diff --git a/.gitignore b/.gitignore
index f42eb1e..10f4f8a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -17,4 +17,3 @@ test_stats_csv.csv
 test_stats_csv.dat
 tmp_tolerance_None.csv
 tolerance.csv
-util/differences.csv
diff --git a/tests/util/test_fof_utils.py b/tests/util/test_fof_utils.py
index 9564311..7794869 100644
--- a/tests/util/test_fof_utils.py
+++ b/tests/util/test_fof_utils.py
@@ -11,9 +11,9 @@
     clean_value,
     compare_arrays,
     compare_var_and_attr_ds,
-    fill_nans_for_float32,
     get_observation_variables,
     get_report_variables,
+    replace_nan_with_sentinel_float64,
     split_feedback_dataset,
 )
 from util.log_handler import initialize_detailed_logger
@@ -194,10 +194,10 @@ def fixture_arr():
 def test_fill_nans_for_float32_nan(arr_nan):
     """
     Test that if an array containing nan is given, these values are replaced
-    by -9.99999e05.
+    by -999999.0.
     """
-    array = fill_nans_for_float32(arr_nan)
-    expected = np.array([1.0, -9.99999e05, 3.0, 4.0, -9.99999e05], dtype=np.float32)
+    array = replace_nan_with_sentinel_float64(arr_nan)
+    expected = np.array([1.0, -999999.0, 3.0, 4.0, -999999.0], dtype=np.float64)
     assert np.array_equal(array, expected)
 
 
@@ -206,7 +206,7 @@ def test_fill_nans_for_float32(arr1):
     Test that if an array without nan is given, the output of the function
     is the same as the input.
     """
-    array = fill_nans_for_float32(arr1)
+    array = replace_nan_with_sentinel_float64(arr1)
     assert np.array_equal(array, arr1)
 
 
diff --git a/util/fof_utils.py b/util/fof_utils.py
index 687d23d..de7eae9 100644
--- a/util/fof_utils.py
+++ b/util/fof_utils.py
@@ -104,13 +104,18 @@ def compare_arrays(arr1, arr2, var_name):
     return total, equal, diff
 
 
-def fill_nans_for_float32(arr):
+def replace_nan_with_sentinel_float64(arr):
     """
-    To make sure nan values are recognised.
+    If the input array has a floating dtype, it is cast to float64
+    and all NaN values are replaced with the sentinel value -999999.0
+    If the array does not have a floating dtype, it is returned unchanged.
     """
-    if arr.dtype == np.float32 and np.isnan(arr).any():
-        return np.where(np.isnan(arr), -999999, arr)
-    return arr
+    if not np.issubdtype(arr.dtype, np.floating):
+        return arr
+
+    arr = arr.astype(np.float64, copy=False)
+
+    return np.where(np.isnan(arr), -999999.0, arr)
 
 
 def clean_value(x):
@@ -203,8 +208,8 @@ def process_var(ds1, ds2, var, detailed_logger):
     number of matching elements.
     """
 
-    arr1 = fill_nans_for_float32(ds1[var].values)
-    arr2 = fill_nans_for_float32(ds2[var].values)
+    arr1 = replace_nan_with_sentinel_float64(ds1[var].values)
+    arr2 = replace_nan_with_sentinel_float64(ds2[var].values)
     if arr1.size == arr2.size:
         t, e, diff = compare_arrays(arr1, arr2, var)
         if diff.size != 0: