forcing.py

###########################################################
# Generate forcing including atmospheric, runoff etc.
###########################################################
import xarray as xr
import numpy as np
import pandas as pd
import glob
import os
import shutil
from .utils import distance_btw_points, closest_point, convert_to_teos10, fix_lon_range, dewpoint_to_specific_humidity, distance_to_bdry, remove_disconnected
from .grid import get_coast_mask, get_icefront_mask
from .ics_obcs import fill_ocean
from .interpolation import regrid_era5_to_cesm2, extend_into_mask, regrid_to_NEMO, neighbours, interp_latlon_cf
from .file_io import find_cesm2_file, find_processed_cesm2_file
from .constants import temp_C2K, rho_fw, cesm2_ensemble_members, sec_per_day, sec_per_hour, months_per_year, hours_per_day, land_ice_point0

# Function subsets global forcing files from the same grid to the new domain, and fills any NaN values with connected 
# nearest neighbour and then fill_val.
# Inputs:
# file_path: string of location of forcing file
# nemo_mask: xarray Dataset of nemo meshmask (must contain tmask)
# fill_ocn: boolean to turn on (or off) the connected nearest neighbour fill
# fill_val: float or NaN value to fill any remaining NaN values with
# Returns an xarray dataset of the original forcing file subset to the new domain and filled
def subset_global(file_path, nemo_mask, fill_ocn=False, fill_val=0):

    # this subset is not generalized for other domains; can fix later
    ds = xr.open_dataset(f'{file_path}').isel(y=slice(0,453)) 

    if fill_ocn:
        for var in list(ds.keys()):
            # Check for values that are NaN and in the ocean and fill with nearest neighbour
            ds_filled = fill_ocean(ds.isel(time_counter=0), var, nemo_mask, dim='2D', fill_val=fill_val)
            ds[var]   = ds[var].dims, ds_filled[var].values[np.newaxis, ...]   

    for var in list(ds.keys()):
        # Then fill any NaN values with fill_val
        ds[var] = xr.where(np.isnan(ds[var]), fill_val, ds[var])
    
    new_file_path = file_path.replace(file_path.split('/')[-1], f"AntArc_{file_path.split('/')[-1]}") 

    # save file with time_counter as unlimited dimension, if time_counter is present
    if 'time_counter' in ds.dims:
        ds.to_netcdf(f"{new_file_path}", unlimited_dims='time_counter')
    else:
        ds.to_netcdf(f"{new_file_path}")
    
    return ds

# Function identifies locations where calving does not occur at the ocean edge of the iceshelf front 
# Inputs:
# calving: xarray variable of calving proportions (2D)
# nemo_mask: xarray Dataset of nemo meshmask (must contain tmask and tmaskutil)
# Returns three 2D arrays of calving that occurs not at the icefront, but further in the ocean (calving_ocn), 
# on land (calving_land), or on an ice shelf grid point (calving_ice)
def calving_at_coastline(calving, nemo_mask):
    
    calving   = xr.where(calving > 0, calving, np.nan)

    # Boolean arrays to identify regions:
    ocean     = (nemo_mask.tmask.isel(time_counter=0, nav_lev=0) == 1)
    iceshelf  = (nemo_mask.tmaskutil.isel(time_counter=0) - nemo_mask.tmask.isel(time_counter=0, nav_lev=0)).astype(bool);
    land      = (nemo_mask.tmaskutil.isel(time_counter=0) == 0)

    # Cases where calving does not occur at the ocean edge of the icefront:
    icefront_mask_ocn = get_icefront_mask(nemo_mask, side='ocean')
    calving_ocn  = xr.where((~icefront_mask_ocn & ocean), calving, np.nan)  # calving occurs in the ocean but not right by the icefront
    calving_land = xr.where(land      , calving, np.nan)  # calving occurs on land
    calving_ice  = xr.where(iceshelf  , calving, np.nan)  # calving occurs on ice shelf

    return calving_ocn, calving_land, calving_ice

# Function shifts the x,y location of a calving point to the nearest iceshelf front ocean point 
# Inputs:
# calving: xarray dataset containing calving proportions 
# mask: locations of cells that need to be shifted
# nemo_mask: xarray dataset of nemo mesh mask file (must contain nav_lon, nav_lat)
# icefront_mask_ocn: mask of ocean points nearest to iceshelf, produced by get_icefront_mask
# max_distance: float of maximum distance (in meters) that an ocean calving point will get moved
# calving_var: string of the calving variable name
def shift_calving(calving, mask, nemo_mask, icefront_mask_ocn, max_distance=2e6, calving_var='soicbclv', ocn=False):
    # NEMO domain grid points
    x, y        = np.meshgrid(nemo_mask.nav_lon.x, nemo_mask.nav_lon.y)
    calving_x   = x[(~np.isnan(mask))]
    calving_y   = y[(~np.isnan(mask))]
    calving_new = np.copy(calving[calving_var].values);

    ice_shelf  = (nemo_mask.tmaskutil.isel(time_counter=0) - nemo_mask.tmask.isel(time_counter=0, nav_lev=0)).astype(bool)
    open_ocean = (nemo_mask.tmask.isel(time_counter=0, nav_lev=0) == 1)
    land       = ~open_ocean
    # Return ocean points with at least 3 ice shelf neighbour
    num_ice_shelf_neighbours = neighbours(ice_shelf, missing_val=0)[-1]
    num_land_neighbours      = neighbours(land, missing_val=0)[-1] # also includes the iceshelf points
    confined_points          = (open_ocean*(num_land_neighbours > 2)).astype(bool)

    # Coordinates of ocean points closest to iceshelf front 
    # increase number of possible points by shifting the icefront mask:
    ocean_neighbours       = neighbours(icefront_mask_ocn, missing_val=0)[-1]
    icefront_mask_extended = (((icefront_mask_ocn)+(ocean_neighbours))*open_ocean).astype(bool)
    icefront_x             = x[icefront_mask_extended]
    icefront_y             = y[icefront_mask_extended]
    icefront_coord         = (nemo_mask.nav_lon.values[icefront_mask_extended], nemo_mask.nav_lat.values[icefront_mask_extended])

    # For each land iceberg calving point, check distance to nearest iceshelf front and move closer if possible:
    for index in list(zip(calving_y, calving_x)):
     
        calving_coord = (nemo_mask.nav_lon.values[index], nemo_mask.nav_lat.values[index])
        distances     = distance_btw_points(calving_coord, icefront_coord)
        distances[distances < 2e4] = np.nan
        # only move cell if it is within a certain distance
        if np.nanmin(np.abs(distances)) < max_distance:     
            new_x         = icefront_x[np.nanargmin(np.abs(distances))]
            new_y         = icefront_y[np.nanargmin(np.abs(distances))]
            # Move calving to the nearest icefront point and add to any pre-existing calving at that point
            calving_new[(new_y, new_x)] = calving_new[(new_y, new_x)] + calving_new[index]    
            calving_new[index]          = 0 # remove calving from originating point

    # For each ocean iceberg calving point, check that it isn't surrounded on three sides by iceshelf points or land points
    # to prevent accumulation of icebergs in small coastal regions
    if ocn:
        # Points that are ocean that have at least three iceshelf neighbour points, these are the ones that I'll need to move
        confined_x = x[confined_points*icefront_mask_ocn]
        confined_y = y[confined_points*icefront_mask_ocn]
        unconfined_x = x[~confined_points*open_ocean]
        unconfined_y = y[~confined_points*open_ocean]
        unconfined_coord = (nemo_mask.nav_lon.values[~confined_points*open_ocean], nemo_mask.nav_lat.values[~confined_points*open_ocean])
        
        for ind in list(zip(confined_y, confined_x)):
            confined_coord = (nemo_mask.nav_lon.values[ind], nemo_mask.nav_lat.values[ind])
            distances      = distance_btw_points(confined_coord, unconfined_coord)#icefront_coord)
            distances[distances < 2e4] = np.nan

            new_x = unconfined_x[np.nanargmin(np.abs(distances))]
            new_y = unconfined_y[np.nanargmin(np.abs(distances))]
            # Move calving to the nearest icefront point and add to any pre-existing calving at that point
            calving_new[new_y, new_x] = calving_new[new_y, new_x] + calving_new[ind]    
            calving_new[ind]        = 0 # remove calving from originating point

    # Write new locations to xarray dataset
    calving_ds = calving.copy()
    calving_ds[calving_var] = calving[calving_var].dims, calving_new

    return calving_ds

# Main function to move pre-existing calving dataset to a new coastline (on the same underlying grid, but subset)
# Inputs:
# nemo_mask: xarray dataset of nemo meshmask file (must contain nav_lon, nav_lat, tmask, tmaskutil)
# calving: xarray dataset of old calving forcing NetCDF file
# calving_var: string of the calving variable name
# new_file_path: string of name and location of new calving forcing file
# Returns: xarray dataset with calving at new locations
def create_calving(calving, nemo_mask, calving_var='soicbclv', new_file_path='./new-calving.nc'):

    # Identify locations with calving that need to be moved
    calving_ocn, calving_land, calving_ice = calving_at_coastline(calving[calving_var], nemo_mask)

    # Mask of ocean grid points nearest to icefront, to move calving to
    icefront_mask_ocn = get_icefront_mask(nemo_mask, side='ocean')
    icefront_mask_ice = get_icefront_mask(nemo_mask, side='ice')

