Aerosol Activation Example

Overview

This example is based on AerosolActivation module which is a part of the CloudMicrophysics.jl package. The AerosolActivation module computes the total number and mass of aerosol particles that get activated and become cloud droplets, given the atmospheric conditions and the initial aerosol size distribution and properties. See the AerosolActivation module docs for derivation and description of all input parameters.

In this example we use the ensemble Kalman methods to learn two parameters that describe the chemical composition of aerosol particles based on the observed total number and mass of activated particles. The AerosolActivation model is used here in a "perfect model" setting, meaning that the observations are generated by the same module we are calibrating.

Prerequisites

The example depends on some standard Julia libraries, as well as the CliMA packages: EnsembleKalmanProcess.jl, CLIMAParameters.jl and CloudMicrophysics.jl. To ensure that all the dependencies are met start Julia using the Project.toml file provided in the example and run the Julia package manager to download all the dependecies.

Example

We begin by importing some standard Julia modules, the Ensemble Kalman Process modules, CLIMA Parameter modules and Aerosol Activation modules.

using Plots
using Distributions
using LinearAlgebra
using Random

rng_seed = 44
rng = Random.seed!(Random.GLOBAL_RNG, rng_seed)

using EnsembleKalmanProcesses
using EnsembleKalmanProcesses.ParameterDistributions
const EKP = EnsembleKalmanProcesses

import CLIMAParameters
const CP = CLIMAParameters
struct EarthParameterSet <: CP.AbstractEarthParameterSet end
const param_set = EarthParameterSet()

import CloudMicrophysics
const AM = CloudMicrophysics.AerosolModel
const AA = CloudMicrophysics.AerosolActivation

import Thermodynamics
const TD = Thermodynamics

Next, we provide the information about the priors of the parameters we want to learn. We are calibrating two parameters decribing the aerosol properties - namely the aerosol molar mass and the osmotic coefficient.

parameter_names = ["molar_mass", "osmotic_coeff"]

In this test we do know the parameter values. We use them to test the convergence of EKP for aerosol activation.

molar_mass_true = 0.058443
osmotic_coeff_true = 0.9
default_params = [molar_mass_true, osmotic_coeff_true]

We must define parameter priors. Both parameters have to be positive definite, therefore we define the constraints to be bounded below by zero. We don't have much other prior knowledge about the parameters. We simply constrain their scale to be loosely of size 1. For more details see constrained_gaussian

prior1 = constrained_gaussian(parameter_names[1], 1, 1, 0, Inf)
prior2 = constrained_gaussian(parameter_names[2], 1, 1, 0, Inf)
priors = combine_distributions([prior1, prior2])

Next we define the atmospheric conditions for which the calibration will take place, (air temperature in K, air pressure in Pa vertical velocity in m/s and vapor specific humidity assuming its saturated in kg/kg) This can be changed later to include more than one $(T, p, w)$ combination in the calibration process

T = 283.15
p = 1e5
w = 5.0
p_vs = TD.saturation_vapor_pressure(param_set, T, TD.Liquid())
q_vs = 1 / (1 - CP.Planet.molmass_ratio(param_set) * (p_vs - p) / p_vs)
q = TD.PhasePartition(q_vs, 0.0, 0.0)

We also define the aerosol size distribution (lognormal, 1 mode) with (mean radius in m, geometric stdev, number concentration 1/m³). These can also be changed later to include different initial size distributions.

r_dry = 0.243e-6
stdev = 1.4
N = 100.0 * 1e6 # since 1/cm³ = 1e6 1/m³

Finally, we define additional parameters that describe the aerosol properties. The chosen aerosol is sea salt.

dissoc_seasalt = 2.0
soluble_mass_frac_seasalt = 1.0
rho_seasalt = 2170.0;

We define a wrapper function that runs the aerosol activation module with two input parameters that will be calibrated by EKP. The output observations are the number and mass of activated aerosol.

function run_activation_model(molar_mass_calibrated, osmotic_coeff_calibrated)

    accum_mode_seasalt = AM.Mode_B(
        r_dry,
        stdev,
        N,
        (1.0,),
        (soluble_mass_frac_seasalt,),
        (osmotic_coeff_calibrated,),
        (molar_mass_calibrated,),
        (dissoc_seasalt,),
        (rho_seasalt,),
        1,
    )

    aerosol_distr = AM.AerosolDistribution((accum_mode_seasalt,))
    N_act = AA.total_N_activated(param_set, aerosol_distr, T, p, w, q)
    M_act = AA.total_M_activated(param_set, aerosol_distr, T, p, w, q)
    return [N_act, M_act]
end

This example is run in a "perfect model setting", meaning the model that we calibrate is also used to generate observations. We use the total number and mass of activated aerosol particles as our observational data.

observation_data_names = ["N_act_and_M_act"];

We generate artificial truth samples based on the default values of parameters we are calibrating.

