 | DO i = 1, klon |
---|
| 67 | d_t(i,k) = 0.0 |
---|
| 68 | d_q(i,k) = 0.0 |
---|
| 69 | ENDDO |
---|
| 70 | ENDDO |
---|
| 71 | c |
---|
| 72 | c preparer les variables d'entree (attention: l'ordre des niveaux |
---|
| 73 | c verticaux augmente du haut vers le bas) |
---|
| 74 | DO k = 1, klev |
---|
| 75 | DO i = 1, klon |
---|
| 76 | pt(i,k) = t(i,klev-k+1) |
---|
| 77 | pq(i,k) = q(i,klev-k+1) |
---|
| 78 | pres(i,k) = pplay(i,klev-k+1) |
---|
| 79 | dp(i,k) = paprs(i,klev+1-k)-paprs(i,klev+1-k+1) |
---|
| 80 | ENDDO |
---|
| 81 | ENDDO |
---|
| 82 | DO i = 1, klon |
---|
| 83 | zgeom(i,klev) = RD * t(i,1) / (0.5*(paprs(i,1)+pplay(i,1))) |
---|
| 84 | . * (paprs(i,1)-pplay(i,1)) |
---|
| 85 | ENDDO |
---|
| 86 | DO k = 2, klev |
---|
| 87 | DO i = 1, klon |
---|
| 88 | zgeom(i,klev+1-k) = zgeom(i,klev+1-k+1) |
---|
| 89 | . + RD * 0.5*(t(i,k-1)+t(i,k)) / paprs(i,k) |
---|
| 90 | . * (pplay(i,k-1)-pplay(i,k)) |
---|
| 91 | ENDDO |
---|
| 92 | ENDDO |
---|
| 93 | c |
---|
| 94 | CALL cmfmca(dtime, pres, dp, zgeom, pt, pq, |
---|
| 95 | $ cmfprt, cmfprs, ntop, nbas) |
---|
| 96 | C |
---|
| 97 | DO k = 1, klev |
---|
| 98 | DO i = 1, klon |
---|
| 99 | d_q(i,klev+1-k) = pq(i,k) - q(i,klev+1-k) |
---|
| 100 | d_t(i,klev+1-k) = pt(i,k) - t(i,klev+1-k) |
---|
| 101 | ENDDO |
---|
| 102 | ENDDO |
---|
| 103 | c |
---|
| 104 | DO i = 1, klon |
---|
| 105 | rain(i) = cmfprt(i) * rhoh2o |
---|
| 106 | snow(i) = cmfprs(i) * rhoh2o |
---|
| 107 | kbascm(i) = klev+1 - nbas(i) |
---|
| 108 | ktopcm(i) = klev+1 - ntop(i) |
---|
| 109 | ENDDO |
---|
| 110 | c |
---|
| 111 | IF (usekuo) THEN |
---|
| 112 | CALL conkuo(dtime, paprs, pplay, t, q, conv_q, |
---|
| 113 | s d_t_bis, d_q_bis, d_ql_bis, rneb_bis, |
---|
| 114 | s rain_bis, snow_bis, ibas_bis, itop_bis) |
---|
| 115 | DO k = 1, klev |
---|
| 116 | DO i = 1, klon |
---|
| 117 | d_t(i,k) = d_t(i,k) + d_t_bis(i,k) |
---|
| 118 | d_q(i,k) = d_q(i,k) + d_q_bis(i,k) |
---|
| 119 | ENDDO |
---|
| 120 | ENDDO |
---|
| 121 | DO i = 1, klon |
---|
| 122 | rain(i) = rain(i) + rain_bis(i) |
---|
| 123 | snow(i) = snow(i) + snow_bis(i) |
---|
| 124 | kbascm(i) = MIN(kbascm(i),ibas_bis(i)) |
---|
| 125 | ktopcm(i) = MAX(ktopcm(i),itop_bis(i)) |
---|
| 126 | ENDDO |
---|
| 127 | DO k = 1, klev ! eau liquide convective est |
---|
| 128 | DO i = 1, klon ! dispersee dans l'air |
---|
| 129 | zlvdcp=RLVTT/RCPD/(1.0+RVTMP2*q(i,k)) |
---|
| 130 | zlsdcp=RLSTT/RCPD/(1.0+RVTMP2*q(i,k)) |
---|
| 131 | zdelta = MAX(0.,SIGN(1.,RTT-t(i,k))) |
---|
| 132 | zz = d_ql_bis(i,k) ! re-evap. de l'eau liquide |
---|
| 133 | zb = MAX(0.0,zz) |
---|
| 134 | za = - MAX(0.0,zz) * (zlvdcp*(1.-zdelta)+zlsdcp*zdelta) |
---|
| 135 | d_t(i,k) = d_t(i,k) + za |
---|
| 136 | d_q(i,k) = d_q(i,k) + zb |
---|
| 137 | ENDDO |
---|
| 138 | ENDDO |
---|
| 139 | ENDIF |
---|
| 140 | c |
---|
| 141 | RETURN |
---|
| 142 | END |
---|
| 143 | SUBROUTINE cmfmca(deltat, p, dp, gz, |
---|
| 144 | $ tb, shb, |
---|
| 145 | $ cmfprt, cmfprs, cnt, cnb) |
---|
| 146 | USE dimphy |
---|
| 147 | IMPLICIT none |
---|
| 148 | C----------------------------------------------------------------------- |
---|
| 149 | C Moist convective mass flux procedure: |
---|
| 150 | C If stratification is unstable to nonentraining parcel ascent, |
---|
| 151 | C complete an adjustment making use of a simple cloud model |
---|
| 152 | C |
---|
| 153 | C Code generalized to allow specification of parcel ("updraft") |
---|
| 154 | C properties, as well as convective transport of an arbitrary |
---|
| 155 | C number of passive constituents (see cmrb array). |
---|
| 156 | C----------------------------Code History------------------------------- |
---|
| 157 | C Original version: J. J. Hack, March 22, 1990 |
---|
| 158 | C Standardized: J. Rosinski, June 1992 |
---|
| 159 | C Reviewed: J. Hack, G. Taylor, August 1992 |
---|
| 160 | c Adaptation au LMD: Z.X. Li, mars 1996 (reference: Hack 1994, |
---|
| 161 | c J. Geophys. Res. vol 99, D3, 5551-5568). J'ai |
---|
| 162 | c introduit les constantes et les fonctions thermo- |
---|
| 163 | c dynamiques du Centre Europeen. J'ai elimine le |
---|
| 164 | c re-indicage du code en esperant que cela pourra |
---|
| 165 | c simplifier la lecture et la comprehension. |
---|
| 166 | C----------------------------------------------------------------------- |