    # Shift calving points to the iceshelf edge
    calv_ocn_new  = shift_calving(calving       , calving_ocn , nemo_mask, icefront_mask_ocn, ocn=True) # from ocean
    calv_land_new = shift_calving(calv_ocn_new  , calving_land, nemo_mask, icefront_mask_ocn, ocn=False) # from land
    calv_ice_new  = shift_calving(calv_land_new , calving_ice , nemo_mask, icefront_mask_ocn, ocn=False) # from ice

    # Check if the icebergs calve in regions shallower than the minimum initial iceberg thickness (40 m)
    calving_depth    = nemo_mask.bathy_metry.squeeze().values*(calv_ice_new[calving_var].squeeze().values.astype(bool))
    if np.sum((calving_depth < 40)*(calving_depth > 0)) != 0:
        print('Warning: number of cells with calving shallower than minimum iceberg thickness is, ', np.sum((calving_depth < 40)*(calving_depth > 0)))

    # Write the calving dataset to a netcdf file:
    calv_ice_new[calving_var] = ('time_counter',) + calving[calving_var].dims, calv_ice_new[calving_var].values[np.newaxis, ...] 
    calv_ice_new.to_netcdf(f"{new_file_path}", unlimited_dims='time_counter')

    # Check that the amount of calving that occurs in the new file is approximately equal to the original file:
    # allow for a tolerance of 0.1% (although it should be essentially equal within machine precision)
    tolerance = 0.001*(np.sum(calv_ice_new[calving_var].values)) 
    if np.abs(np.sum(calv_ice_new[calving_var].values) - np.sum(calving[calving_var].values)) >= tolerance:
        raise Exception('The total amount of calving in the new file is not equal to the original total', \
                np.sum(calv_ice_new[calving_var].values), np.sum(calving[calving_var].values))
    
    return calv_ice_new

# Process ocean conditions from CESM2 scenarios for a single variable and single ensemble member (for initial and boundary conditions).
def cesm2_ocn_forcing (expt, var, ens, out_dir, start_year=1850, end_year=2100):

    if expt not in ['LE2']:
        raise Exception(f'Invalid experiment {expt}')

    freq = 'monthly'
    for year in range(start_year, end_year+1):
        # read in the data and subset to the specified year
        if var in ['aice','sithick','sisnthick']:
            if year > 1850:
                file_path_prev = find_cesm2_file(expt, var, 'ice', freq, ens, year-1)
            file_path = find_cesm2_file(expt, var, 'ice', freq, ens, year)
        else:
            if year > 1850:
                file_path_prev = find_cesm2_file(expt, var, 'ocn', freq, ens, year-1) # load last month from previous file
            file_path = find_cesm2_file(expt, var, 'ocn', freq, ens, year)
        
        if year==1850:
            ds = xr.open_dataset(file_path)
        else:
            if file_path_prev != file_path:
                ds = xr.open_mfdataset([file_path_prev, file_path])
            else:
                ds = xr.open_dataset(file_path)
        data = ds[var].isel(time=(ds.time.dt.year == year))

        # Unit conversions ### need to check that these are still consistent between CESM1 and CESM2
        if var in ['SSH','UVEL','VVEL']:
            # convert length units from cm to m
            data *= 0.01 
        # Convert from practical salinity to absolute salinity??
        elif var == 'TEMP':
            # Convert from potential temperature to conservative temperature
            # need to load absolute salinity from file
            salt_path = find_cesm2_file(expt, 'SALT', 'ocn', freq, ens, year)
            salt_ds   = xr.open_dataset(salt_path)
            salt_data = salt_ds['SALT'].isel(time=(salt_ds.time.dt.year == year))

            dsc  = xr.Dataset({'PotTemp':data, 'AbsSal':salt_data, 'depth':data.z_t, 'lon':data.TLONG, 'lat':data.TLAT})
            data = convert_to_teos10(dsc, var='PotTemp')

        # Convert calendar to Gregorian
        data = data.convert_calendar('gregorian')
        # Convert longitude range from 0-360 to degrees east
        if var in ['SSH','SALT','TEMP']:
            lon_name = 'TLONG'
        elif var in ['aice','sithick','sisnthick']:
            lon_name = 'TLON'
        elif var in ['UVEL','VVEL']:
            lon_name = 'ULONG'
        data[lon_name] = fix_lon_range(data[lon_name])
        
        # Convert depth (z_t) from cm to m
        if var in ['SALT','TEMP','UVEL','VVEL']:
            data['z_t'] = data['z_t']*0.01
            data['z_t'].attrs['units'] = 'meters'
            data['z_t'].attrs['long_name'] = 'depth from surface to midpoint of layer'
  
        # Mask sea ice conditions based on tmask (so that land is NaN and no ice areas are zero)
        if var in ['sithick','sisnthick']:
            data = data.fillna(0)
            data = data.where((ds.isel(time=(ds.time.dt.year == year)).tmask.fillna(0).values) != 0)

        # Change variable names and units in the dataset:
        if var=='TEMP':
            varname = 'ConsTemp'
            data.attrs['long_name'] ='Conservative Temperature'
        elif var=='SALT':
            varname = 'AbsSal'
            data = data.rename(varname)
            data.attrs['long_name'] ='Absolute Salinity'
        else:
            varname=var
            if var=='SSH':
                data.attrs['units'] = 'meters'
            elif var in ['UVEL','VVEL']:
                data.attrs['units'] = 'm/s'

        # Write data
        out_file_name = f'{out_dir}CESM2-{expt}_ens{ens}_{varname}_y{year}.nc'
        data.to_netcdf(out_file_name, unlimited_dims='time')

    return

# Convert wind velocities from the lowest atmospheric model level grid cell to the 10 m wind level 
# (CESM2 output UBOT and VBOT is at 992 hPa) by using the wind speed magnitude from the variable U10 and the wind directions from UBOT and VBOT
def UBOT_to_U10_wind(UBOT, VBOT, U10):

    # calculate the angle between the UBOT and VBOT wind vectors
    theta = np.arctan2(VBOT, UBOT) 
    # then use this angle to create the x and y wind vector components that sum to the magnitude of the U10 wind speed
    U10x = U10*np.cos(theta)
    U10y = U10*np.sin(theta)

    return U10x, U10y

# Get CESM2 wind velocities from the wind speed at 10 meters and the direction from the surface stress 
# (because wind speed vectors are not available for the CESM2 single forcing experiments at daily frequency)
def TAU_to_U10xy(TAUX, TAUY, U10):

    # calculate the angle between the surface stress vectors (need to take opposite direction because of definition difference)
    theta = np.arctan2(-1*TAUY, -1*TAUX)
    # then use this angle to create the x and y wind vector components that sum to the magnitude of the U10 wind speed
    U10x = U10*np.cos(theta)
    U10y = U10*np.sin(theta)

    return U10x, U10y

# Process atmospheric forcing from CESM2 scenarios (LE2, etc.) for a single variable and single ensemble member.
# expt='LE2', var='PRECT', ens='1011.001' etc.
def cesm2_atm_forcing (expt, var, ens, out_dir, start_year=1850, end_year=2100, year_ens_start=1750, freq='daily', shift_wind=False):

    if expt not in ['LE2', 'piControl', 'SF-xAER', 'SF-AAER', 'SF-BMB', 'SF-GHG', 'SF-EE']:
        raise Exception('Invalid experiment {expt}')

    # for each year, read in the data, process it, and save the processed forcing to a file
    for year in range(start_year,  end_year+1):
        
        # helper function to load CESM2 file and select variable for the specified year
        def load_cesm2_file(var, expt=expt, freq=freq, ens=ens, year=year):
            file_path = find_cesm2_file(expt, var, 'atm', freq, ens, year)
            ds_vel    = xr.open_dataset(file_path)
            data_vel  = ds_vel[var].isel(time=(ds_vel.time.dt.year == year))
            return data_vel

        # load cesm2 land-ocean mask
        land_mask  = find_cesm2_file(expt, 'LANDFRAC', 'atm', 'monthly', ens, year)
        cesm2_mask = xr.open_dataset(land_mask).LANDFRAC.isel(time=0)
        # read in variables from forcing files
        if var=='PRECS': # snowfall
            if expt=='piControl': freq='monthly' # only monthly files available
            data_conv  = load_cesm2_file('PRECSC') # convective snow rate
            data_large = load_cesm2_file('PRECSL') # large-scale snow rate
        elif var=='U10':
            if expt=='piControl': var_U = 'U'; var_V='V';
            else: var_U='UBOT'; var_V='VBOT';

            data_U10  = load_cesm2_file('U10')
            data_UBOT = load_cesm2_file(var_U)
            data_VBOT = load_cesm2_file(var_V)

            # For the piControl experiment, only full column winds are available, so select the bottom wind
            if expt=='piControl':
                data_UBOT = data_UBOT.isel(lev=-1) # bottom wind is the last entry (992 hPa)
                data_VBOT = data_VBOT.isel(lev=-1)

            # Convert the wind to the 10 m wind (corrected height)
            if shift_wind:
                Ux, Uy    = UBOT_to_U10_wind(data_UBOT, data_VBOT, data_U10)
                data_UBOT = Ux.rename('U10x')
                data_VBOT = Uy.rename('U10y')
        else:
            data = load_cesm2_file(var)
 
