diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index 46a5197..850c042 100644
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -23,6 +23,7 @@ The changes listed in this file are categorised as follows:
 
 ### Fixed
 
+- ``Cmip6MeteorDataGetter.validate_pattern_scaling_cache`` now accepts an optional ``variable`` argument matching the per-variable suffix ``MeteorPatternScaling`` saves under (``cmip6-{model}-aer-{variable}``). Without it, a cache that ``MeteorPatternScaling.__init__`` loads successfully by filename was reported as "invalid" by the outer validator, causing ``MeteorInterface`` to run ``prepare_pattern_scaling_training_data`` (which fetches and assembles CMIP6 data over the network) on every generation call and immediately discard the result. Threading ``variable=`` through ``_train_pattern_scaling`` fixes the mismatch and eliminates the redundant work (~13–18 s per generation call in the NorESM2-MM profiling workloads).
 - Variable length timeseries now works also when don't have "year" as time dimension.
 - Fixes to generate annual and monthly gridded ensembles with unified noise and preserving more of the variance.
 
diff --git a/src/meteor/cmip6_meteor_data_getter.py b/src/meteor/cmip6_meteor_data_getter.py
index a119625..5f605a6 100644
--- a/src/meteor/cmip6_meteor_data_getter.py
+++ b/src/meteor/cmip6_meteor_data_getter.py
@@ -1083,7 +1083,9 @@ def prepare_pattern_scaling_training_data(
         )
         return training_data
 
-    def validate_pattern_scaling_cache(self, cache_file, model_name, scenario="aer"):
+    def validate_pattern_scaling_cache(
+        self, cache_file, model_name, scenario="aer", variable=None
+    ):
         """
         Validate a cached pattern scaling model file.
 
@@ -1098,6 +1100,11 @@ def validate_pattern_scaling_cache(self, cache_file, model_name, scenario="aer")
             Expected model name
         scenario : str, optional
             Scenario suffix for expected model name. Default is "aer".
+        variable : str, optional
+            Variable suffix appended to the expected model name (e.g. "tas").
+            Must match the name ``MeteorPatternScaling`` uses when saving the
+            per-variable model. When omitted, the legacy (no-suffix) form is
+            used, but this will mismatch caches saved with per-variable names.
 
         Returns
         -------
@@ -1121,7 +1128,14 @@ def validate_pattern_scaling_cache(self, cache_file, model_name, scenario="aer")
         ...     print(f"✅ {info['message']}")
         """
         expected_name = f"cmip6-{model_name}-{scenario}"
-        expected_vars = set(self.flds)
+        if variable is not None:
+            expected_name = f"{expected_name}-{variable}"
+            # Per-variable caches only carry that one variable; asking for all
+            # of self.flds here would spuriously reject a valid single-variable
+            # cache and force a re-fit.
+            expected_vars = {variable}
+        else:
+            expected_vars = set(self.flds)
 
         info = {
             "expected_name": expected_name,
diff --git a/src/meteor/meteor_interface.py b/src/meteor/meteor_interface.py
index 3da2183..3d9af58 100644
--- a/src/meteor/meteor_interface.py
+++ b/src/meteor/meteor_interface.py
@@ -405,9 +405,14 @@ def _train_pattern_scaling(self, variable, config, verbose=True):
             self.model, variable=variable
         )
 
-        # Check cache
+        # Check cache. The variable suffix must match the name
+        # MeteorPatternScaling saves under (``cmip6-{model}-aer-{variable}``);
+        # without it the validator returns "invalid" for a cache that then
+        # loads successfully by filename, causing wasted training-data prep.
         is_valid, cached_model, info = (  # pylint: disable=unused-variable
-            self.data_getter.validate_pattern_scaling_cache(cache_file, self.model)
+            self.data_getter.validate_pattern_scaling_cache(
+                cache_file, self.model, variable=variable
+            )
         )
 
         if is_valid:
diff --git a/tests/unit/test_cmip6_meteor_data_getter.py b/tests/unit/test_cmip6_meteor_data_getter.py
index 8798cec..0ae5d02 100644
--- a/tests/unit/test_cmip6_meteor_data_getter.py
+++ b/tests/unit/test_cmip6_meteor_data_getter.py
@@ -773,6 +773,35 @@ def test_caching(tmp_path):
     assert cached_model is None
     assert "Error reading cached file" in info["message"]
 
+    # variable= must be honored so that per-variable caches (which
+    # MeteorPatternScaling saves under `cmip6-{model}-aer-{variable}` and which
+    # only carry that one variable in patternflds) round-trip.
+    # Without this, MeteorInterface prepares training data on every call because
+    # this check returns "invalid" even when the pkl loads fine by filename.
+    per_var_file = tmp_path / "per_variable.pkl"
+    per_var_name = "cmip6-ModelA-aer-tas"
+    with open(per_var_file, "wb") as handle:
+        pickle.dump(
+            {"name": per_var_name, "patternflds": {"tas": None}},
+            handle,
+        )
+
+    # No variable= arg: legacy validator, mismatches per-variable name.
+    is_valid, _, info = data_getter.validate_pattern_scaling_cache(
+        str(per_var_file), "ModelA"
+    )
+    assert is_valid is False
+    assert "Model name mismatch" in info["message"]
+
+    # variable="tas": matches name AND expected_vars is narrowed to just
+    # {"tas"} instead of self.flds, so the single-variable pkl is accepted.
+    is_valid, cached_model, info = data_getter.validate_pattern_scaling_cache(
+        str(per_var_file), "ModelA", variable="tas"
+    )
+    assert is_valid is True
+    assert cached_model is not None
+    assert "Cache valid" in info["message"]
+
 
 # def test_error_handling_for_invalid_experiments_and_fields():
 #     """Test error handling for invalid experiments and fields to hit lines 585-593."""