---|
| 167 | cym#include "dimensions.h" |
---|
| 168 | cym#include "dimphy.h" |
---|
| 169 | INTEGER pcnst ! nombre de traceurs passifs |
---|
| 170 | PARAMETER (pcnst=1) |
---|
| 171 | C------------------------------Arguments-------------------------------- |
---|
| 172 | C Input arguments |
---|
| 173 | C |
---|
| 174 | REAL deltat ! time step (seconds) |
---|
| 175 | REAL p(klon,klev) ! pressure |
---|
| 176 | REAL dp(klon,klev) ! delta-p |
---|
| 177 | REAL gz(klon,klev) ! geopotential (a partir du sol) |
---|
| 178 | c |
---|
| 179 | REAL thtap(klon) ! PBL perturbation theta |
---|
| 180 | REAL shp(klon) ! PBL perturbation specific humidity |
---|
| 181 | REAL pblh(klon) ! PBL height (provided by PBL routine) |
---|
| 182 | REAL cmrp(klon,pcnst) ! constituent perturbations in PBL |
---|
| 183 | c |
---|
| 184 | c Updated arguments: |
---|
| 185 | c |
---|
| 186 | REAL tb(klon,klev) ! temperature (t bar) |
---|
| 187 | REAL shb(klon,klev) ! specific humidity (sh bar) |
---|
| 188 | REAL cmrb(klon,klev,pcnst) ! constituent mixing ratios (cmr bar) |
---|
| 189 | C |
---|
| 190 | C Output arguments |
---|
| 191 | C |
---|
| 192 | REAL cmfdt(klon,klev) ! dT/dt due to moist convection |
---|
| 193 | REAL cmfdq(klon,klev) ! dq/dt due to moist convection |
---|
| 194 | REAL cmfmc(klon,klev ) ! moist convection cloud mass flux |
---|
| 195 | REAL cmfdqr(klon,klev) ! dq/dt due to convective rainout |
---|
| 196 | REAL cmfsl(klon,klev ) ! convective lw static energy flux |
---|
| 197 | REAL cmflq(klon,klev ) ! convective total water flux |
---|
| 198 | REAL cmfprt(klon) ! convective precipitation rate |
---|
| 199 | REAL cmfprs(klon) ! convective snowfall rate |
---|
| 200 | REAL qc(klon,klev) ! dq/dt due to rainout terms |
---|
| 201 | INTEGER cnt(klon) ! top level of convective activity |
---|
| 202 | INTEGER cnb(klon) ! bottom level of convective activity |
---|
| 203 | C------------------------------Parameters------------------------------- |
---|
| 204 | REAL c0 ! rain water autoconversion coefficient |
---|
| 205 | PARAMETER (c0=1.0e-4) |
---|
| 206 | REAL dzmin ! minimum convective depth for precipitation |
---|
| 207 | PARAMETER (dzmin=0.0) |
---|
| 208 | REAL betamn ! minimum overshoot parameter |
---|
| 209 | PARAMETER (betamn=0.10) |
---|
| 210 | REAL cmftau ! characteristic adjustment time scale |
---|
| 211 | PARAMETER (cmftau=3600.) |
---|
| 212 | INTEGER limcnv ! top interface level limit for convection |
---|
| 213 | PARAMETER (limcnv=1) |
---|
| 214 | REAL tpmax ! maximum acceptable t perturbation (degrees C) |
---|
| 215 | PARAMETER (tpmax=1.50) |
---|
| 216 | REAL shpmax ! maximum acceptable q perturbation (g/g) |
---|
| 217 | PARAMETER (shpmax=1.50e-3) |
---|
| 218 | REAL tiny ! arbitrary small num used in transport estimates |
---|
| 219 | PARAMETER (tiny=1.0e-36) |
---|
| 220 | REAL eps ! convergence criteria (machine dependent) |
---|
| 221 | PARAMETER (eps=1.0e-13) |
---|
| 222 | REAL tmelt ! freezing point of water(req'd for rain vs snow) |
---|
| 223 | PARAMETER (tmelt=273.15) |
---|
| 224 | REAL ssfac ! supersaturation bound (detrained air) |
---|
| 225 | PARAMETER (ssfac=1.001) |
---|
| 226 | C |
---|
| 227 | C---------------------------Local workspace----------------------------- |
---|
| 228 | REAL gam(klon,klev) ! L/cp (d(qsat)/dT) |
---|
| 229 | REAL sb(klon,klev) ! dry static energy (s bar) |
---|
| 230 | REAL hb(klon,klev) ! moist static energy (h bar) |
---|
| 231 | REAL shbs(klon,klev) ! sat. specific humidity (sh bar star) |
---|
| 232 | REAL hbs(klon,klev) ! sat. moist static energy (h bar star) |
---|
| 233 | REAL shbh(klon,klev+1) ! specific humidity on interfaces |
---|
| 234 | REAL sbh(klon,klev+1) ! s bar on interfaces |
---|
| 235 | REAL hbh(klon,klev+1) ! h bar on interfaces |
---|
| 236 | REAL cmrh(klon,klev+1) ! interface constituent mixing ratio |
---|
| 237 | REAL prec(klon) ! instantaneous total precipitation |
---|
| 238 | REAL dzcld(klon) ! depth of convective layer (m) |
---|
| 239 | REAL beta(klon) ! overshoot parameter (fraction) |
---|
| 240 | REAL betamx ! local maximum on overshoot |
---|
| 241 | REAL eta(klon) ! convective mass flux (kg/m^2 s) |
---|
| 242 | REAL etagdt ! eta*grav*deltat |
---|
| 243 | REAL cldwtr(klon) ! cloud water (mass) |