        # Unit conversions #
        # notes: don't think I need to swap FSDS,FLDS signs like Kaitlin did for CESM1, qrefht is specific 
        #        humidity so don't need to convert, but specify read in option in namelist_cfg
        if var=='PRECT': # total precipitation
            # Convert from m/s to kg/m2/s
            data *= rho_fw
        elif var=='PRECS': # snowfall
            # Combine convective and large scale snowfall rates and convert from m of water equivalent to kg/m2/s
            data  = (data_conv + data_large) * rho_fw        
            data  = data.rename('PRECS')

        if var=='U10':
            data_arrays = [data_UBOT, data_VBOT]
        else:
            data_arrays = [data]       

        for arr in data_arrays:
            if var in ['QREFHT','TREFHT','FSDS','FLDS','PSL','PRECT','PRECS','UBOT','VBOT']: 
                # Mask atmospheric forcing over land based on cesm2 land mask (since land values might not be representative for the ocean areas)
                arr = xr.where(cesm2_mask.values != 0, -9999, arr)
                # And then fill masked areas with nearest non-NaN latitude neighbour
                print(f'Filling land for variable {var} in year {year}')
                var_filled_array = np.empty(arr.shape)
                for tind, t in enumerate(arr.time):
                    var_filled_array[tind,:,:] = extend_into_mask(arr.isel(time=tind).values, missing_val=-9999, fill_val=np.nan, use_2d=True, use_3d=False, num_iters=100)
                   
                arr.data = var_filled_array 
     
            # Convert longitude range from (0,360) to (-180,180) degrees east
            arr['lon'] = fix_lon_range(arr['lon'])
            # CESM2 does not do leap years, but NEMO does, so fill 02-29 with 02-28        
            # Also convert calendar to Gregorian
            fill_value = arr.isel(time=((arr.time.dt.month==2)*(arr.time.dt.day==28)))
            if freq=='daily':
                arr = arr.convert_calendar('gregorian', dim='time', missing=fill_value)
            elif freq=='3-hourly': # need to do something slightly different because you need to fill a slice of times
                arr = arr.convert_calendar('gregorian', dim='time', missing=fill_value.isel(time=0))
                #if leap year, properly replace the time slice:
                if sum((arr.time.dt.month==2)*(arr.time.dt.day==29)) > 0:
                    print('leap year')
                    arr.loc[dict(time=(arr.time.dt.month==2)*(arr.time.dt.day==29))] = fill_value.data

            # Daily variables of ensemble members 004-009 of the SF-xAER experiment are missing the day 2014-12-31 --- fill with the conditions from the previous day
            if expt=='SF-xAER':
                if ens in ['004','005','006','007','008','009']:
                    fill_value = arr.isel(time=((arr.time.dt.month==12)*(arr.time.dt.day==30)))
                    if freq=='daily':
                        arr.loc[dict(time=(arr.time.dt.month==12)*(arr.time.dt.day==31))] = fill_value.data

            # Change variable names and units in the dataset:
            varname = arr.name 
            if var=='PRECS':
                arr.attrs['long_name'] ='Total snowfall (convective + large-scale)'
                arr.attrs['units'] = 'kg/m2/s'
            elif var=='PRECT':
                arr.attrs['units'] = 'kg/m2/s'
            elif var=='wind':
                if varname=='U10x':
                    arr.attrs['long_name'] = 'zonal wind at 10 m'
                elif varname=='U10y':
                    arr.attrs['long_name'] = 'meridional wind at 10 m'

            # Write data
            out_file_name = f'{out_dir}CESM2-{expt}_ens{ens}_{freq}_{varname}_y{year}.nc'
            arr.to_netcdf(out_file_name, unlimited_dims='time')
    return arr


# Create CESM2 atmospheric forcing for the given scenario, for all variables and ensemble members.
# ens_strs : list of strings of ensemble member names
def cesm2_expt_all_atm_forcing (expt, ens_strs=None, out_dir=None, start_year=1850, end_year=2100, year_ens_start=1750, shift_wind=False):
    
    if out_dir is None:
        raise Exception('Please specify an output directory via optional argument out_dir')

    var_names = ['UBOT','VBOT','FSDS','FLDS','TREFHT','QREFHT','PRECT','PSL','PRECS'] 
    for ens in ens_strs:
        print(f'Processing ensemble member {ens}')
        out_dir = f'{out_dir}ens{ens}/'
        for var in var_names:
            print(f'Processing {var}')
            # specify the forcing frequency to read in
            if var in ['UBOT', 'VBOT']:
                freq='3-hourly'
            else:
                freq='daily'
            cesm2_atm_forcing(expt, var, ens, out_dir, start_year=start_year, end_year=end_year, freq=freq, year_ens_start=year_ens_start, shift_wind=shift_wind)

    return

# Create CESM2 ocean forcing for the given scenario, for all variables and ensemble members.
# ens_strs : list of strings of ensemble member names
def cesm2_expt_all_ocn_forcing(expt, ens_strs=None, out_dir=None, start_year=1850, end_year=2100):

    if out_dir is None:
        raise Exception('Please specify an output directory via optional argument out_dir')

    ocn_var_names = ['TEMP','SALT','UVEL','VVEL','SSH']
    ice_var_names = ['aice','sithick','sisnthick']
    var_names = ocn_var_names + ice_var_names
 
    for ens in ens_strs:
        print(f'Processing ensemble member {ens}')
        for var in var_names:
            print(f'Processing {var}')
            cesm2_ocn_forcing(expt, var, ens, out_dir, start_year=start_year, end_year=end_year)

    return


# Helper function calculates the monthly time-mean over specified year range for ERA5 output (for bias correction)
# Note that the year 1996 is excluded from climatology due to an issue with a cyclone in the Amundsen Sea in ERA5
# Input: 
# - variable : string of forcing variable name (in ERA5 naming convention)
# - (optional) year_start : start year for time averaging
# - (optional) end_year   : end year for time averaging
def era5_time_mean_forcing(variable, year_start=1979, year_end=2024, freq='daily', lat_slice=slice(-90,-50),
                           era5_folder='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/ERA5-forcing/', era5_folder_in=None, processed=True, varname=None):

    if era5_folder_in is None:
        if freq=='3-hourly':
            era5_folder_in = f'{era5_folder}hourly/processed/'
        else:
            era5_folder_in = f'{era5_folder}{freq}/processed/'
    else:
        if era5_folder is None:
            era5_folder = era5_folder_in
        
    era5_folder_out = f'{era5_folder}climatology/'
    if varname is None:
        varname = variable
    if processed:
        time_dim = 'time'
        lat_dim = 'lat'
    else:
        time_dim = 'valid_time'
        lat_dim = 'latitude'

    if freq=='daily' or freq=='hourly':
        if variable in ['wind_speed', 'wind_angle']:
            era5_u  = xr.open_mfdataset(f'{era5_folder_in}u10_*')['u10'].sortby(lat_dim).sel({lat_dim:lat_slice})
            era5_v  = xr.open_mfdataset(f'{era5_folder_in}v10_*')['v10'].sortby(lat_dim).sel({lat_dim:lat_slice})
            era5_u  = era5_u.isel({time_dim:((era5_u[time_dim].dt.year <= year_end)*(era5_u[time_dim].dt.year >= year_start)*(era5_u[time_dim].dt.year != 1996))})
            era5_v  = era5_v.isel({time_dim:((era5_v[time_dim].dt.year <= year_end)*(era5_v[time_dim].dt.year >= year_start)*(era5_v[time_dim].dt.year != 1996))})
            if variable == 'wind_speed':
                era5_ds = np.hypot(era5_u, era5_v).rename(variable).to_dataset()
            elif variable == 'wind_angle':
                era5_ds = np.arctan2(era5_v, era5_u).rename(variable).to_dataset()
        else:
            try:
                era5_ds = xr.open_mfdataset(f'{era5_folder_in}{variable}_*.nc')[varname].sortby(lat_dim).sel({lat_dim:lat_slice})
            except(KeyError):
                if variable == 'sph2m':
                    varname = 'specific_humidity'
                era5_ds = xr.open_mfdataset(f'{era5_folder_in}{variable}_*.nc')[varname].sortby(lat_dim).sel({lat_dim:lat_slice})
            era5_ds = era5_ds.isel({time_dim:((era5_ds[time_dim].dt.year <= year_end)*(era5_ds[time_dim].dt.year >= year_start)*(era5_ds[time_dim].dt.year != 1996))})

        time_mean = era5_ds.groupby(time_dim+'.month').mean(dim=time_dim)

    elif freq=='3-hourly':
        # slow due to resampling if you use the same approach as for daily, so do the below instead
        for year in range(year_start, year_end+1):
            if year == 1996:
                print('skipping year 1996')
            else:
                print(year)
                if variable in ['wind_speed', 'wind_angle']:
                    era5_u = xr.open_dataset(f'{era5_folder_in}u10_y{year}.nc')['u10'].sortby(lat_dim).sel({lat_dim:lat_slice})
                    era5_v = xr.open_dataset(f'{era5_folder_in}v10_y{year}.nc')['v10'].sortby(lat_dim).sel({lat_dim:lat_slice})
                    era5_u = era5_u.resample({time_dim:'3h'}).mean()
                    era5_v = era5_v.resample({time_dim:'3h'}).mean()           
                    if variable == 'wind_speed':
                        era5_ds = np.hypot(era5_u, era5_v).rename(variable).groupby(time_dim+'.month').mean(dim=time_dim).to_dataset()
                    elif variable == 'wind_angle':
                        era5_ds = np.arctan2(era5_v, era5_u).rename(variable).groupby(time_dim+'.month').mean(dim=time_dim).to_dataset()
                else:
                    try:
                        era5_ds = xr.open_dataset(f'{era5_folder_in}{variable}_y{year}.nc')[varname].sortby(lat_dim).sel({lat_dim:lat_slice})
                    except(KeyError):
                        if variable == 'sph2m':
                            varname = 'specific_humidity'
                        era5_ds = xr.open_dataset(f'{era5_folder_in}{variable}_y{year}.nc')[varname].sortby(lat_dim).sel({lat_dim:lat_slice})
                    era5_ds = era5_ds.resample({time_dim:'3h'}).mean().groupby(time_dim+'.month').mean(dim=time_dim).to_dataset()
               