G_t = run_activation_model(molar_mass_true, osmotic_coeff_true)

Γy = convert(Array, LinearAlgebra.Diagonal([0.01 * G_t[1], 0.01 * G_t[2]]))
μ = zeros(length(G_t));

And add noise to the generated truth sample.

y_t = G_t .+ rand(Distributions.MvNormal(μ, Γy))

observation = EKP.Observation(Dict("samples" => y_t, "covariances" => Γy, "names" => observation_data_names))

We use 50 ensemble members and do 10 iterations.

N_ens = 50
N_iter = 10

initial_par = EKP.construct_initial_ensemble(rng, priors, N_ens)
ekiobj = EKP.EnsembleKalmanProcess(initial_par, observation, EKP.Inversion())

Finally, we can run the Ensemble Kalman Process calibration.

ϕ_n_values = []
final_it = [N_iter]
for n in 1:N_iter
    ϕ_n = EKP.get_ϕ_final(priors, ekiobj)
    G_n = [run_activation_model(ϕ_n[:, i]...) for i in 1:N_ens]
    G_ens = hcat(G_n...)
    terminate = EKP.update_ensemble!(ekiobj, G_ens)
    if !isnothing(terminate)
        final_it[1] = n - 1
        break
    end
    global ϕ_n_values = vcat(ϕ_n_values, [ϕ_n])
end
N_iter = final_it[1]

[ Info: Termination condition of scheduler `DataMisfitController` will be exceeded during the next iteration.
┌ Warning: Termination condition of scheduler `DataMisfitController` has been exceeded, returning `true` from `update_ensemble!` and preventing futher updates
│  Set on_terminate="continue" in `DataMisfitController` to ignore termination
└ @ EnsembleKalmanProcesses ~/work/EnsembleKalmanProcesses.jl/EnsembleKalmanProcesses.jl/src/LearningRateSchedulers.jl:300

We define some simple functions for plotting the data.

function plot_ensemble_scatter(id)

    ensemble_member = 1:N_ens

    if id == 1
        ylabel = "Molar mass [kg/mol]"
        filename = "molar_mass_scatter.pdf"
    elseif id == 2
        ylabel = "Osmotic coefficient [-]"
        filename = "osmotic_coeff_scatter.pdf"
    end

    plot(
        ensemble_member,
        ϕ_n_values[1][id, 1:N_ens],
        seriestype = :scatter,
        xlabel = "Ensemble Number",
        ylabel = ylabel,
        legend = false,
    )

    for it in 2:N_iter
        plot!(ensemble_member .+ ((it - 1) * 50), ϕ_n_values[it][id, 1:N_ens], seriestype = :scatter, legend = false)
    end

    current()
    savefig(filename)
end

function plot_ensemble_means(id)

    number_of_iters = 1:N_iter
    means = zeros(N_iter)

    for it in 1:N_iter
        means[it] = mean(ϕ_n_values[it][id, 1:N_ens])
    end

    if id == 1
        ylabel = "Molar mass [kg/mol]"
        filename = "molar_mass_average.pdf"
    end
    if id == 2
        ylabel = "Osmotic coefficient [-]"
        filename = "osmotic_coeff_average.pdf"
    end

    plot(
        number_of_iters,
        means,
        markershape = :star5,
        xticks = number_of_iters,
        xlabel = "Iteration Number",
        ylabel = ylabel,
        label = "Ensemble Mean",
    )
    hline!([default_params[id]], label = "true value")

    savefig(filename)
end

We plot the ensemble members and the ensemble mean for the molar mass and osmotic coefficient.

plot_ensemble_scatter(1)
plot_ensemble_means(1)
plot_ensemble_scatter(2)
plot_ensemble_means(2)

"/home/runner/work/EnsembleKalmanProcesses.jl/EnsembleKalmanProcesses.jl/docs/build/literated/osmotic_coeff_average.pdf"

Finally, we test that the parameter values obtained via EnsembleKalmanProcesses.jl are close to the known true parameter values.

molar_mass_ekp = round(mean(ϕ_n_values[N_iter][1, 1:N_ens]), digits = 6)
osmotic_coeff_ekp = round(mean(ϕ_n_values[N_iter][2, 1:N_ens]), digits = 6)

println("Molar mass [kg/mol]: ", molar_mass_ekp, " vs ", molar_mass_true)
println("Osmotic coefficient [-]: ", osmotic_coeff_ekp, " vs ", osmotic_coeff_true)

Molar mass [kg/mol]: 0.110116 vs 0.058443
Osmotic coefficient [-]: 0.717813 vs 0.9

This page was generated using Literate.jl.