---|
| 244 | REAL rnwtr(klon) ! rain water (mass) |
---|
| 245 | REAL sc (klon) ! dry static energy ("in-cloud") |
---|
| 246 | REAL shc (klon) ! specific humidity ("in-cloud") |
---|
| 247 | REAL hc (klon) ! moist static energy ("in-cloud") |
---|
| 248 | REAL cmrc(klon) ! constituent mix rat ("in-cloud") |
---|
| 249 | REAL dq1(klon) ! shb convective change (lower lvl) |
---|
| 250 | REAL dq2(klon) ! shb convective change (mid level) |
---|
| 251 | REAL dq3(klon) ! shb convective change (upper lvl) |
---|
| 252 | REAL ds1(klon) ! sb convective change (lower lvl) |
---|
| 253 | REAL ds2(klon) ! sb convective change (mid level) |
---|
| 254 | REAL ds3(klon) ! sb convective change (upper lvl) |
---|
| 255 | REAL dcmr1(klon) ! cmrb convective change (lower lvl) |
---|
| 256 | REAL dcmr2(klon) ! cmrb convective change (mid level) |
---|
| 257 | REAL dcmr3(klon) ! cmrb convective change (upper lvl) |
---|
| 258 | REAL flotab(klon) ! hc - hbs (mesure d'instabilite) |
---|
| 259 | LOGICAL ldcum(klon) ! .true. si la convection existe |
---|
| 260 | LOGICAL etagt0 ! true if eta > 0.0 |
---|
| 261 | REAL dt ! current 2 delta-t (model time step) |
---|
| 262 | REAL cats ! modified characteristic adj. time |
---|
| 263 | REAL rdt ! 1./dt |
---|
| 264 | REAL qprime ! modified specific humidity pert. |
---|
| 265 | REAL tprime ! modified thermal perturbation |
---|
| 266 | REAL pblhgt ! bounded pbl height (max[pblh,1m]) |
---|
| 267 | REAL fac1 ! intermediate scratch variable |
---|
| 268 | REAL shprme ! intermediate specific humidity pert. |
---|
| 269 | REAL qsattp ! saturation mixing ratio for |
---|
| 270 | C ! thermally perturbed PBL parcels |
---|
| 271 | REAL dz ! local layer depth |
---|
| 272 | REAL b1 ! bouyancy measure in detrainment lvl |
---|
| 273 | REAL b2 ! bouyancy measure in condensation lvl |
---|
| 274 | REAL g ! bounded vertical gradient of hb |
---|
| 275 | REAL tmass ! total mass available for convective exchange |
---|
| 276 | REAL denom ! intermediate scratch variable |
---|
| 277 | REAL qtest1! used in negative q test (middle lvl) |
---|
| 278 | REAL qtest2! used in negative q test (lower lvl) |
---|
| 279 | REAL fslkp ! flux lw static energy (bot interface) |
---|
| 280 | REAL fslkm ! flux lw static energy (top interface) |
---|
| 281 | REAL fqlkp ! flux total water (bottom interface) |
---|
| 282 | REAL fqlkm ! flux total water (top interface) |
---|
| 283 | REAL botflx! bottom constituent mixing ratio flux |
---|
| 284 | REAL topflx! top constituent mixing ratio flux |
---|
| 285 | REAL efac1 ! ratio cmrb to convectively induced change (bot lvl) |
---|
| 286 | REAL efac2 ! ratio cmrb to convectively induced change (mid lvl) |
---|
| 287 | REAL efac3 ! ratio cmrb to convectively induced change (top lvl) |
---|
| 288 | c |
---|
| 289 | INTEGER i,k ! indices horizontal et vertical |
---|
| 290 | INTEGER km1 ! k-1 (index offset) |
---|
| 291 | INTEGER kp1 ! k+1 (index offset) |
---|
| 292 | INTEGER m ! constituent index |
---|
| 293 | INTEGER ktp ! temporary index used to track top |
---|
| 294 | INTEGER is ! nombre de points a ajuster |
---|
| 295 | C |
---|
| 296 | REAL tmp1, tmp2, tmp3, tmp4 |
---|
| 297 | REAL zx_t, zx_p, zx_q, zx_qs, zx_gam |
---|
| 298 | REAL zcor, zdelta, zcvm5 |
---|
| 299 | C |
---|
| 300 | REAL qhalf, sh1, sh2, shbs1, shbs2 |
---|
| 301 | #include "YOMCST.h" |
---|
| 302 | #include "YOETHF.h" |
---|
| 303 | #include "FCTTRE.h" |
---|
| 304 | qhalf(sh1,sh2,shbs1,shbs2) = MIN(MAX(sh1,sh2), |
---|
| 305 | $ (shbs2*sh1 + shbs1*sh2)/(shbs1+shbs2)) |
---|
| 306 | C |
---|
| 307 | C----------------------------------------------------------------------- |
---|
| 308 | c pas de traceur pour l'instant |
---|
| 309 | DO m = 1, pcnst |
---|
| 310 | DO k = 1, klev |
---|
| 311 | DO i = 1, klon |
---|
| 312 | cmrb(i,k,m) = 0.0 |
---|
| 313 | ENDDO |
---|
| 314 | ENDDO |
---|
| 315 | ENDDO |
---|
| 316 | c |
---|
| 317 | c Les perturbations de la couche limite sont zero pour l'instant |
---|
| 318 | c |
---|
| 319 | DO m = 1, pcnst |
---|
| 320 | DO i = 1, klon |
---|
| 321 | cmrp(i,m) = 0.0 |
---|
| 322 | ENDDO |
---|
| 323 | ENDDO |
---|
| 324 | DO i = 1, klon |
---|
| 325 | thtap(i) = 0.0 |
---|
| 326 | shp(i) = 0.0 |
---|
| 327 | pblh(i) = 1.0 |
---|
| 328 | ENDDO |
---|
| 329 | C |
---|