                era5_ds.to_netcdf(f'{era5_folder_out}ERA5_{variable}_{freq}_monthly_mean_y{year}.nc')
 
        time_mean = xr.open_mfdataset(f'{era5_folder_out}ERA5_{variable}_{freq}_monthly_mean_y*', concat_dim='year', combine='nested')[varname].mean(dim='year')
    
    time_mean.rename(variable).to_netcdf(f'{era5_folder_out}ERA5_{variable}_{freq}_{year_start}-{year_end}_mean_monthly.nc')

    return


# Wrapper for the above to calculate ERA5 climatology for comparison with UKESM, for one variable (set var_name) or looping over all variables (default)
def era5_clim_for_ukesm (era5_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ERA5_hourly/', var_name=None):

    if var_name is not None:
        var_names = [var_name]
    else:
        var_names = ['t2m', 'sph2m', 'wind_speed', 'wind_angle', 'mtpr', 'msr', 'msl', 'msdwswrf', 'msdwlwrf']

    for variable in var_names:
        print('Processing '+variable)
        if variable == 'mtpr':
            varname = 'avg_tprate'
        elif variable == 'msr':
            varname = 'avg_tsrwe'
        elif variable == 'msdwswrf':
            varname = 'avg_sdswrf'
        elif variable == 'msdwlwrf':
            varname = 'avg_sdlwrf'
        else:
            varname = variable
        era5_time_mean_forcing(variable, year_start=1979, year_end=2014, freq='3-hourly', era5_folder_in=era5_dir, era5_folder=None, processed=False, varname=varname)


# Calculate monthly climatology for 3-hourly UKESM historical atmospheric forcing, one ensemble member (suite=cy690, cy691, cy692, cy693). 
# Options for variable = tair, qair, wind_speed, wind_angle, precip, snow, pair, swrad, lwrad
def ukesm_hist_forcing_monthly_clim (variable, suite, base_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/nemo_forcing_files/', out_dir=None, start_year=1979, end_year=2014):

    in_dir = base_dir + suite + '/'
    if out_dir is None:
        out_dir = in_dir + '/climatology/'

    if variable in ['wind_speed', 'wind_angle']:
        data_u = xr.open_mfdataset(f'{in_dir}uwind_*')['uwind']
        data_v = xr.open_mfdataset(f'{in_dir}vwind_*')['vwind']
        # Interpolate to tracer grid
        data_t = xr.open_mfdataset(f'{in_dir}tair_*')['tair']
        def interp_tgrid (data, gtype):
            # Interpolate once as usual; this will produce a stripe of missing values in the middle where longitude jumps from 180 to -180
            data_interp1 = data.rename({'longitude_'+gtype:'longitude', 'latitude_'+gtype:'latitude'}).interp_like(data_t)
            # Now move longitude to the range 0-360
            data.coords['longitude_'+gtype] = fix_lon_range(data.coords['longitude_'+gtype], max_lon=360)
            data_t.coords['longitude'] = fix_lon_range(data_t.coords['longitude'], max_lon=360)
            # Interpolate again
            data_interp2 = data.rename({'longitude_'+gtype:'longitude', 'latitude_'+gtype:'latitude'}).interp_like(data_t)
            # Move longitude back to -180 to 180
            data_interp2.coords['longitude'] = fix_lon_range(data_interp2.coords['longitude'], max_lon=180)
            data_t.coords['longitude'] = fix_lon_range(data_t.coords['longitude'], max_lon=180)
            # Fill in that gap
            data_interp = xr.where(data_interp1.isnull(), data_interp2, data_interp1)
            return data_interp
        data_u = interp_tgrid(data_u, 'u')
        data_v = interp_tgrid(data_v, 'v')        
        time_years = data_u['time'].dt.year
        time_sel = (time_years >= start_year)*(time_years <= end_year)*(time_years != 1996)
        data_u = data_u.isel({'time':time_sel})
        data_v = data_v.isel({'time':time_sel})
        if variable == 'wind_speed':
            data_ds = np.hypot(data_u, data_v).rename(variable).to_dataset()
        elif variable == 'wind_angle':
            data_ds = np.arctan2(data_v, data_u).rename(variable).to_dataset()
    else:
        data = xr.open_mfdataset(f'{in_dir}{variable}_*.nc')[variable]
        time_years = data['time'].dt.year
        time_sel = (time_years >= start_year)*(time_years <= end_year)*(time_years != 1996)
        data_ds = data.isel({'time':time_sel})
    time_mean = data_ds.groupby('time.month').mean(dim='time')
    time_mean.to_netcdf(f'{out_dir}{variable}_{start_year}-{end_year}_mean_monthly.nc')
    

# Function calculates the monthly time-mean over specified year range for mean of all CESM2 ensemble members in the specified experiment (for bias correction)
# Input:
# - expt : string of CESM2 experiment name (e.g. 'LE2')
# - variable : string of forcing variable name
# - out_dir : directory to write ensemble mean files to
# - (optional) year_start : start year for time averaging
# - (optional) end_year   : end year for time averaging
# - (optional) ensemble_members : list of strings of ensemble members to average (defaults to all the ones that have been downloaded)
def cesm2_ensemble_time_mean_forcing(expt, variable, out_dir, year_start=1979, year_end=2024, ensemble_members=cesm2_ensemble_members, freq='daily', lat_slice=slice(-90,-50)):

    print(f'Calculating ensemble mean for variable {variable} at frequency {freq} from {year_start}-{year_end} for {expt} ensemble members:', ensemble_members)

    # helper function to read in processed CESM2 file for all ensemble members, and subset domain
    def load_cesm2_ensemble(variable, year, expt=expt, freq=freq, ensemble_members=ensemble_members):
        for e, ens in enumerate(ensemble_members):
            cesm2_file = find_processed_cesm2_file(expt, variable, ens, year, freq=freq)
            cesm2_ds   = xr.open_dataset(cesm2_file)
            if e==0:
                cesm2_ds_ens = cesm2_ds.copy().expand_dims(dim='ens')
            else:
                cesm2_ds_ens = xr.concat([cesm2_ds_ens, cesm2_ds], dim='ens')

        cesm2_ds_ens = cesm2_ds_ens.sortby('lat').sel(lat=lat_slice)

        return cesm2_ds_ens

    # For each year within the specified range, calculate the monthly ensemble mean and write to file
    for year in range(year_start, year_end+1):
        print(year)
        if variable in ['wind_speed', 'wind_angle']:
            cesm2_u_ens = load_cesm2_ensemble('UBOT', year)
            cesm2_v_ens = load_cesm2_ensemble('VBOT', year)
            if variable == 'wind_speed':
                cesm2_ds_ens = np.hypot(cesm2_u_ens.UBOT, cesm2_v_ens.VBOT).rename(variable)
            elif variable == 'wind_angle':
                cesm2_ds_ens = np.arctan2(cesm2_v_ens.VBOT, cesm2_u_ens.UBOT).rename(variable)
        else:
            cesm2_ds_ens = load_cesm2_ensemble(variable, year)      

        cesm2_ds_mean = cesm2_ds_ens.mean(dim='ens').groupby("time.month").mean(dim="time")
        cesm2_ds_mean['lon'] = fix_lon_range(cesm2_ds_mean['lon'])
        cesm2_ds_mean = cesm2_ds_mean.sortby('lon')
        cesm2_ds_mean.to_netcdf(f'{out_dir}CESM2-LE2_{variable}_{freq}_y{year}_ensemble_mean_monthly.nc')

    # Next, read in the above ensemble mean files for all years and calculate the mean over the full timeseries
    file_list = [f'{out_dir}CESM2-LE2_{variable}_{freq}_y{year}_ensemble_mean_monthly.nc' for year in range(year_start, year_end+1)]
    ds_ens = xr.open_mfdataset(file_list, concat_dim='year', combine='nested')[variable]
    ds_ens.mean(dim='year').to_netcdf(f'{out_dir}CESM2-LE2_{variable}_{freq}_ensemble_{year_start}-{year_end}_mean_monthly.nc')            

    return

# wrapper function to calculate ensemble mean for each of the atmospheric variables we're interested in:
def cesm2_all_variables_ensemble_mean(expt, out_dir, year_start=1979, year_end=2024, ensemble_members=cesm2_ensemble_members):

    for variable in ['TREFHT','QREFHT','FSDS','FLDS','PRECT','PRECS', 'PSL', 'wind_speed', 'wind_angle']: 
        if variable in ['wind_speed','wind_angle','UBOT','VBOT']:
            freq='3-hourly'
        else:
            freq='daily'
        cesm2_ensemble_time_mean_forcing(expt, variable, out_dir, year_start=year_start, year_end=year_end, ensemble_members=ensemble_members, freq=freq)

    return

# Function calculate the bias correction for the atmospheric variable from the specified source type based on 
# the difference between its mean state and the ERA5 mean state. Reads in fields that were already regridded with the NEMO WEIGHTS tool
# ex. CESM2 vs. ERA5. Assumes you've ran the functions: cesm2_all_variables_ensemble_mean() to calculate the ensemble mean of CESM2 variables
# Input:
# - source : string of source type (currently only set up for 'CESM2') 
# - variable : string of the variable from the source dataset to be corrected
# - (optional) expt : cesm2 experiment set
# - (optional) year_start : start year for time averaging
# - (optional) end_year   : end year for time averaging
# - (optional) fill_land  : boolean whether or not to fill grid cells that are land in CESM2 with ocean values along lines of latitudes
# - (optional) freq       : frequency of underlying forcing files (daily, 3-hourly, or hourly)
# - (optional) monthly    : boolean whether to do monthly bias correction or one mean field without seasonality
# - (optional) era5_folder, source_folder : string paths to era5 forcing climatology and source to correct (CESM2) forcing climatology/ensemble mean
# - (optional) out_folder : string to location to save the bias correction file
def calc_bias_correction(source, variable, expt='LE2', year_start=1979, year_end=2024, freq='daily', fill_land=False, monthly=False,
                         era5_folder='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/ERA5-forcing/climatology/',
                         source_folder='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/climate-forcing/CESM2/LE2/ensemble_mean/',
                         out_folder='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/climate-forcing/CESM2/LE2/ensemble_mean/bias_corr/'):

    if source=='CESM2':
        
