diff --git a/src/multiego/ensemble.py b/src/multiego/ensemble.py
index 3c5712fa..fa2b44e9 100644
--- a/src/multiego/ensemble.py
+++ b/src/multiego/ensemble.py
@@ -1101,47 +1101,26 @@ def generate_LJ(meGO_ensemble, train_dataset, check_dataset, parameters):
 
     for sym in symmetries:
         for atypes in itertools.combinations(sym[1:], 2):
-            ai_rows = meGO_LJ[meGO_LJ["ai"].str.startswith(f"{atypes[0]}_") & (mglj_resn_ai == sym[0])].copy()
-            aj_rows = meGO_LJ[meGO_LJ["aj"].str.startswith(f"{atypes[0]}_") & (mglj_resn_aj == sym[0])].copy()
-            ai_rows.loc[:, "ai"] = (
-                atypes[1] + "_" + ai_rows["ai"].str.split("_").str[1] + "_" + ai_rows["ai"].str.split("_").str[2]
+            stmp_df_ai = meGO_LJ[meGO_LJ["ai"].str.startswith(f"{atypes[0]}_") & (mglj_resn_ai == sym[0])].copy()
+            stmp_df_aj = meGO_LJ[meGO_LJ["aj"].str.startswith(f"{atypes[0]}_") & (mglj_resn_aj == sym[0])].copy()
+            stmp_df_ai.loc[:, "ai"] = (
+                atypes[1] + "_" + stmp_df_ai["ai"].str.split("_").str[1] + "_" + stmp_df_ai["ai"].str.split("_").str[2]
             )
-            aj_rows.loc[:, "aj"] = (
-                atypes[1] + "_" + aj_rows["aj"].str.split("_").str[1] + "_" + aj_rows["aj"].str.split("_").str[2]
+            stmp_df_aj.loc[:, "aj"] = (
+                atypes[1] + "_" + stmp_df_aj["aj"].str.split("_").str[1] + "_" + stmp_df_aj["aj"].str.split("_").str[2]
             )
-
-            ai_rows["ai"], ai_rows["aj"] = np.where(
-                ai_rows["ai"].str.split("_").str[2] < ai_rows["aj"].str.split("_").str[2],
-                (ai_rows["ai"], ai_rows["aj"]),
-                (ai_rows["aj"], ai_rows["ai"]),
-            )
-            aj_rows["ai"], aj_rows["aj"] = np.where(
-                aj_rows["ai"].str.split("_").str[2] < aj_rows["aj"].str.split("_").str[2],
-                (aj_rows["ai"], aj_rows["aj"]),
-                (aj_rows["aj"], aj_rows["ai"]),
-            )
-            tmp_df = pd.concat([tmp_df, ai_rows, aj_rows])
-
-            ai_rows = meGO_LJ[meGO_LJ["ai"].str.startswith(f"{atypes[1]}_") & (mglj_resn_ai == sym[0])].copy()
-            aj_rows = meGO_LJ[meGO_LJ["aj"].str.startswith(f"{atypes[1]}_") & (mglj_resn_aj == sym[0])].copy()
-            ai_rows.loc[:, "ai"] = (
-                atypes[0] + "_" + ai_rows["ai"].str.split("_").str[1] + "_" + ai_rows["ai"].str.split("_").str[2]
+            if tmp_df.empty:
+                continue
+            tmp_df_i = tmp_df[tmp_df["ai"].str.startswith(f"{atypes[0]}_") & (mglj_resn_ai == sym[0])].copy()
+            tmp_df_j = tmp_df[tmp_df["aj"].str.startswith(f"{atypes[0]}_") & (mglj_resn_aj == sym[0])].copy()
+            tmp_df_i.loc[:, "ai"] = (
+                atypes[1] + "_" + tmp_df_i["ai"].str.split("_").str[1] + "_" + tmp_df_i["ai"].str.split("_").str[2]
             )
-            aj_rows.loc[:, "aj"] = (
-                atypes[0] + "_" + aj_rows["aj"].str.split("_").str[1] + "_" + aj_rows["aj"].str.split("_").str[2]
+            tmp_df_j.loc[:, "aj"] = (
+                atypes[1] + "_" + tmp_df_j["aj"].str.split("_").str[1] + "_" + tmp_df_j["aj"].str.split("_").str[2]
             )
 
-            ai_rows["ai"], ai_rows["aj"] = np.where(
-                ai_rows["ai"].str.split("_").str[2] < ai_rows["aj"].str.split("_").str[2],
-                (ai_rows["ai"], ai_rows["aj"]),
-                (ai_rows["aj"], ai_rows["ai"]),
-            )
-            aj_rows["ai"], aj_rows["aj"] = np.where(
-                aj_rows["ai"].str.split("_").str[2] < aj_rows["aj"].str.split("_").str[2],
-                (aj_rows["ai"], aj_rows["aj"]),
-                (aj_rows["aj"], aj_rows["ai"]),
-            )
-            tmp_df = pd.concat([tmp_df, ai_rows, aj_rows])
+            tmp_df = pd.concat([tmp_df, stmp_df_ai, stmp_df_aj, tmp_df_i, tmp_df_j])
 
     meGO_LJ = pd.concat([meGO_LJ, tmp_df])
 