| 330 | C Ensure that characteristic adjustment time scale (cmftau) assumed |
---|
| 331 | C in estimate of eta isn't smaller than model time scale (deltat) |
---|
| 332 | C |
---|
| 333 | dt = deltat |
---|
| 334 | cats = MAX(dt,cmftau) |
---|
| 335 | rdt = 1.0/dt |
---|
| 336 | C |
---|
| 337 | C Compute sb,hb,shbs,hbs |
---|
| 338 | C |
---|
| 339 | DO k = 1, klev |
---|
| 340 | DO i = 1, klon |
---|
| 341 | zx_t = tb(i,k) |
---|
| 342 | zx_p = p(i,k) |
---|
| 343 | zx_q = shb(i,k) |
---|
| 344 | zdelta=MAX(0.,SIGN(1.,RTT-zx_t)) |
---|
| 345 | zcvm5 = R5LES*RLVTT*(1.-zdelta) + R5IES*RLSTT*zdelta |
---|
| 346 | zcvm5 = zcvm5 / RCPD / (1.0+RVTMP2*zx_q) |
---|
| 347 | zx_qs= r2es * FOEEW(zx_t,zdelta)/zx_p |
---|
| 348 | zx_qs=MIN(0.5,zx_qs) |
---|
| 349 | zcor=1./(1.-retv*zx_qs) |
---|
| 350 | zx_qs=zx_qs*zcor |
---|
| 351 | zx_gam = FOEDE(zx_t,zdelta,zcvm5,zx_qs,zcor) |
---|
| 352 | shbs(i,k) = zx_qs |
---|
| 353 | gam(i,k) = zx_gam |
---|
| 354 | ENDDO |
---|
| 355 | ENDDO |
---|
| 356 | C |
---|
| 357 | DO k=limcnv,klev |
---|
| 358 | DO i=1,klon |
---|
| 359 | sb (i,k) = RCPD*tb(i,k) + gz(i,k) |
---|
| 360 | hb (i,k) = sb(i,k) + RLVTT*shb(i,k) |
---|
| 361 | hbs(i,k) = sb(i,k) + RLVTT*shbs(i,k) |
---|
| 362 | ENDDO |
---|
| 363 | ENDDO |
---|
| 364 | C |
---|
| 365 | C Compute sbh, shbh |
---|
| 366 | C |
---|
| 367 | DO k=limcnv+1,klev |
---|
| 368 | km1 = k - 1 |
---|
| 369 | DO i=1,klon |
---|
| 370 | sbh (i,k) =0.5*(sb(i,km1) + sb(i,k)) |
---|
| 371 | shbh(i,k) =qhalf(shb(i,km1),shb(i,k),shbs(i,km1),shbs(i,k)) |
---|
| 372 | hbh (i,k) =sbh(i,k) + RLVTT*shbh(i,k) |
---|
| 373 | ENDDO |
---|
| 374 | ENDDO |
---|
| 375 | C |
---|
| 376 | C Specify properties at top of model (not used, but filling anyway) |
---|
| 377 | C |
---|
| 378 | DO i=1,klon |
---|
| 379 | sbh (i,limcnv) = sb(i,limcnv) |
---|
| 380 | shbh(i,limcnv) = shb(i,limcnv) |
---|
| 381 | hbh (i,limcnv) = hb(i,limcnv) |
---|
| 382 | ENDDO |
---|
| 383 | C |
---|
| 384 | C Zero vertically independent control, tendency & diagnostic arrays |
---|
| 385 | C |
---|
| 386 | DO i=1,klon |
---|
| 387 | prec(i) = 0.0 |
---|
| 388 | dzcld(i) = 0.0 |
---|
| 389 | cnb(i) = 0 |
---|
| 390 | cnt(i) = klev+1 |
---|
| 391 | ENDDO |
---|
| 392 | |
---|
| 393 | DO k = 1, klev |
---|
| 394 | DO i = 1,klon |
---|
| 395 | cmfdt(i,k) = 0. |
---|
| 396 | cmfdq(i,k) = 0. |
---|
| 397 | cmfdqr(i,k) = 0. |
---|
| 398 | cmfmc(i,k) = 0. |
---|
| 399 | cmfsl(i,k) = 0. |
---|
| 400 | cmflq(i,k) = 0. |
---|
| 401 | ENDDO |
---|
| 402 | ENDDO |
---|
| 403 | C |
---|
| 404 | C Begin moist convective mass flux adjustment procedure. |
---|
| 405 | C Formalism ensures that negative cloud liquid water can never occur |
---|
| 406 | C |
---|
| 407 | DO 70 k=klev-1,limcnv+1,-1 |
---|
| 408 | km1 = k - 1 |
---|
| 409 | kp1 = k + 1 |
---|
| 410 | DO 10 i=1,klon |
---|
| 411 | eta (i) = 0.0 |
---|
| 412 | beta (i) = 0.0 |
---|
| 413 | ds1 (i) = 0.0 |
---|
| 414 | ds2 (i) = 0.0 |
---|
| 415 | ds3 (i) = 0.0 |
---|
| 416 | dq1 (i) = 0.0 |
---|
| 417 | dq2 (i) = 0.0 |
---|
| 418 | dq3 (i) = 0.0 |
---|
| 419 | C |
---|
| 420 | C Specification of "cloud base" conditions |
---|
| 421 | C |
---|
| 422 | qprime = 0.0 |
---|
| 423 | tprime = 0.0 |
---|
| 424 | C |
---|
| 425 | C Assign tprime within the PBL to be proportional to the quantity |
---|
| 426 | C thtap (which will be bounded by tpmax), passed to this routine by |
---|
| 427 | C the PBL routine. Don't allow perturbation to produce a dry |
---|
| 428 | C adiabatically unstable parcel. Assign qprime within the PBL to be |
---|
| 429 | C an appropriately modified value of the quantity shp (which will be |
---|
| 430 | C bounded by shpmax) passed to this routine by the PBL routine. The |
---|
| 431 | C quantity qprime should be less than the local saturation value |
---|
| 432 | C (qsattp=qsat[t+tprime,p]). In both cases, thtap and shp are |
---|
| 433 | C linearly reduced toward zero as the PBL top is approached. |
---|
| 434 | C |
---|
| 435 | pblhgt = MAX(pblh(i),1.0) |
---|
| 436 | IF (gz(i,kp1)/RG.LE.pblhgt .AND. dzcld(i).EQ.0.0) THEN |
---|
| 437 | fac1 = MAX(0.0,1.0-gz(i,kp1)/RG/pblhgt) |
---|
| 438 | tprime = MIN(thtap(i),tpmax)*fac1 |
---|
| 439 | qsattp = shbs(i,kp1) + RCPD/RLVTT*gam(i,kp1)*tprime |
---|
| 440 | shprme = MIN(MIN(shp(i),shpmax)*fac1, |