        # Load regridded CESM2 ensemble mean
        CESM2_ensemble_mean = xr.open_dataset(f'{source_folder}CESM2-{expt}_eANT025_{variable}_{freq}_ensemble_{year_start}-{year_end}_mean_monthly.nc')

        # Load regridded ERA5 climatology
        CESM2_to_ERA5_varnames = {'TREFHT':'t2m','FSDS':'msdwswrf','FLDS':'msdwlwrf','QREFHT':'sph2m', 'PRECS':'msr', 'PRECT':'mtpr', \
                'PSL':'msl', 'wind_speed':'wind_speed', 'wind_angle':'wind_angle', 'UBOT':'u10', 'VBOT':'v10'}
        filevarname = CESM2_to_ERA5_varnames[variable]
        if variable=='QREFHT':
           varname='specific_humidity' # file with dewpoint temperature converted to specific humidity
        else:
           varname=filevarname
        ERA5_climatology = xr.open_dataset(f'{era5_folder}ERA5_eANT025_{filevarname}_{freq}_{year_start}-{year_end}_mean_monthly.nc').rename({varname:variable})
 
        # thermo(dynamic) correction
        if variable in ['TREFHT','QREFHT','FLDS','FSDS','PRECS','PRECT','wind_speed','wind_angle','PSL','UBOT','VBOT']:
            if monthly:
                out_file = f'{out_folder}{source}-{expt}_{variable}_{freq}_bias_corr_monthly.nc'
            else:
                CESM2_ensemble_mean = CESM2_ensemble_mean.mean(dim='month')
                ERA5_climatology    = ERA5_climatology.mean(dim='month')
                out_file = f'{out_folder}{source}-{expt}_{variable}_{freq}_bias_corr.nc'
            
            # Call (thermo)dynamic correction:
            thermo_dynamic_correction(CESM2_ensemble_mean, ERA5_climatology, variable, out_file, fill_land=fill_land, monthly=monthly)
        else:
            raise Exception(f'Variable {variable} does not need bias correction. Is this true?')
    else:
        raise Exception("Bias correction currently only set up to correct CESM2. Sorry you'll need to write some more code!")

    return

# Function to calculate the bias and associated (thermo)dynamic correction between source (CESM2) and ERA5 mean fields
# Inputs:
# - source_mean : xarray Dataset containing the ensemble and time mean of the source dataset variable (currently CESM2)
# - ERA5_mean   : xarray Dataset containing the time mean of the ERA5 variable
# - out_file    : string to path to write NetCDF file to
# - (optional) fill_land  : boolean indicating whether to fill areas that are land in the source mask with nearest values or whether to just leave it as is
# - (optional) monthly    : boolean indicating whether bias correction fields are monthly or not
# - (optional) dist_coast : 
# - (optional) coastal_correction_limit : float indicating the distance from the coastline (in km) considered for the coastal wind angle correction
def thermo_dynamic_correction(source_mean, ERA5_mean, variable, out_file, fill_land=True, monthly=False, coastal=False, 
                              dist_coast='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/bathymetry/distance_coast-20250715.nc', coastal_correction_limit=400):
    if variable in ['wind_speed','UBOT','VBOT']:
        print('Correcting dynamics')
        bias = ERA5_mean / source_mean
    elif variable=='wind_angle':
        print('Correcting dynamics') 
        
        angle_bias = ERA5_mean - source_mean
        
        if coastal: # apply angle correction only within a certain distance of the coastline
            # open file with distance from coastline for the configuration
            distcoast = xr.open_dataset(dist_coast)
            # apply correction to distance within coastal_correction_limit (excluding south america)
            region_to_correct = (distcoast.distance_coast < coastal_correction_limit)*(distcoast.nav_lat <-59)
            angle_bias = xr.where(region_to_correct, angle_bias, 0)
            angle_bias = xr.where(distcoast.distance_coast ==0, 0, angle_bias) # mask land
            # apply a cosine taper to the angle bias to gradually transition from strength 1 to 0
            bias = xr.where(region_to_correct, angle_bias*np.cos((np.pi/2)*(distcoast.distance_coast/coastal_correction_limit)), 0)
        else:
            bias = angle_bias
    else:
        print('Correcting thermodynamics')
        # Calculate difference:
        bias = ERA5_mean - source_mean
    # Fill land regions along latitudes
    if fill_land:
       # bias = bias.interpolate_na(dim='lat', method='nearest', fill_value="extrapolate")
       src_to_fill = xr.where(np.isnan(bias), -9999, bias) # which cells need to be filled
       var_filled  = extend_into_mask(src_to_fill[variable].values, missing_val=-9999, fill_val=np.nan, use_2d=True, use_3d=False, num_iters=100)
       if monthly:
           bias[variable] = (('time','lat','lon'), var_filled)
       else:
           bias[variable] = (('lat','lon'), var_filled)

    # write to file
    if monthly:
        bias.to_netcdf(out_file, unlimited_dims='month')
    else:
        bias.to_netcdf(out_file)
    
    return

# Main function to apply bias correction files produced from calc_bias_correction to the source files
# Inputs:
# - variable : string of variable name
# - ens      : ensemble member to bias correct
# - (optional) expt       : string specifying CESM2 experiment to correct
# - (optional) start_year : integer start year for files to process
# - (optional) end_year   : integer end year for files to process
# - (optional) monthly    : boolean specifying whether to do monthly bias correction or otherwise a constant spatial field
# - (optional) freq       : string specifying underlying forcing frequency from which bias correction was derived (hourly, 3-hourly, or hourly)
# - (optional) highres    : boolean whether the files to be read in and bias corrected are on the eANT025 (high resolution) grid
# - (optional) out_dir    : string path where to save bias corrected forcing files
# - (optional) bias_dir   : string path to bias correction field files produced by the function calc_bias_correction
def apply_bias_correction(variable, ens, expt='LE2', start_year=1900, end_year=2050, monthly=True, freq='daily', highres=True,
                          out_dir='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/climate-forcing/CESM2/LE2/bias-corrected/',
                          bias_dir='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/climate-forcing/CESM2/LE2/ensemble_mean/bias_corr/'):
    from tqdm import tqdm
    print(f'Processing {variable} files for ensemble member {ens} from {start_year}-{end_year} with bias correction file from {freq} means')

    time_coder = xr.coders.CFDatetimeCoder(use_cftime=True)

    out_dir = f'{out_dir}ens{ens}/'
    bias_ds = xr.open_dataset(f'{bias_dir}CESM2-LE2_{variable}_{freq}_bias_corr_monthly.nc')[variable]
    if variable=='wind_speed': # also load wind angle bias correction file
        bias_ds_angle = xr.open_dataset(f'{bias_dir}CESM2-LE2_wind_angle_{freq}_bias_corr_monthly.nc').wind_angle

    # helper functions for applying bias correction
    def apply_multiplier(angle):
        speed = angle * bias_ds.sel(month=(angle.time_counter.dt.month[0].values))
        return speed
    def add_bias(ds_var):
        ds_var_corrected = ds_var + bias_ds.sel(month=(ds_var.time_counter.dt.month[0].values))
        return ds_var_corrected
    def add_angle(ds_var):
        ds_var_corrected = ds_var + bias_ds_angle.sel(month=(ds_var.time_counter.dt.month[0].values))
        return ds_var_corrected

    # loop over each year to apply bias correction
    for year in tqdm(range(start_year, end_year+1)):

        # for the u and v wind velocities, bias correction is based on a wind speed multiplier
        if variable=='wind_speed':
            file_pathx = find_processed_cesm2_file(expt, 'UBOT', ens, year, freq='3-hourly', highres=highres)
            file_pathy = find_processed_cesm2_file(expt, 'VBOT', ens, year, freq='3-hourly', highres=highres)
            if isinstance(file_pathx, str):
                file_pathx = [file_pathx]; file_pathy = [file_pathy]

            for filex, filey in zip(file_pathx,file_pathy):
                dsx = xr.open_dataset(filex, decode_times=time_coder)
                dsy = xr.open_dataset(filey, decode_times=time_coder)
                if monthly:
                    dsx['UBOT'] = dsx.UBOT.groupby('time_counter.month').apply(apply_multiplier)
                    dsy['VBOT'] = dsy.VBOT.groupby('time_counter.month').apply(apply_multiplier)
                else:
                    dsx['UBOT'] = dsx.UBOT * bias_ds.mean(dim='month')
                    dsy['VBOT'] = dsy.VBOT * bias_ds.mean(dim='month')
              
                if np.unique(dsx.time_counter.dt.month.values).size==1:
                    month = dsx.time_counter.dt.month.isel(time_counter=0).values
                    dsx.to_netcdf(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_UBOT_bias_corr_{'monthly_' if monthly else ''}2D_y{year}m{month:02}.nc", \
                            unlimited_dims=['time_counter'])
                    dsy.to_netcdf(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_VBOT_bias_corr_{'monthly_' if monthly else ''}2D_y{year}m{month:02}.nc", \
                            unlimited_dims=['time_counter'])
                else:
                    dsx.to_netcdf(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_UBOT_bias_corr_{'monthly_' if monthly else ''}2D_y{year}.nc", 
                            unlimited_dims=['time_counter'])
                    dsy.to_netcdf(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_VBOT_bias_corr_{'monthly_' if monthly else ''}2D_y{year}.nc", \
                            unlimited_dims=['time_counter'])

