aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorAaron LI <aly@aaronly.me>2017-10-19 13:21:01 +0800
committerAaron LI <aly@aaronly.me>2017-10-19 13:21:01 +0800
commit128d5e3745a19c956556bfbecbd66ee8f1fb5dc2 (patch)
tree4bea5242158821dd583bbb3ee59d9caaa677298c
parent08e33b8617f4f702a97e3aeacc5260b2254c7523 (diff)
downloadfg21sim-128d5e3745a19c956556bfbecbd66ee8f1fb5dc2.tar.bz2
clusters: Update configuration defaults and description
-rw-r--r--fg21sim/configs/20-extragalactic.conf.spec24
1 files changed, 12 insertions, 12 deletions
diff --git a/fg21sim/configs/20-extragalactic.conf.spec b/fg21sim/configs/20-extragalactic.conf.spec
index df5ea86..2c879ac 100644
--- a/fg21sim/configs/20-extragalactic.conf.spec
+++ b/fg21sim/configs/20-extragalactic.conf.spec
@@ -20,12 +20,12 @@
# The model of the fitting function for halo mass distribution
# For all models and more details:
# https://hmf.readthedocs.io/en/latest/_autosummary/hmf.fitting_functions.html
- model = option("jenkins", "ps", "smt", default="smt")
+ model = option("smt", "jenkins", "ps", default="smt")
- # The minimum (inclusive) and maximum (exclusive) halo mass (dark
+ # The minimum (inclusive) and maximum (exclusive!) halo mass (dark
# matter only) within which to calculate the halo mass distribution.
# Unit: [Msun]
- M_min = float(default=1e13, min=1e10, max=1e14)
+ M_min = float(default=1e12, min=1e10, max=1e14)
M_max = float(default=1e16, min=1e14, max=1e18)
# The logarithmic (base 10) step size for the halo masses; therefore
# the number of intervals is: (log10(M_max) - log10(M_min)) / M_step
@@ -60,8 +60,8 @@
# mass, redshift, position, shape, and the recent major merger info.
catalog_outfile = string(default=None)
- # Directly use the (previously simulated) catalog file specified
- # as the above "catalog_outfile" option.
+ # Whether to directly use the (previously simulated) catalog file
+ # specified as the above "catalog_outfile" option?
# NOTE:
# By using an existing catalog, the steps to derive these data are
# simply skipped.
@@ -96,7 +96,7 @@
# The minimum mass for clusters when to determine the galaxy clusters
# total counts and their distributions.
# Unit: [Msun]
- mass_min = float(default=2e14, min=1e12)
+ mass_min = float(default=1e14, min=1e12)
# Boost the number of expected cluster number within the sky coverage
# by the specified times.
@@ -150,13 +150,13 @@
gamma_max = float(default=1e5)
# Number of momentum points/cells for solving the Fokker-Planck
# equation.
- gamma_np = integer(default=200, min=50)
+ gamma_np = integer(default=200, min=100)
- # Number of grid points used as the buffer region near the lower
- # boundary, and the value within this buffer region will be fixed to
- # avoid unphysical pile-up of low-energy electrons.
- # Reference: Donnert & Brunetti 2014, MNRAS, 443, 3564, Sec.(3.3)
- buffer_np = integer(default=5, min=0)
+ # Number of cells used as the buffer regions near both the lower
+ # and upper boundaries, and the value within the buffer regions will
+ # be fixed to avoid unphysical pile-ups.
+ # It is suggested to be about 5%-10% of the above ``gamma_np``.
+ buffer_np = integer(default=10, min=0)
# Time step for solving the Fokker-Planck equation
# Unit: [Gyr]