---|
| 441 | $ MAX(qsattp-shb(i,kp1),0.0)) |
---|
| 442 | qprime = MAX(qprime,shprme) |
---|
| 443 | ELSE |
---|
| 444 | tprime = 0.0 |
---|
| 445 | qprime = 0.0 |
---|
| 446 | ENDIF |
---|
| 447 | C |
---|
| 448 | C Specify "updraft" (in-cloud) thermodynamic properties |
---|
| 449 | C |
---|
| 450 | sc (i) = sb (i,kp1) + RCPD*tprime |
---|
| 451 | shc(i) = shb(i,kp1) + qprime |
---|
| 452 | hc (i) = sc (i ) + RLVTT*shc(i) |
---|
| 453 | flotab(i) = hc(i) - hbs(i,k) |
---|
| 454 | dz = dp(i,k)*RD*tb(i,k)/RG/p(i,k) |
---|
| 455 | IF (flotab(i).gt.0.0) THEN |
---|
| 456 | dzcld(i) = dzcld(i) + dz |
---|
| 457 | ELSE |
---|
| 458 | dzcld(i) = 0.0 |
---|
| 459 | ENDIF |
---|
| 460 | 10 CONTINUE |
---|
| 461 | C |
---|
| 462 | C Check on moist convective instability |
---|
| 463 | C |
---|
| 464 | is = 0 |
---|
| 465 | DO i = 1, klon |
---|
| 466 | IF (flotab(i).GT.0.0) THEN |
---|
| 467 | ldcum(i) = .TRUE. |
---|
| 468 | is = is + 1 |
---|
| 469 | ELSE |
---|
| 470 | ldcum(i) = .FALSE. |
---|
| 471 | ENDIF |
---|
| 472 | ENDDO |
---|
| 473 | C |
---|
| 474 | IF (is.EQ.0) THEN |
---|
| 475 | DO i=1,klon |
---|
| 476 | dzcld(i) = 0.0 |
---|
| 477 | ENDDO |
---|
| 478 | GOTO 70 |
---|
| 479 | ENDIF |
---|
| 480 | C |
---|
| 481 | C Current level just below top level => no overshoot |
---|
| 482 | C |
---|
| 483 | IF (k.le.limcnv+1) THEN |
---|
| 484 | DO i=1,klon |
---|
| 485 | IF (ldcum(i)) THEN |
---|
| 486 | cldwtr(i) = sb(i,k)-sc(i)+flotab(i)/(1.0+gam(i,k)) |
---|
| 487 | cldwtr(i) = MAX(0.0,cldwtr(i)) |
---|
| 488 | beta(i) = 0.0 |
---|
| 489 | ENDIF |
---|
| 490 | ENDDO |
---|
| 491 | GOTO 20 |
---|
| 492 | ENDIF |
---|
| 493 | C |
---|
| 494 | C First guess at overshoot parameter using crude buoyancy closure |
---|
| 495 | C 10% overshoot assumed as a minimum and 1-c0*dz maximum to start |
---|
| 496 | C If pre-existing supersaturation in detrainment layer, beta=0 |
---|
| 497 | C cldwtr is temporarily equal to RLVTT*l (l=> liquid water) |
---|
| 498 | C |
---|
| 499 | DO i=1,klon |
---|
| 500 | IF (ldcum(i)) THEN |
---|
| 501 | cldwtr(i) = sb(i,k)-sc(i)+flotab(i)/(1.0+gam(i,k)) |
---|
| 502 | cldwtr(i) = MAX(0.0,cldwtr(i)) |
---|
| 503 | betamx = 1.0 - c0*MAX(0.0,(dzcld(i)-dzmin)) |
---|
| 504 | b1 = (hc(i) - hbs(i,km1))*dp(i,km1) |
---|
| 505 | b2 = (hc(i) - hbs(i,k ))*dp(i,k ) |
---|
| 506 | beta(i) = MAX(betamn,MIN(betamx, 1.0+b1/b2)) |
---|
| 507 | IF (hbs(i,km1).le.hb(i,km1)) beta(i) = 0.0 |
---|
| 508 | ENDIF |
---|
| 509 | ENDDO |
---|
| 510 | C |
---|
| 511 | C Bound maximum beta to ensure physically realistic solutions |
---|
| 512 | C |
---|
| 513 | C First check constrains beta so that eta remains positive |
---|
| 514 | C (assuming that eta is already positive for beta equal zero) |
---|
| 515 | c La premiere contrainte de beta est que le flux eta doit etre positif. |
---|
| 516 | C |
---|
| 517 | DO i=1,klon |
---|
| 518 | IF (ldcum(i)) THEN |
---|
| 519 | tmp1 = (1.0+gam(i,k))*(sc(i)-sbh(i,kp1) + cldwtr(i)) |
---|
| 520 | $ - (hbh(i,kp1)-hc(i))*dp(i,k)/dp(i,kp1) |
---|
| 521 | tmp2 = (1.0+gam(i,k))*(sc(i)-sbh(i,k)) |
---|
| 522 | IF ((beta(i)*tmp2-tmp1).GT.0.0) THEN |
---|
| 523 | betamx = 0.99*(tmp1/tmp2) |
---|
| 524 | beta(i) = MAX(0.0,MIN(betamx,beta(i))) |
---|
| 525 | ENDIF |
---|
| 526 | C |
---|
| 527 | C Second check involves supersaturation of "detrainment layer" |
---|
| 528 | C small amount of supersaturation acceptable (by ssfac factor) |
---|
| 529 | c La 2e contrainte est que la convection ne doit pas sursaturer |
---|
| 530 | c la "detrainment layer", Neanmoins, une petite sursaturation |
---|
| 531 | c est acceptee (facteur ssfac). |
---|
| 532 | C |
---|
| 533 | IF (hb(i,km1).lt.hbs(i,km1)) THEN |
---|
| 534 | tmp1 = (1.0+gam(i,k))*(sc(i)-sbh(i,kp1) + cldwtr(i)) |
---|
| 535 | $ - (hbh(i,kp1)-hc(i))*dp(i,k)/dp(i,kp1) |
---|
| 536 | tmp1 = tmp1/dp(i,k) |
---|
| 537 | tmp2 = gam(i,km1)*(sbh(i,k)-sc(i) + cldwtr(i)) - |
---|
| 538 | $ hbh(i,k) + hc(i) - sc(i) + sbh(i,k) |
---|
| 539 | tmp3 = (1.0+gam(i,k))*(sc(i)-sbh(i,k))/dp(i,k) |
---|
| 540 | tmp4 = (dt/cats)*(hc(i)-hbs(i,k))*tmp2 |
---|
| 541 | $ / (dp(i,km1)*(hbs(i,km1)-hb(i,km1))) + tmp3 |
---|
| 542 | IF ((beta(i)*tmp4-tmp1).GT.0.0) THEN |
---|
| 543 | betamx = ssfac*(tmp1/tmp4) |
---|
| 544 | beta(i) = MAX(0.0,MIN(betamx,beta(i))) |