        # every other variable has an addition spatial bias field correction
        else:
            file_path = find_processed_cesm2_file(expt, variable, ens, year, freq=freq, highres=highres)
            if isinstance(file_path,str):
                file_path = [file_path]

            for file in file_path:
                ds = xr.open_dataset(file, decode_times=time_coder)             
                if not highres:
                    ds = ds.rename({'time':'time_counter'})

                if variable in ['UBOT','VBOT']:
                    if monthly:
                        ds[variable] = ds[variable].groupby(f'time_counter.month').apply(apply_multiplier)
                    else:
                        ds[variable] = ds[variable] * bias_ds.mean(dim='month')
                else:
                    if monthly:
                        ds[variable] = ds[variable].groupby(f'time_counter.month').apply(add_bias)
                    else:
                        ds[variable] += bias_corr[variable].mean(dim='month')

                if variable=='PRECS':
                    # prevent the snowfall rate from exceeding the total precipitation rate
                    # read in bias-corrected PRECT file
                    try:
                        ds_PRECT = xr.open_dataset(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_PRECT_bias_corr_{'monthly_' if monthly else ''}y{year}.nc", decode_times=time_coder)
                    except:
                        raise Exception('Need to bias correct PRECT before PRECS for consistency of precipitation rates')
                
                    # set the snowfall rate equal to the total precipitation rate if it is higher than it, otherwise leave as is
                    ds['PRECS'] = xr.where(ds['PRECS'] > ds_PRECT['PRECT'], ds_PRECT['PRECT'], ds['PRECS'])

                if variable in ['PRECS', 'PRECT', 'FLDS', 'FSDS', 'QREFHT']:                    
                    ds[variable] = xr.where(ds[variable] < 0, 0, ds[variable]) # should not be negative as a result of bias correction

                if np.unique(ds.time_counter.dt.month.values).size==1:
                    month = ds.time_counter.dt.month.isel(time_counter=0).values
                    ds.to_netcdf(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_{variable}_bias_corr_{'monthly_' if monthly else ''}y{year}m{month:02}.nc", \
                            unlimited_dims=['time_counter'])
                else:
                    ds.to_netcdf(f"{out_dir}CESM2-{expt}_ens{ens}_{'eANT025_' if highres else ''}{freq}_{variable}_bias_corr_{'monthly_' if monthly else ''}y{year}.nc", \
                            unlimited_dims=['time_counter'])

    return

# Function to pre-process ERA5 atmospheric forcing datasets
# Inputs:
# - variable : string of ERA5 variable name
# - (optional) year_start  : start year of range to process
# - (optional) year_end    : end year
# - (optinoal) era5_folder : path to ERA5 forcing files
def process_era5_forcing(variable, year_start=1979, year_end=2024, era5_folder='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/ERA5-forcing/daily/'):

    for year in range(year_start,year_end+1):
        if variable=='d2m':
            print('Convert dewpoint temperature')
            # convert dewpoint temperature to specific humidity and write to file (for bias correction calc)
            dewpoint_to_specific_humidity(file_dew=f'd2m_y{year}.nc', variable_dew='d2m',
                                          file_slp=f'msl_y{year}.nc', variable_slp='msl',
                                          dataset='ERA5', folder=f'{era5_folder}raw/')

        # fill land with nearest connected point (assume land mask is more or less constant over the reanalysis period)
        landmask = xr.open_dataset(f'{era5_folder}../climatology/land_sea_mask.nc').isel(time=0).lsm.rename({'longitude':'lon','latitude':'lat'})
        landmask['lon'] = fix_lon_range(landmask['lon'])
        landmask = landmask.sortby('lon')

        # convert time dimension to unlimited so that NEMO reads in the calendar correctly
        for filename in glob.glob(f'{era5_folder}raw/{variable}*y{year}.nc'):
            with xr.open_dataset(filename, mode='a') as data:
                print('Processing', filename)
                try:
                    data = data.rename({'valid_time':'time'})
                except:
                    pass

                # tidy up the longitude range and naming conventions
                data        = data.rename({'longitude':'lon', 'latitude':'lat'})
                data['lon'] = fix_lon_range(data['lon'])
                data        = data.sortby('lon')
                name_remap  = {'msdwlwrf':'avg_sdlwrf', 'msdwswrf':'avg_sdswrf', 'mtpr':'avg_tprate', 'msr':'avg_tsrwe'}
                try:
                    data = data.rename({f'{name_remap[variable]}':variable})
                except:
                    pass
                
                variable = filename.split('raw/')[1].split('_')[0]
                if variable in ['msdwlwrf','msdwswrf','t2m','sph2m','d2m','msl','msr','mtpr','t2m','u10','v10']: 
                    print(f'Filling land for variable {variable} year {year}')
                    if variable=='sph2m':
                        varname='specific_humidity'
                    else:
                        varname=variable
                    src_to_fill = xr.where(landmask!=0, -9999, data[varname]) # which cells need to be filled
                    var_filled_array = np.empty(src_to_fill.shape)
                    for tind, t in enumerate(src_to_fill.time):
                        var_filled_array[:,:,tind] = extend_into_mask(src_to_fill.isel(time=tind).values, missing_val=-9999, fill_val=np.nan, 
                                                                      use_2d=True, use_3d=False, num_iters=200)
                    data[varname] = (('lat','lon','time'), var_filled_array)
                    data = data.transpose('time','lat','lon')

                data.to_netcdf(f'{era5_folder}processed/{variable}_y{year}.nc', unlimited_dims={'time':True})

    return


# As above but only convert specific humidity, don't do the flood-filling; and do so looping over 24 chunks of the file to save memory.
def convert_era5_specific_humidity (year_start=1979, year_end=2024, era5_folder='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ERA5_hourly/'):

    for year in range(year_start, year_end+1):
        print('Processing '+str(year))
        ds_dew = xr.open_dataset(era5_folder+'/d2m_y'+str(year)+'.nc')
        ds_slp = xr.open_dataset(era5_folder+'/msl_y'+str(year)+'.nc')
        ds_sph = None
        num_time = ds_dew.sizes['valid_time']
        chunk = num_time//hours_per_day
        for t in range(0, num_time, chunk):
            ds_tmp = dewpoint_to_specific_humidity(ds_dew=ds_dew.isel(valid_time=slice(t,t+chunk)), ds_slp=ds_slp.isel(valid_time=slice(t,t+chunk)))
            if ds_sph is None:
                ds_sph = ds_tmp
            else:
                ds_sph = xr.concat([ds_sph, ds_tmp], dim='valid_time')
        ds_sph.to_netcdf(era5_folder+'/sph2m_y'+str(year)+'.nc')
            
        


# Convert one suite of UKESM forcing (3-hourly, pp files, 9 variables) to NEMO forcing files.
def ukesm_atm_forcing_3h (suite, in_dir=None, out_dir='./', lat_max=-50, flood_fill=True, mask_file='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/masks_um_ukesm1.2.nc'):

    import iris
    import warnings

    var_names = ['air_temperature', 'specific_humidity', 'air_pressure_at_sea_level', 'x_wind', 'y_wind', 'precipitation_flux', 'snowfall_flux', 'surface_downwelling_shortwave_flux_in_air', 'surface_downwelling_longwave_flux_in_air']
    var_names_snapshot = ['air_temperature', 'specific_humidity', 'air_pressure_at_sea_level']  # Variables which are only available as 3 hour snapshots, not time-means
    # Output files should have shorter variable names, as otherwise some of them overflow in NEMO
    var_name_remapping = {'air_temperature' : 'tair',
                          'specific_humidity' : 'qair',
                          'air_pressure_at_sea_level' : 'pair',
                          'x_wind' : 'uwind',
                          'y_wind' : 'vwind',
                          'precipitation_flux' : 'precip',
                          'snowfall_flux' : 'snow',
                          'surface_downwelling_shortwave_flux_in_air' : 'swrad',
                          'surface_downwelling_longwave_flux_in_air' : 'lwrad'} 
    lat_buffer = 1
    if flood_fill:
        ds_masks = xr.open_dataset(mask_file)
        missing_val = -9999

    if in_dir is None:
        in_dir = './'+suite+'/'
    file_head = suite+'a.p8'
    file_tail = '.pp'

