diff --git a/apax/train/loss.py b/apax/train/loss.py
index 56096455..cd28786e 100644
--- a/apax/train/loss.py
+++ b/apax/train/loss.py
@@ -52,6 +52,13 @@ def force_angle_exponential_weight(
     return (1.0 - dotp) * jnp.exp(-F_0_norm) / divisor
 
 
+def stress_tril(label, prediction, divisor=1.0):
+    idxs = jnp.tril_indices(3)
+    label_tril = label[:, idxs[0], idxs[1]]
+    prediction_tril = prediction[:, idxs[0], idxs[1]]
+    return (label_tril - prediction_tril) ** 2 / divisor
+
+
 loss_functions = {
     "molecules": weighted_squared_error,
     "structures": weighted_squared_error,
@@ -59,6 +66,7 @@ def force_angle_exponential_weight(
     "cosine_sim": force_angle_loss,
     "cosine_sim_div_magnitude": force_angle_div_force_label,
     "cosine_sim_exp_magnitude": force_angle_exponential_weight,
+    "tril": stress_tril,
 }
 
 
@@ -101,6 +109,7 @@ def determine_divisor(self, n_atoms: jnp.array) -> jnp.array:
                 n_atoms, "batch -> batch 1 1"
             ),
             "stress_structures": einops.repeat(n_atoms**2, "batch -> batch 1 1"),
+            "stress_tril": einops.repeat(n_atoms**2, "batch -> batch 1 1"),
             "stress_vibrations": einops.repeat(n_atoms, "batch -> batch 1 1"),
         }
         divisor = divisor_dict.get(divisor_id, jnp.array(1.0))