---|
| 545 | ENDIF |
---|
| 546 | ELSE |
---|
| 547 | beta(i) = 0.0 |
---|
| 548 | ENDIF |
---|
| 549 | C |
---|
| 550 | C Third check to avoid introducing 2 delta x thermodynamic |
---|
| 551 | C noise in the vertical ... constrain adjusted h (or theta e) |
---|
| 552 | C so that the adjustment doesn't contribute to "kinks" in h |
---|
| 553 | C |
---|
| 554 | g = MIN(0.0,hb(i,k)-hb(i,km1)) |
---|
| 555 | tmp3 = (hb(i,k)-hb(i,km1)-g)*(cats/dt) / (hc(i)-hbs(i,k)) |
---|
| 556 | tmp1 = (1.0+gam(i,k))*(sc(i)-sbh(i,kp1) + cldwtr(i)) |
---|
| 557 | $ - (hbh(i,kp1)-hc(i))*dp(i,k)/dp(i,kp1) |
---|
| 558 | tmp1 = tmp1/dp(i,k) |
---|
| 559 | tmp1 = tmp3*tmp1 + (hc(i) - hbh(i,kp1))/dp(i,k) |
---|
| 560 | tmp2 = tmp3*(1.0+gam(i,k))*(sc(i)-sbh(i,k))/dp(i,k) |
---|
| 561 | $ + (hc(i)-hbh(i,k)-cldwtr(i)) |
---|
| 562 | $ *(1.0/dp(i,k)+1.0/dp(i,kp1)) |
---|
| 563 | IF ((beta(i)*tmp2-tmp1).GT.0.0) THEN |
---|
| 564 | betamx = 0.0 |
---|
| 565 | IF (tmp2.NE.0.0) betamx = tmp1/tmp2 |
---|
| 566 | beta(i) = MAX(0.0,MIN(betamx,beta(i))) |
---|
| 567 | ENDIF |
---|
| 568 | ENDIF |
---|
| 569 | ENDDO |
---|
| 570 | C |
---|
| 571 | C Calculate mass flux required for stabilization. |
---|
| 572 | C |
---|
| 573 | C Ensure that the convective mass flux, eta, is positive by |
---|
| 574 | C setting negative values of eta to zero.. |
---|
| 575 | C Ensure that estimated mass flux cannot move more than the |
---|
| 576 | C minimum of total mass contained in either layer k or layer k+1. |
---|
| 577 | C Also test for other pathological cases that result in non- |
---|
| 578 | C physical states and adjust eta accordingly. |
---|
| 579 | C |
---|
| 580 | 20 CONTINUE |
---|
| 581 | DO i=1,klon |
---|
| 582 | IF (ldcum(i)) THEN |
---|
| 583 | beta(i) = MAX(0.0,beta(i)) |
---|
| 584 | tmp1 = hc(i) - hbs(i,k) |
---|
| 585 | tmp2 = ((1.0+gam(i,k))*(sc(i)-sbh(i,kp1)+cldwtr(i)) - |
---|
| 586 | $ beta(i)*(1.0+gam(i,k))*(sc(i)-sbh(i,k)))/dp(i,k) - |
---|
| 587 | $ (hbh(i,kp1)-hc(i))/dp(i,kp1) |
---|
| 588 | eta(i) = tmp1/(tmp2*RG*cats) |
---|
| 589 | tmass = MIN(dp(i,k),dp(i,kp1))/RG |
---|
| 590 | IF (eta(i).GT.tmass*rdt .OR. eta(i).LE.0.0) eta(i) = 0.0 |
---|
| 591 | C |
---|
| 592 | C Check on negative q in top layer (bound beta) |
---|
| 593 | C |
---|
| 594 | IF(shc(i)-shbh(i,k).LT.0.0 .AND. beta(i)*eta(i).NE.0.0)THEN |
---|
| 595 | denom = eta(i)*RG*dt*(shc(i) - shbh(i,k))/dp(i,km1) |
---|
| 596 | beta(i) = MAX(0.0,MIN(-0.999*shb(i,km1)/denom,beta(i))) |
---|
| 597 | ENDIF |
---|
| 598 | C |
---|
| 599 | C Check on negative q in middle layer (zero eta) |
---|
| 600 | C |
---|
| 601 | qtest1 = shb(i,k) + eta(i)*RG*dt*((shc(i) - shbh(i,kp1)) - |
---|
| 602 | $ (1.0 - beta(i))*cldwtr(i)/RLVTT - |
---|
| 603 | $ beta(i)*(shc(i) - shbh(i,k)))/dp(i,k) |
---|
| 604 | IF (qtest1.le.0.0) eta(i) = 0.0 |
---|
| 605 | C |
---|
| 606 | C Check on negative q in lower layer (bound eta) |
---|
| 607 | C |
---|
| 608 | fac1 = -(shbh(i,kp1) - shc(i))/dp(i,kp1) |
---|
| 609 | qtest2 = shb(i,kp1) - eta(i)*RG*dt*fac1 |
---|
| 610 | IF (qtest2 .lt. 0.0) THEN |
---|
| 611 | eta(i) = 0.99*shb(i,kp1)/(RG*dt*fac1) |
---|
| 612 | ENDIF |
---|
| 613 | ENDIF |
---|
| 614 | ENDDO |
---|
| 615 | C |
---|
| 616 | C |
---|
| 617 | C Calculate cloud water, rain water, and thermodynamic changes |
---|
| 618 | C |
---|
| 619 | DO 30 i=1,klon |
---|
| 620 | IF (ldcum(i)) THEN |
---|
| 621 | etagdt = eta(i)*RG*dt |
---|
| 622 | cldwtr(i) = etagdt*cldwtr(i)/RLVTT/RG |
---|
| 623 | rnwtr(i) = (1.0 - beta(i))*cldwtr(i) |
---|
| 624 | ds1(i) = etagdt*(sbh(i,kp1) - sc(i))/dp(i,kp1) |
---|
| 625 | dq1(i) = etagdt*(shbh(i,kp1) - shc(i))/dp(i,kp1) |
---|
| 626 | ds2(i) = (etagdt*(sc(i) - sbh(i,kp1)) + |
---|
| 627 | $ RLVTT*RG*cldwtr(i) - beta(i)*etagdt* |
---|
| 628 | $ (sc(i) - sbh(i,k)))/dp(i,k) |
---|
| 629 | dq2(i) = (etagdt*(shc(i) - shbh(i,kp1)) - |
---|
| 630 | $ RG*rnwtr(i) - beta(i)*etagdt* |
---|
| 631 | $ (shc(i) - shbh(i,k)))/dp(i,k) |
---|
| 632 | ds3(i) = beta(i)*(etagdt*(sc(i) - sbh(i,k)) - |
---|
| 633 | $ RLVTT*RG*cldwtr(i))/dp(i,km1) |
---|
| 634 | dq3(i) = beta(i)*etagdt*(shc(i) - shbh(i,k))/dp(i,km1) |
---|
| 635 | C |
---|
| 636 | C Isolate convective fluxes for later diagnostics |
---|
| 637 | C |
---|
| 638 | fslkp = eta(i)*(sc(i) - sbh(i,kp1)) |
---|