    # Find all matching files in directory
    file_names = []
    for f in os.listdir(in_dir):
        if f.startswith(file_head) and f.endswith(file_tail):
            file_names.append(f)
    file_names.sort()
    # Find range of years to process
    # Inner function to extract year and month from filename
    def get_date_stamp (f):
        date = f[len(file_head):-len(file_tail)]
        year = int(date[:4])
        month = int(date[4:6])
        return year, month
    # First loop forwards until we find a January
    for f in file_names:
        year, month = get_date_stamp(f)
        if month == 1:
            start_year = year
            break
    # Then loop backwards until we find a December
    for f in file_names[::-1]:
        year, month = get_date_stamp(f)
        if month == months_per_year:
            end_year = year
            break

    # Inner function to use later, to check if a cube is a time-mean
    def check_mean (cube):
        for cm in cube.cell_methods:
            if cm.method=='mean' and 'time' in cm.coord_names:
                return True
        return False
    # Inner function to add variable to given dataset
    def add_var (ds, data, var):
        if ds is None:
            ds = xr.Dataset({var:data})
        else:
            ds = ds.assign({var:data})
        return ds
    # Inner function to concatenate dataset to master dataset
    def concat_ds (ds_master, ds):
        if ds_master is None:
            if 'time' not in ds.dims:
                ds = ds.expand_dims(dim='time')
            return ds
        else:
            return xr.concat([ds_master, ds], dim='time')

    # Loop over years and months
    # Will have two rolling datasets to keep track of time-mean and time-snapshot variables
    ds_mean = None
    ds_snapshot = None
    index = 0
    for year in range(start_year, end_year+1):
        for month in range(1, months_per_year+1):
            # Read files until we've covered this month
            while True:
                fname = file_names[index]
                # Check year on this file and the next file
                this_year = get_date_stamp(fname)[0]
                if index == len(file_names)-1:
                    next_year = None
                else:
                    next_year = get_date_stamp(file_names[index+1])[0]
                # Prepare for next iteration of loop
                index += 1
                if next_year is not None and next_year < start_year:
                    # Too early; skip this file
                    continue
                # Check if this is the last file before start_year
                is_last_file_before = this_year < start_year and next_year == start_year
                print('Reading '+in_dir+fname)
                with warnings.catch_warnings():
                    warnings.simplefilter('ignore')                    
                    cubes = iris.load(in_dir+fname)
                ds_mean_tmp = None
                ds_snapshot_tmp = None
                # Loop over variables
                for var in var_names:
                    if is_last_file_before and var not in var_names_snapshot:
                        # Only need to read this file for snapshot variables
                        continue
                    # Find how many cubes match this variable name
                    matches = [cube.standard_name==var for cube in cubes]
                    num_matches = np.count_nonzero(matches)
                    if num_matches == 0:
                        raise Exception('Missing variable '+var+' from file '+fname)
                    elif num_matches == 2:
                        # Two options; expect one is time-mean and one is snapshot
                        is_mean = [check_mean(cube) for cube in cubes]
                        match_mean = np.argwhere(np.array(is_mean)*np.array(matches))[0]
                        if len(match_mean) != 1:
                            raise Exception('Found '+str(len(match_mean))+' options for time-mean for variable '+var)
                        cube = cubes[match_mean[0]]
                        # Will add to time-mean dataset
                        method = 'mean'
                    elif num_matches == 1:
                        cube = cubes[matches.index(True)]
                        # Add to either time-mean or snapshot dataset
                        if check_mean(cube):
                            method = 'mean'
                        else:
                            method = 'snapshot'
                    else:
                        raise Exception('Unexpected number of matches ('+str(num_matches)+') for variable '+var+' in file '+fname)
                    # Now convert to xarray
                    with warnings.catch_warnings():
                        warnings.simplefilter('ignore')
                        data = xr.DataArray.from_iris(cube)
                    if flood_fill:
                        # Apply land mask
                        if var == 'x_wind':
                            mask_var = 'aum3'
                        elif var == 'y_wind':
                            mask_var = 'avm3'
                        else:
                            mask_var = 'atm3'
                        mask = ds_masks[mask_var+'.msk'].rename({'y_'+mask_var:'latitude', 'x_'+mask_var:'longitude'})
                        data = xr.where(mask==0, data, missing_val)
                    # Trim latitude, with a 1-degree buffer
                    data = data.where(data.latitude < lat_max + 1, drop=True)
                    if 'time' in data.dims:
                        data = data.transpose('time','latitude','longitude')
                    else:
                        data = data.expand_dims(dim='time')
                    # Make sure longitude is in the range -180 to 180
                    data['longitude'] = fix_lon_range(data['longitude'])
                    data.load()
                    if flood_fill:
                        # Fill land mask with nearest neighbours
                        data_filled = np.empty(data.shape)
                        for t in range(data.sizes['time']):
                            data_filled[t,:] = extend_into_mask(data.isel(time=t).data, missing_val=missing_val, use_2d=True, num_iters=data.sizes['latitude'])
                        data.data = data_filled
                    if var == 'x_wind':
                        data = data.rename({'longitude':'longitude_u', 'latitude':'latitude_u'})
                    elif var == 'y_wind':
                        data = data.rename({'longitude':'longitude_v', 'latitude':'latitude_v'})
                    # Add to the correct dataset
                    if method == 'mean':
                        ds_mean_tmp = add_var(ds_mean_tmp, data, var)
                    elif method == 'snapshot':
                        ds_snapshot_tmp = add_var(ds_snapshot_tmp, data, var)
                # After all variables, concatenate to master datasets
                if ds_mean_tmp is not None:
                    ds_mean = concat_ds(ds_mean, ds_mean_tmp)
                ds_snapshot = concat_ds(ds_snapshot, ds_snapshot_tmp)
                # Now check if we have covered this month yet: will need at least one snapshot into the next month
                end_time = ds_snapshot.time[-1]
                if (end_time.dt.year > year) or (end_time.dt.year == year and end_time.dt.month > month):
                    break
            # Select data for this month and remove it from master array
            # Time-mean array: simple
            time_range = (ds_mean.time.dt.year==year)*(ds_mean.time.dt.month==month)
            ds_mean_month = ds_mean.where(time_range, drop=True)
            ds_mean = ds_mean.where(~time_range, drop=True)
            # Snapshot array: first interpolate to time-mean axis (equivalent to midpoints of snapshots)
            ds_snapshot_mean_month = ds_snapshot.interp(time=ds_mean_month.time, method='linear')
            # Trim the snapshots to only keep the ones after the last time-mean
            ds_snapshot = ds_snapshot.where(ds_snapshot.time > ds_mean_month.time[-1], drop=True)
            # Drop unused coordinates which may confuse NEMO
            ds_mean_month = ds_mean_month.drop_vars({'forecast_reference_time', 'forecast_period'})
            ds_snapshot_mean_month = ds_snapshot_mean_month.drop_vars({'forecast_reference_time', 'forecast_period'})
            # Write each variable to a file with the correct naming convention
            for var in var_names:
                var_new = var_name_remapping[var]
                out_file = out_dir+'/'+var_new+'_y'+str(year)+'m'+str(month).zfill(2)+'.nc'
                print('Writing '+out_file)
                if var in ds_mean_month:
                    data = ds_mean_month[var]
                else:
                    data = ds_snapshot_mean_month[var]
                data = data.rename(var_new)
                if data.sizes['time'] != 30*hours_per_day/3:
                    raise Exception('Invalid length of time axis ('+str(data.sizes['time'])+' for file '+out_file)
                data.to_netcdf(out_file, unlimited_dims='time')


# Calculate bias correction files on the UM grid, for testing online bias correction code (6 thermodynamic variables only) with no time dependence (annual mean). Don't worry about flood filling or proper interpolation against periodic boundaries; this is just a test of online vs offline.
def ukesm_bias_corrections_test (ukesm_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ensemble_mean_climatology/', era5_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ERA5_hourly/climatology/', out_dir='./'):

    ukesm_var_names = ['tair', 'qair', 'precip', 'snow', 'swrad', 'lwrad']
    era5_var_names = ['t2m', 'sph2m', 'mtpr', 'msr', 'msdwswrf', 'msdwlwrf']
    ukesm_tail = '_1979-2014_mean_monthly.nc'
    era5_head = 'ERA5_'
    era5_tail = '_3-hourly_1979-2014_mean_monthly.nc'

    for var_u, var_e in zip(ukesm_var_names, era5_var_names):
        ds_ukesm = xr.open_dataset(ukesm_dir+'/'+var_u+ukesm_tail).mean(dim='month')
        ds_era5 = xr.open_dataset(era5_dir+'/'+era5_head+var_e+era5_tail).rename({var_e:var_u}).mean(dim='month')
        ds_era5.coords['longitude'] = fix_lon_range(ds_era5.coords['longitude'], max_lon=180)
        # Regrid ERA5 to UM grid
        ds_era5_interp = ds_era5.interp_like(ds_ukesm)
        ds_correction = ds_era5_interp - ds_ukesm
        # Fill missing row at north with zeros
        ds_correction = ds_correction.fillna(0)
        out_file = out_dir+'/'+var_u+'_bias_correction.nc'
        print('Writing '+out_file)
        ds_correction.to_netcdf(out_file)


# Apply the bias corrections calculated above to a short period of UM forcing, using annually averaged fields (no time dependence).
def ukesm_apply_bias_corrections_test (forcing_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/nemo_forcing_files/cy691/', out_dir=None, bias_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/bias_corrections/test/', start_year=1978, end_year=1979):

    var_names = ['tair', 'qair', 'precip', 'snow', 'swrad', 'lwrad']  # Only the ones to correct
    file_tail = '_bias_correction.nc'
    if out_dir is None:
        out_dir = forcing_dir + '/bias_corrected/'

