CPLUOS Astrophysics team meeting

Monday 24 Jan 2022, 16:30 → 18:10 Asia/Seoul

물리학과

https://uos-ac-kr.zoom.us/j/8264902652

16:30 → 17:50 COSMOLOGY¶

16:30 Dr.Sangnam Park's Report - FDM_Offset¶

Speaker: Sangnam Park

16:40 Hyeonmo's report - FDM vs CDM Halo Collision¶

Speaker: Hyeonmo Koo (University of Seoul)

16:50 Young's report. FoF and MST algorithm¶

Speaker: Young Ju

FoF and MST algorithm

• Controlled random data 2 : Halo mass function + NFW profile
  • Get NFW profile : radius range = 1~10^4 (kpc)
  • 50 halo ; total number of galaxies = 4512
  • Increase number of halo : 50 → 500 ; total number of galaxies = 50257
  • Linking-Length = 4.0
    • For FoF : mean-separation x 0.74
    • Member > 50, (MST, FoF, MGS, DBSCAN, Hierarchical)
      • (455, 397, 475, 483, 453)
  • fResolution - number of cluster graph(MGS), black line = 500
  • Linking-Length - number of cluster graph(FoF), black line = 500

17:00 CHOA(Cosmology of High-Order Statistics)¶

Speaker: Sumi Kim (University of Seoul)

17:10 Seyeon's report¶

Speaker: Se Yeon Hwang

CLML(Cosmology with Large scale structure using Machine Learning)

• COSMOLOGICAL PARAMETER ESTIMATION FROM LARGE-SCALE STRUCTURE DEEP LEARNING
• Parameter option
  • 0.16 <= Ω_m <= 0.46 (step size = 0.01) >> total number = 31
  • 0.4 <= σ_B <= 1.1 (step size = 0.02) >> total number = 36
  • 1) Parameter space
    • Total number of simulation : 1116(=31*36)
      • Used 1000 for training, 116 for test)
  • 2) Different random seed at each case
• 3d distribution of dark matter halos
  • Ω_m = 0.23, σ_B = 0.84
  • Ω_m = 0.39, σ_B = 0.56
• Data preparation
  • 1) From the Pinocchio
    • Get particle position(x, y, z)
    • data.shape = [data_length, 3]
  • 2) Density field
    • np.histogramdd(file_name, bins = (32, 32, 32))
    • data.shape = [32, 32, 32]
• About Pinocchio resolution
  • 1) Box size = 256(Mpc/h), Grid size = 128
    • Total time of 1116 simulations : 1.9hr
  • 2) Box size = 256(Mpc/h), Grid size = 256
    • Total time of 1116 simulations : 8hr
• Model
  • In the paper..[arXiv : 1908.10590]
  • Input data = 3d histogram
  • Output data = 2parameter(Ω_m, σ_B)
• Box size : 256, Grid size : 128, Epoch : 100, Learning rate : 0.0001
• Box size : 256, Grid size : 128, Epoch : 100, Learning rate : 0.0005
• Box size : 256, Grid size : 128, Epoch : 100, Learning rate : 0.005
• Next step
  • 256(Mpc/h) → Make 32 * 32 * 32 histogram → So, one voxel includes 8(Mpc/h)'s information
• Try one voxel includes 2(Mpc/h)'s information by cutting the histogram and see the result.

17:20 Hannah's Report - Weak Lensing with Machine Learning¶

Speaker: Hannah Jhee (University of Seoul)

Weak Lensing Mass Reconstruction with cGAN

• Our goal
  • Our goal is expect the mass of the lens!
  • Compared to KS93(analytic), CNN finds the center better.
  • How about cGAN(conditional GAN)?
  • First try : U-Net + PatchGAN + L1_loss.
• Generator
  • U-net = Downsampling + (Upsampling + Skip)
  •                   downsize               upsize     concatenate?
  • Weak-lensing map (3, 400, 400) channel(g1, g2, error) →  Concatenate → [FAKE] Reconstruction  map (400x400)
  • (64, 200, 200) → Concatenate → (64, 200, 200)
  • (128, 100, 100) → Concatenate → (128, 100, 100)
  • 25x25
  • U-net = Downsampling + (Upsampling + Skip) → 'tanh' graph
• Discriminator
• Loss
  • G = Generator
  • D = Discriminator
  • λ = 101
  • y = true image
  • z = noise vector
  • Batch size 16(can't afford bigger)
  • Learning rate 2e-4
  • Took more than 30 min per epoch.
  • Maybe generator is giving up?
• Should check loss function

17:35 Dr.Sabiu's Report - N-PCF¶

Speaker: Dr Cristiano Sabiu (University of Seoul)

17:30 → 17:40 Discussion¶