| 639 | fslkm = beta(i)*(eta(i)*(sc(i) - sbh(i,k)) - |
---|
| 640 | $ RLVTT*cldwtr(i)*rdt) |
---|
| 641 | fqlkp = eta(i)*(shc(i) - shbh(i,kp1)) |
---|
| 642 | fqlkm = beta(i)*eta(i)*(shc(i) - shbh(i,k)) |
---|
| 643 | C |
---|
| 644 | C |
---|
| 645 | C Update thermodynamic profile (update sb, hb, & hbs later) |
---|
| 646 | C |
---|
| 647 | tb (i,kp1) = tb(i,kp1) + ds1(i) / RCPD |
---|
| 648 | tb (i,k ) = tb(i,k ) + ds2(i) / RCPD |
---|
| 649 | tb (i,km1) = tb(i,km1) + ds3(i) / RCPD |
---|
| 650 | shb(i,kp1) = shb(i,kp1) + dq1(i) |
---|
| 651 | shb(i,k ) = shb(i,k ) + dq2(i) |
---|
| 652 | shb(i,km1) = shb(i,km1) + dq3(i) |
---|
| 653 | prec(i) = prec(i) + rnwtr(i)/rhoh2o |
---|
| 654 | C |
---|
| 655 | C Update diagnostic information for final budget |
---|
| 656 | C Tracking temperature & specific humidity tendencies, |
---|
| 657 | C rainout term, convective mass flux, convective liquid |
---|
| 658 | C water static energy flux, and convective total water flux |
---|
| 659 | C |
---|
| 660 | cmfdt (i,kp1) = cmfdt (i,kp1) + ds1(i)/RCPD*rdt |
---|
| 661 | cmfdt (i,k ) = cmfdt (i,k ) + ds2(i)/RCPD*rdt |
---|
| 662 | cmfdt (i,km1) = cmfdt (i,km1) + ds3(i)/RCPD*rdt |
---|
| 663 | cmfdq (i,kp1) = cmfdq (i,kp1) + dq1(i)*rdt |
---|
| 664 | cmfdq (i,k ) = cmfdq (i,k ) + dq2(i)*rdt |
---|
| 665 | cmfdq (i,km1) = cmfdq (i,km1) + dq3(i)*rdt |
---|
| 666 | cmfdqr(i,k ) = cmfdqr(i,k ) + (RG*rnwtr(i)/dp(i,k))*rdt |
---|
| 667 | cmfmc (i,kp1) = cmfmc (i,kp1) + eta(i) |
---|
| 668 | cmfmc (i,k ) = cmfmc (i,k ) + beta(i)*eta(i) |
---|
| 669 | cmfsl (i,kp1) = cmfsl (i,kp1) + fslkp |
---|
| 670 | cmfsl (i,k ) = cmfsl (i,k ) + fslkm |
---|
| 671 | cmflq (i,kp1) = cmflq (i,kp1) + RLVTT*fqlkp |
---|
| 672 | cmflq (i,k ) = cmflq (i,k ) + RLVTT*fqlkm |
---|
| 673 | qc (i,k ) = (RG*rnwtr(i)/dp(i,k))*rdt |
---|
| 674 | ENDIF |
---|
| 675 | 30 CONTINUE |
---|
| 676 | C |
---|
| 677 | C Next, convectively modify passive constituents |
---|
| 678 | C |
---|
| 679 | DO 50 m=1,pcnst |
---|
| 680 | DO 40 i=1,klon |
---|
| 681 | IF (ldcum(i)) THEN |
---|
| 682 | C |
---|
| 683 | C If any of the reported values of the constituent is negative in |
---|
| 684 | C the three adjacent levels, nothing will be done to the profile |
---|
| 685 | C |
---|
| 686 | IF ((cmrb(i,kp1,m).LT.0.0) .OR. |
---|
| 687 | $ (cmrb(i,k,m).LT.0.0) .OR. |
---|
| 688 | $ (cmrb(i,km1,m).LT.0.0)) GOTO 40 |
---|
| 689 | C |
---|
| 690 | C Specify constituent interface values (linear interpolation) |
---|
| 691 | C |
---|
| 692 | cmrh(i,k ) = 0.5*(cmrb(i,km1,m) + cmrb(i,k ,m)) |
---|
| 693 | cmrh(i,kp1) = 0.5*(cmrb(i,k ,m) + cmrb(i,kp1,m)) |
---|
| 694 | C |
---|
| 695 | C Specify perturbation properties of constituents in PBL |
---|
| 696 | C |
---|
| 697 | pblhgt = MAX(pblh(i),1.0) |
---|
| 698 | IF (gz(i,kp1)/RG.LE.pblhgt .AND. dzcld(i).EQ.0.) THEN |
---|
| 699 | fac1 = MAX(0.0,1.0-gz(i,kp1)/RG/pblhgt) |
---|
| 700 | cmrc(i) = cmrb(i,kp1,m) + cmrp(i,m)*fac1 |
---|
| 701 | ELSE |
---|
| 702 | cmrc(i) = cmrb(i,kp1,m) |
---|
| 703 | ENDIF |
---|
| 704 | C |
---|
| 705 | C Determine fluxes, flux divergence => changes due to convection |
---|
| 706 | C Logic must be included to avoid producing negative values. A bit |
---|
| 707 | C messy since there are no a priori assumptions about profiles. |
---|
| 708 | C Tendency is modified (reduced) when pending disaster detected. |
---|
| 709 | C |
---|
| 710 | etagdt = eta(i)*RG*dt |
---|
| 711 | botflx = etagdt*(cmrc(i) - cmrh(i,kp1)) |
---|
| 712 | topflx = beta(i)*etagdt*(cmrc(i)-cmrh(i,k)) |
---|
| 713 | dcmr1(i) = -botflx/dp(i,kp1) |
---|
| 714 | efac1 = 1.0 |
---|
| 715 | efac2 = 1.0 |
---|
| 716 | efac3 = 1.0 |
---|
| 717 | C |
---|
| 718 | IF (cmrb(i,kp1,m)+dcmr1(i) .LT. 0.0) THEN |
---|
| 719 | efac1 = MAX(tiny,ABS(cmrb(i,kp1,m)/dcmr1(i)) - eps) |
---|
| 720 | ENDIF |
---|
| 721 | C |
---|
| 722 | IF (efac1.EQ.tiny .OR. efac1.GT.1.0) efac1 = 0.0 |
---|
| 723 | dcmr1(i) = -efac1*botflx/dp(i,kp1) |
---|
| 724 | dcmr2(i) = (efac1*botflx - topflx)/dp(i,k) |
---|
| 725 | C |
---|
| 726 | IF (cmrb(i,k,m)+dcmr2(i) .LT. 0.0) THEN |
---|
| 727 | efac2 = MAX(tiny,ABS(cmrb(i,k ,m)/dcmr2(i)) - eps) |
---|
| 728 | ENDIF |
---|
| 729 | C |
---|
| 730 | IF (efac2.EQ.tiny .OR. efac2.GT.1.0) efac2 = 0.0 |
---|
| 731 | dcmr2(i) = (efac1*botflx - efac2*topflx)/dp(i,k) |