    # Read bias corrections and merge into one dataset
    ds_corr = None
    for var in var_names:
        ds = xr.open_dataset(bias_dir+'/'+var+file_tail).drop_vars({'height'})
        if ds_corr is None:
            ds_corr = ds
        else:
            ds_corr = ds_corr.merge(ds)

    # Loop over files in the forcing directory
    for f in os.listdir(forcing_dir):
        # Skip anything that's not a NetCDF file; this includes directories
        if not f.endswith('.nc'):
            continue
        # Extract year
        i0 = f.index('_y')+2
        year = int(f[i0:i0+4])
        # Check if in date range
        if year < start_year or year > end_year:            
            continue
        file_path_in = forcing_dir+'/'+f
        file_path_out = out_dir+'/'+f
        print('Processing '+file_path_in)
        # Check if this variable needs correcting
        if any([var in f for var in var_names]):
            # Extract variable
            var = f[:f.index('_y')]
            # Read the data
            ds = xr.open_dataset(file_path_in)
            # Apply bias correction; minimum of 0 will happen within NEMO code per timestep
            ds[var] = ds[var] + ds_corr[var]
            # Save file
            ds.to_netcdf(file_path_out)
        else:
            # Variable doesn't need correcting; just copy
            shutil.copyfile(file_path_in, file_path_out)


# Make proper bias correction files: monthly, on the ERA5 grid, flood filled, proper interpolation with cf. These will be used to correct biases online.
def ukesm_bias_corrections_thermo (ukesm_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ensemble_mean_climatology/', era5_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ERA5_hourly/climatology/', out_dir='./', era5_mask_file='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/ERA5-forcing/climatology/land_sea_mask.nc'):

    ukesm_var_names = ['tair', 'qair', 'precip', 'snow', 'swrad', 'lwrad']
    era5_var_names = ['t2m', 'sph2m', 'mtpr', 'msr', 'msdwswrf', 'msdwlwrf']
    ukesm_tail = '_1979-2014_mean_monthly.nc'
    era5_head = 'ERA5_'
    era5_tail = '_3-hourly_1979-2014_mean_monthly.nc'
    missing_val = -9999

    # Read ERA5 mask and trim to existing climatology file
    ds_mask = xr.open_dataset(era5_mask_file).isel(time=0).drop_vars({'time'})
    # Flip order of latitude to agree with ERA5 climatology files
    ds_mask = ds_mask.reindex(latitude=ds_mask['latitude'][::-1])
    mask_era5 = ds_mask['lsm']  # Open ocean is zero   

    for var_u, var_e in zip(ukesm_var_names, era5_var_names):
        # Read UKESM climatology; land already flood filled by ukesm_atm_forcing_3h
        ds_ukesm = xr.open_dataset(ukesm_dir+'/'+var_u+ukesm_tail).transpose('month', 'latitude', 'longitude')
        # Read ERA5
        ds_era5 = xr.open_dataset(era5_dir+'/'+era5_head+var_e+era5_tail).rename({var_e:var_u}).transpose('month', 'latitude', 'longitude')
        # Regrid UM to the ERA5 grid
        ds_ukesm_interp = interp_latlon_cf(ds_ukesm, ds_era5, source_type='other', target_type='other', pster_src=False, pster_target=False, periodic_src=True, periodic_target=True, method='linear', time_dim='month')
        # Calculate bias correction
        data_correction = ds_era5[var_u] - ds_ukesm_interp[var_u]
        # Flood fill ERA5 land
        # First trim mask if needed on first variable
        if mask_era5.sizes['latitude'] > ds_era5.sizes['latitude']:
            mask_era5 = mask_era5.isel(latitude=slice(0,ds_era5.sizes['latitude']))
        data_tmp = xr.where(mask_era5==0, data_correction, missing_val).transpose('month', 'latitude', 'longitude')
        data_filled = np.empty(data_tmp.shape)
        for t in range(data_tmp.sizes['month']):
            data_filled[t,:] = extend_into_mask(data_tmp.isel(month=t).data, missing_val=missing_val, use_2d=True, num_iters=data_tmp.sizes['latitude'])
        data_correction.data = data_filled
        # Fill northern part with zeros (outside of domain)
        if ds_mask.sizes['latitude'] > ds_era5.sizes['latitude']:
            data_buffer = 0*ds_mask['lsm'].isel(latitude=slice(ds_era5.sizes['latitude'],None))
            data_correction = xr.concat([data_correction, data_buffer], dim='latitude')
        # Swap order of latitude to agree with ERA5 forcing files
        data_correction = data_correction.reindex(latitude=data_correction['latitude'][::-1])
        out_file = out_dir+'/'+var_u+'_bias_correction.nc'
        print('Writing '+out_file)
        data_correction.to_netcdf(out_file)


# As above, but for coastal winds in polar coordinates.
# Will correct over <taper_dist> km from the coast, scaling the speed by a factor not exceeding <scale_cap>, and rotating the angle.
def ukesm_bias_corrections_winds (ukesm_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ensemble_mean_climatology/', era5_dir='/gws/ssde/j25b/terrafirma/kaight/NEMO_AIS/UKESM_forcing/ERA5_hourly/climatology/', out_dir='./', era5_mask_file='/gws/ssde/j25b/anthrofail/birgal/NEMO_AIS/ERA5-forcing/climatology/land_sea_mask.nc', taper_dist=150, scale_cap=3):

    var_names = ['wind_speed', 'wind_angle']
    ukesm_tail = '_1979-2014_mean_monthly.nc'
    era5_head = 'ERA5_'
    era5_tail = '_3-hourly_1979-2014_mean_monthly.nc'
    missing_val = -9999

    print('Reading ERA5 masks')
    # Read ERA5 mask and trim to existing climatology file
    ds_mask = xr.open_dataset(era5_mask_file).isel(time=0).drop_vars({'time'})
    # Flip order of latitude to agree with ERA5 climatology files
    ds_mask = ds_mask.reindex(latitude=ds_mask['latitude'][::-1])
    mask_era5 = ds_mask['lsm']  # Open ocean is zero
    taper = None

    for var in var_names:
        print('Correcting '+var)
        
        # Read both datasets
        ds_ukesm = xr.open_dataset(ukesm_dir+'/'+var+ukesm_tail).transpose('month', 'latitude', 'longitude')
        ds_era5 = xr.open_dataset(era5_dir+'/'+era5_head+var+era5_tail).transpose('month', 'latitude', 'longitude')
        # Regrid UM to the ERA5 grid
        ds_ukesm_interp = interp_latlon_cf(ds_ukesm, ds_era5, source_type='other', target_type='other', pster_src=False, pster_target=False, periodic_src=True, periodic_target=True, method='linear', time_dim='month')
        
        # Trim ERA5 mask if needed
        if mask_era5.sizes['latitude'] > ds_era5.sizes['latitude']:
            mask_era5 = mask_era5.isel(latitude=slice(0,ds_era5.sizes['latitude']))
        if taper is None:
            print('Calculating distance from the coast')
            # Remove islands from ERA5 mask
            point0 = closest_point(mask_era5, land_ice_point0)
            mask_era5.data = remove_disconnected(mask_era5, point0)
            # Find distance to coast (boundary of ERA5 mask)
            dist_to_coast = distance_to_bdry(mask_era5['longitude'], mask_era5['latitude'], mask_era5)
            # Now fill land mask with zeros
            dist_to_coast = xr.where(mask_era5, 0, dist_to_coast)
            # Set up tapering function
            taper = (dist_to_coast < taper_dist)*np.cos(np.pi/2*dist_to_coast/taper_dist)            
            
        if var == 'wind_speed':
            # Take the ratio of ERA5 to UKESM winds, and apply scale_cap if needed
            data_correction = np.minimum(ds_era5[var]/ds_ukesm_interp[var], scale_cap)
            # Taper towards 1
            taper0 = 1
            
        elif var == 'wind_angle':
            # Take difference in angles
            rotate = ds_era5[var] - ds_ukesm_interp[var]
            # Take mod 2pi when necessary
            rotate = xr.where(rotate < -np.pi, rotate+2*np.pi, rotate)
            rotate = xr.where(rotate > 2*np.pi, rotate-2*np.pi, rotate)
            data_correction = rotate
            # Taper towards 0
            taper0 = 0

        # Now apply the tapering function
        data_correction = (data_correction - taper0)*taper + taper0
        # Fill any missing values (stripe at top) with taper0
        data_correction = xr.where(data_correction.isnull(), taper0, data_correction)

        # Flood fill ERA5 land
        data_tmp = xr.where(mask_era5==0, data_correction, missing_val).transpose('month', 'latitude', 'longitude')
        data_filled = np.empty(data_tmp.shape)
        for t in range(data_tmp.sizes['month']):
            data_filled[t,:] = extend_into_mask(data_tmp.isel(month=t).data, missing_val=missing_val, use_2d=True, num_iters=data_tmp.sizes['latitude'])
        data_correction.data = data_filled

        # Fill northern part with zeros or ones (outside of domain)
        if ds_mask.sizes['latitude'] > ds_era5.sizes['latitude']:
            data_buffer = 0*ds_mask['lsm'].isel(latitude=slice(ds_era5.sizes['latitude'],None)) + taper0
            data_correction = xr.concat([data_correction, data_buffer], dim='latitude')
        # Swap order of latitude to agree with ERA5 forcing files
        data_correction = data_correction.reindex(latitude=data_correction['latitude'][::-1])
        
        data_correction.name = var        
        out_file = out_dir+'/'+var+'_bias_correction.nc'
        print('Writing '+out_file)
        data_correction.to_netcdf(out_file)