[NDSL] Aer Activation

[FlorianDeconinck]

Port the Aer Activation component of Moist to pyMoist

Original code: aer_actv_single_moment.F90:AerActivation

Parent: GEOS-ESM/SMT-Nebulae#57

---

• Port Aer Activation to pyMoist
• Generate dataset that actually looks at non-clamped values for NACTL and non-zero values for NWFA
• Validate on finer dataset

---

Breakdown:

Aer_Activation

• GetActFrac
  • ActFrac_Mat

[mattweil11]

State of Validation:

• Data set not good enough for NACTL or NWFA validation
  • Line 285 in the Fortran acts as a clamping feature making NACTL hard to validate. Need different data set to truly validate NACTL.
  • NWFA validation data is always zero. No variance in the validation data.
• Variables from Fortran that are serialized that are never used in the Python Port according to the translate tests
  • AERO_F_DUST
  • AERO_F_ORGANIC
  • PLE
  • AERO_F_SOOT
  • ZLE0
  • ZL0
  • AERO_DENSITY
  • n_modes --> technically a global constant, but due to how the translate tests is setup, has to be included in the tests
  • SH
  • EVAP
  • KPBL
• Potential dead code in the Fortran:
  • Lines 238-244 (JM/IM loop) is reran in lines 252-257 (IM/JM/KM loop) so lines 238-244 does not do anything to our understanding
  • Literals 'WC', 'BB' and 'RAUX' are never used - Defined in lines 233-235

[FlorianDeconinck]

PR up: #964

[FlorianDeconinck]

4D fields are not handled by the Translate test. We are doing allocation by hand, resulting in not quantities in code and breaking of the auto-uploaded to GPU.

[mattweil11]

GPU and CPU benchmarks before reintegration:

CPU Range: 8.113 ms - 14.404 ms / median time: 11.2585 ms
GPU Range: 2.727 ms - 2.905 ms / median time: 2.816 ms

Range of Efficiencies:
Median times: GPU 74.98% faster / 4.00x
Minimum times: GPU 66.4% faster / 2.976x
Maximum times: GPU 79.8% faster / 4.957x

Spec (Discover)
CPU: AMD EPYC 7402 (24 cores, released 2020)
GPU: Nvidia A100 40GB (released 2021)

[FlorianDeconinck]

Handled 4D field by duplicating the IJK quantity factory and adding an extra dimension. This means we still need to do the Translate test memory translation by hand, but at least the code has been re-written with the full Quantity standard and is now ready for cpu/gpu.

[FlorianDeconinck]

Orchestration do not validate
GPU do not validate
dace:cpu do not validate

[FlorianDeconinck]

C180_L137 Run:

• unstable, crashes w/ floating point exception after 1 calculation in Microphysics downstream
• is it the numerics? is it the bridge?

[FlorianDeconinck]

Bridge fixed.
Stable for 48h of simulation on CPU.
Performance on par with Fortran on gt:dace:cpu

[FlorianDeconinck]

Test on GPU for lower resolution shows stability on gt:dace:gpu

[FlorianDeconinck]

Validation with multi-model metric + in-situ simulation shows good stability.