---|
| 732 | dcmr3(i) = efac2*topflx/dp(i,km1) |
---|
| 733 | C |
---|
| 734 | IF (cmrb(i,km1,m)+dcmr3(i) .LT. 0.0) THEN |
---|
| 735 | efac3 = MAX(tiny,ABS(cmrb(i,km1,m)/dcmr3(i)) - eps) |
---|
| 736 | ENDIF |
---|
| 737 | C |
---|
| 738 | IF (efac3.EQ.tiny .OR. efac3.GT.1.0) efac3 = 0.0 |
---|
| 739 | efac3 = MIN(efac2,efac3) |
---|
| 740 | dcmr2(i) = (efac1*botflx - efac3*topflx)/dp(i,k) |
---|
| 741 | dcmr3(i) = efac3*topflx/dp(i,km1) |
---|
| 742 | C |
---|
| 743 | cmrb(i,kp1,m) = cmrb(i,kp1,m) + dcmr1(i) |
---|
| 744 | cmrb(i,k ,m) = cmrb(i,k ,m) + dcmr2(i) |
---|
| 745 | cmrb(i,km1,m) = cmrb(i,km1,m) + dcmr3(i) |
---|
| 746 | ENDIF |
---|
| 747 | 40 CONTINUE |
---|
| 748 | 50 CONTINUE ! end of m=1,pcnst loop |
---|
| 749 | C |
---|
| 750 | IF (k.EQ.limcnv+1) GOTO 60 ! on ne pourra plus glisser |
---|
| 751 | c |
---|
| 752 | c Dans la procedure de glissage ascendant, les variables thermo- |
---|
| 753 | c dynamiques des couches k et km1 servent au calcul des couches |
---|
| 754 | c superieures. Elles ont donc besoin d'une mise-a-jour. |
---|
| 755 | C |
---|
| 756 | DO i = 1, klon |
---|
| 757 | IF (ldcum(i)) THEN |
---|
| 758 | zx_t = tb(i,k) |
---|
| 759 | zx_p = p(i,k) |
---|
| 760 | zx_q = shb(i,k) |
---|
| 761 | zdelta=MAX(0.,SIGN(1.,RTT-zx_t)) |
---|
| 762 | zcvm5 = R5LES*RLVTT*(1.-zdelta) + R5IES*RLSTT*zdelta |
---|
| 763 | zcvm5 = zcvm5 / RCPD / (1.0+RVTMP2*zx_q) |
---|
| 764 | zx_qs= r2es * FOEEW(zx_t,zdelta)/zx_p |
---|
| 765 | zx_qs=MIN(0.5,zx_qs) |
---|
| 766 | zcor=1./(1.-retv*zx_qs) |
---|
| 767 | zx_qs=zx_qs*zcor |
---|
| 768 | zx_gam = FOEDE(zx_t,zdelta,zcvm5,zx_qs,zcor) |
---|
| 769 | shbs(i,k) = zx_qs |
---|
| 770 | gam(i,k) = zx_gam |
---|
| 771 | c |
---|
| 772 | zx_t = tb(i,km1) |
---|
| 773 | zx_p = p(i,km1) |
---|
| 774 | zx_q = shb(i,km1) |
---|
| 775 | zdelta=MAX(0.,SIGN(1.,RTT-zx_t)) |
---|
| 776 | zcvm5 = R5LES*RLVTT*(1.-zdelta) + R5IES*RLSTT*zdelta |
---|
| 777 | zcvm5 = zcvm5 / RCPD / (1.0+RVTMP2*zx_q) |
---|
| 778 | zx_qs= r2es * FOEEW(zx_t,zdelta)/zx_p |
---|
| 779 | zx_qs=MIN(0.5,zx_qs) |
---|
| 780 | zcor=1./(1.-retv*zx_qs) |
---|
| 781 | zx_qs=zx_qs*zcor |
---|
| 782 | zx_gam = FOEDE(zx_t,zdelta,zcvm5,zx_qs,zcor) |
---|
| 783 | shbs(i,km1) = zx_qs |
---|
| 784 | gam(i,km1) = zx_gam |
---|
| 785 | C |
---|
| 786 | sb (i,k ) = sb(i,k ) + ds2(i) |
---|
| 787 | sb (i,km1) = sb(i,km1) + ds3(i) |
---|
| 788 | hb (i,k ) = sb(i,k ) + RLVTT*shb(i,k) |
---|
| 789 | hb (i,km1) = sb(i,km1) + RLVTT*shb(i,km1) |
---|
| 790 | hbs(i,k ) = sb(i,k ) + RLVTT*shbs(i,k ) |
---|
| 791 | hbs(i,km1) = sb(i,km1) + RLVTT*shbs(i,km1) |
---|
| 792 | C |
---|
| 793 | sbh (i,k) = 0.5*(sb(i,k) + sb(i,km1)) |
---|
| 794 | shbh(i,k) = qhalf(shb(i,km1),shb(i,k) |
---|
| 795 | $ ,shbs(i,km1),shbs(i,k)) |
---|
| 796 | hbh (i,k) = sbh(i,k) + RLVTT*shbh(i,k) |
---|
| 797 | sbh (i,km1) = 0.5*(sb(i,km1) + sb(i,k-2)) |
---|
| 798 | shbh(i,km1) = qhalf(shb(i,k-2),shb(i,km1), |
---|
| 799 | $ shbs(i,k-2),shbs(i,km1)) |
---|
| 800 | hbh (i,km1) = sbh(i,km1) + RLVTT*shbh(i,km1) |
---|
| 801 | ENDIF |
---|
| 802 | ENDDO |
---|
| 803 | C |
---|
| 804 | C Ensure that dzcld is reset if convective mass flux zero |
---|
| 805 | C specify the current vertical extent of the convective activity |
---|
| 806 | C top of convective layer determined by size of overshoot param. |
---|
| 807 | C |
---|
| 808 | 60 CONTINUE |
---|
| 809 | DO i=1,klon |
---|
| 810 | etagt0 = eta(i).gt.0.0 |
---|
| 811 | IF (.not.etagt0) dzcld(i) = 0.0 |
---|
| 812 | IF (etagt0 .and. beta(i).gt.betamn) THEN |
---|
| 813 | ktp = km1 |
---|
| 814 | ELSE |
---|
| 815 | ktp = k |
---|
| 816 | ENDIF |
---|
| 817 | IF (etagt0) THEN |
---|
| 818 | cnt(i) = MIN(cnt(i),ktp) |
---|
| 819 | cnb(i) = MAX(cnb(i),k) |
---|
| 820 | ENDIF |
---|
| 821 | ENDDO |
---|
| 822 | 70 CONTINUE ! end of k loop |
---|
| 823 | C |
---|
| 824 | C determine whether precipitation, prec, is frozen (snow) or not |
---|
| 825 | C |
---|
| 826 | DO i=1,klon |
---|
| 827 | IF (tb(i,klev).LT.tmelt .AND. tb(i,klev-1).lt.tmelt) THEN |
---|
| 828 | cmfprs(i) = prec(i)*rdt |
---|
| 829 | ELSE |
---|
| 830 | cmfprt(i) = prec(i)*rdt |
---|
| 831 | ENDIF |
---|
| 832 | ENDDO |
---|
| 833 | C |
---|
| 834 | RETURN ! we're all done ... return to calling procedure |
---|
| 835 | END |
---|