diff --git a/scripts/sp2xr_pipeline.py b/scripts/sp2xr_pipeline.py
index 9c4e325..f7525c7 100644
--- a/scripts/sp2xr_pipeline.py
+++ b/scripts/sp2xr_pipeline.py
@@ -213,7 +213,7 @@ def main():
                         "Sample Flow Controller Read (vccm)": pd.Series(
                             dtype="float64"
                         ),
-                        "date": pd.Series(dtype="datetime64[ns]"),
+                        "date": pd.Series(dtype="str"),
                         "hour": pd.Series(dtype="int64"),
                     },
                     index=pd.DatetimeIndex([]),
@@ -340,7 +340,7 @@ def main():
                 delete_partition_if_exists(
                     output_path=f"{run_config['output']}/pbp_calibrated",
                     partition_values={
-                        "date": chunk_start.strftime("%Y-%m-%d 00:00:00"),
+                        "date": chunk_start.strftime("%Y-%m-%d"),
                         "hour": chunk_start.hour,
                     },
                 )
@@ -411,9 +411,8 @@ def main():
                     )
 
                     # 2) cast partition columns *before* Dask strips them off
-                    ddf_conc["date"] = dd.to_datetime(ddf_conc["date"]).astype(
-                        "datetime64[ns]"
-                    )
+                    # Keep date as string to avoid Windows path issues with datetime partitions
+                    ddf_conc["date"] = ddf_conc["date"].astype("str")
                     ddf_conc["hour"] = ddf_conc["hour"].astype("int64")
 
                     conc_future = ddf_conc.to_parquet(
diff --git a/src/sp2xr/calibration.py b/src/sp2xr/calibration.py
index 31b18de..7b377c2 100644
--- a/src/sp2xr/calibration.py
+++ b/src/sp2xr/calibration.py
@@ -182,7 +182,9 @@ def calibrate_particle_data(
         curve_type=scatt_conf.get("curve_type"),
         params=scatt_conf.get("parameters", []),
     )
-    df["date"] = df.index.date.astype("datetime64[ns]")
+    # Use string format for date to avoid Windows path issues with datetime partitions
+    # PyArrow creates partition dirs like "date=2019-05-08 00:00:00" which has colons (invalid on Windows)
+    df["date"] = df.index.strftime("%Y-%m-%d")
     df["hour"] = df.index.hour.astype("int8")
     return df
 
@@ -218,8 +220,7 @@ def calibrate_single_particle(
     meta_cal = ddf_raw._meta.copy()
     meta_cal["BC mass"] = pd.Series([], dtype="float64")
     meta_cal["Opt diam"] = pd.Series([], dtype="float64")
-    meta_cal["time_lag"] = pd.Series([], dtype="int8")
-    meta_cal["date"] = pd.Series([], dtype="datetime64[ns]")
+    meta_cal["date"] = pd.Series([], dtype="str")
     meta_cal["hour"] = pd.Series([], dtype="int8")
 
     ddf_cal = ddf_raw.map_partitions(
diff --git a/src/sp2xr/distribution.py b/src/sp2xr/distribution.py
index c017fb6..5d5b787 100644
--- a/src/sp2xr/distribution.py
+++ b/src/sp2xr/distribution.py
@@ -427,7 +427,8 @@ def process_histograms(
         # merged_ddf["date"] = dd.to_datetime(merged_ddf.index.to_series()).dt.normalize()
         # merged_ddf["hour"] = merged_ddf["hour"].astype("int64")
         time_index = dd.to_datetime(merged_ddf.index.to_series())
-        merged_ddf["date"] = time_index.dt.normalize()  # works on Series
+        # Use string format to avoid Windows path issues with datetime partitions
+        merged_ddf["date"] = time_index.dt.strftime("%Y-%m-%d")
         merged_ddf["hour"] = time_index.dt.hour.astype("int64")
 
         # --- Save hists to parquet
diff --git a/src/sp2xr/resample_pbp_hk.py b/src/sp2xr/resample_pbp_hk.py
index 62bfba6..c8c71bd 100644
--- a/src/sp2xr/resample_pbp_hk.py
+++ b/src/sp2xr/resample_pbp_hk.py
@@ -92,7 +92,8 @@ def build_dt_summary(pdf: pd.DataFrame, dt_s: int = 1) -> pd.DataFrame:
     out = out[out["original_idx"] != 0].drop(columns="original_idx")
 
     # add date / hour helper cols
-    out["date"] = out.index.normalize()
+    # Use string format to avoid Windows path issues with datetime partitions
+    out["date"] = out.index.strftime("%Y-%m-%d")
     out["hour"] = out.index.hour.astype("int64")
 
     return out
@@ -130,7 +131,8 @@ def resample_hk_partition(pdf: pd.DataFrame, dt="1s") -> pd.DataFrame:
     out = out.asfreq(dt).ffill(limit=1)
 
     # --- add date/hour columns ---
-    out["date"] = out.index.normalize()
+    # Use string format to avoid Windows path issues with datetime partitions
+    out["date"] = out.index.strftime("%Y-%m-%d")
     out["hour"] = out.index.hour.astype("int64")
 
     return out
@@ -176,7 +178,8 @@ def aggregate_dt(ddf_pbp_dt, ddf_hk_dt, run_config):
     )
     time_index = dd.to_datetime(ddf_pbp_hk_dt.index.to_series())
 
-    ddf_pbp_hk_dt["date"] = time_index.dt.normalize()  # works on Series
+    # Use string format to avoid Windows path issues with datetime partitions
+    ddf_pbp_hk_dt["date"] = time_index.dt.strftime("%Y-%m-%d")
     ddf_pbp_hk_dt["hour"] = time_index.dt.hour.astype("int64")
 
     # Optionally drop the old columns
diff --git a/src/sp2xr/schema.py b/src/sp2xr/schema.py
index 48f77d0..3c27640 100644
--- a/src/sp2xr/schema.py
+++ b/src/sp2xr/schema.py
@@ -6,7 +6,8 @@ from typing import Mapping
 # schema.py  ── central truth table ───────────────────────────────────────────
 CANONICAL_DTYPES = {
     # identifiers / file bookkeeping
-    "date": "datetime64[ns]",
+    # Use string for date to avoid Windows path issues with datetime partitions
+    "date": "string[pyarrow]",
     "hour": pd.Int64Dtype(),
     "file": "string[pyarrow]",
     "folder_name": "string[pyarrow]",