@@ -1267,57 +1246,33 @@ def generate_LJ(meGO_ensemble, train_dataset, check_dataset, parameters):
         mglj_resn_aj = meGO_check_contacts["aj"].map(dict_sbtype_to_resname)
         tmp_df = pd.DataFrame()
         # apply symmetries to check contacts
+        tmp_df = pd.DataFrame()
         for sym in symmetries:
             for atypes in itertools.combinations(sym[1:], 2):
-                ai_rows = meGO_check_contacts[
+                stmp_df_ai = meGO_check_contacts[
                     meGO_check_contacts["ai"].str.startswith(f"{atypes[0]}_") & (mglj_resn_ai == sym[0])
                 ].copy()
-                aj_rows = meGO_check_contacts[
+                stmp_df_aj = meGO_check_contacts[
                     meGO_check_contacts["aj"].str.startswith(f"{atypes[0]}_") & (mglj_resn_aj == sym[0])
                 ].copy()
-                ai_rows.loc[:, "ai"] = (
-                    atypes[1] + "_" + ai_rows["ai"].str.split("_").str[1] + "_" + ai_rows["ai"].str.split("_").str[2]
+                stmp_df_ai.loc[:, "ai"] = (
+                    atypes[1] + "_" + stmp_df_ai["ai"].str.split("_").str[1] + "_" + stmp_df_ai["ai"].str.split("_").str[2]
                 )
-                aj_rows.loc[:, "aj"] = (
-                    atypes[1] + "_" + aj_rows["aj"].str.split("_").str[1] + "_" + aj_rows["aj"].str.split("_").str[2]
+                stmp_df_aj.loc[:, "aj"] = (
+                    atypes[1] + "_" + stmp_df_aj["aj"].str.split("_").str[1] + "_" + stmp_df_aj["aj"].str.split("_").str[2]
                 )
-
-                ai_rows["ai"], ai_rows["aj"] = np.where(
-                    ai_rows["ai"].str.split("_").str[2] < ai_rows["aj"].str.split("_").str[2],
-                    (ai_rows["ai"], ai_rows["aj"]),
-                    (ai_rows["aj"], ai_rows["ai"]),
+                if tmp_df.empty:
+                    continue
+                tmp_df_i = tmp_df[tmp_df["ai"].str.startswith(f"{atypes[0]}_") & (mglj_resn_ai == sym[0])].copy()
+                tmp_df_j = tmp_df[tmp_df["aj"].str.startswith(f"{atypes[0]}_") & (mglj_resn_aj == sym[0])].copy()
+                tmp_df_i.loc[:, "ai"] = (
+                    atypes[1] + "_" + tmp_df_i["ai"].str.split("_").str[1] + "_" + tmp_df_i["ai"].str.split("_").str[2]
                 )
-                aj_rows["ai"], aj_rows["aj"] = np.where(
-                    aj_rows["ai"].str.split("_").str[2] < aj_rows["aj"].str.split("_").str[2],
-                    (aj_rows["ai"], aj_rows["aj"]),
-                    (aj_rows["aj"], aj_rows["ai"]),
+                tmp_df_j.loc[:, "aj"] = (
+                    atypes[1] + "_" + tmp_df_j["aj"].str.split("_").str[1] + "_" + tmp_df_j["aj"].str.split("_").str[2]
                 )
-                tmp_df = pd.concat([tmp_df, ai_rows, aj_rows])
 
-                ai_rows = meGO_check_contacts[
-                    meGO_check_contacts["ai"].str.startswith(f"{atypes[1]}_") & (mglj_resn_ai == sym[0])
-                ].copy()
-                aj_rows = meGO_check_contacts[
-                    meGO_check_contacts["aj"].str.startswith(f"{atypes[1]}_") & (mglj_resn_aj == sym[0])
-                ].copy()
-                ai_rows.loc[:, "ai"] = (
-                    atypes[0] + "_" + ai_rows["ai"].str.split("_").str[1] + "_" + ai_rows["ai"].str.split("_").str[2]
-                )
-                aj_rows.loc[:, "aj"] = (
-                    atypes[0] + "_" + aj_rows["aj"].str.split("_").str[1] + "_" + aj_rows["aj"].str.split("_").str[2]
-                )
-
-                ai_rows["ai"], ai_rows["aj"] = np.where(
-                    ai_rows["ai"].str.split("_").str[2] < ai_rows["aj"].str.split("_").str[2],
-                    (ai_rows["ai"], ai_rows["aj"]),
-                    (ai_rows["aj"], ai_rows["ai"]),
-                )
-                aj_rows["ai"], aj_rows["aj"] = np.where(
-                    aj_rows["ai"].str.split("_").str[2] < aj_rows["aj"].str.split("_").str[2],
-                    (aj_rows["ai"], aj_rows["aj"]),
-                    (aj_rows["aj"], aj_rows["ai"]),
-                )
-                tmp_df = pd.concat([tmp_df, ai_rows, aj_rows])
+                tmp_df = pd.concat([tmp_df, stmp_df_ai, stmp_df_aj, tmp_df_i, tmp_df_j])
 
         meGO_check_contacts = pd.concat([meGO_check_contacts, tmp_df])
         # check contacts are all repulsive so among duplicates we keep the one with shortest distance