diff --git a/Source/Particles/Collision/BinaryCollision/DSMC/CollisionFilterFunc.H b/Source/Particles/Collision/BinaryCollision/DSMC/CollisionFilterFunc.H
index 46b228b049e..c5bd2e1cec6 100644
--- a/Source/Particles/Collision/BinaryCollision/DSMC/CollisionFilterFunc.H
+++ b/Source/Particles/Collision/BinaryCollision/DSMC/CollisionFilterFunc.H
@@ -34,7 +34,7 @@
  * @param[in] scattering processes an array of scattering processes included for consideration.
  * @param[in] engine the random engine.
  */
-template <typename index_type>
+template <int max_process_count, typename index_type>
 AMREX_GPU_HOST_DEVICE AMREX_INLINE
 void CollisionPairFilter (const amrex::ParticleReal u1x, const amrex::ParticleReal u1y,
                           const amrex::ParticleReal u1z, const amrex::ParticleReal u2x,
@@ -65,11 +65,11 @@ void CollisionPairFilter (const amrex::ParticleReal u1x, const amrex::ParticleRe
 
     // Evaluate the cross-section for each scattering process to determine
     // the total collision probability.
-    AMREX_ASSERT_WITH_MESSAGE(
-        (process_count < 4), "Too many scattering processes in DSMC routine."
-    );
-    int coll_type[4] = {0, 0, 0, 0};
-    amrex::ParticleReal sigma_sums[4] = {0._prt, 0._prt, 0._prt, 0._prt};
+
+    // The size of the arrays below is a compile-time constant (template parameter)
+    // for performance reasons: it avoids dynamic memory allocation on the GPU.
+    int coll_type[max_process_count] = {0};
+    amrex::ParticleReal sigma_sums[max_process_count] = {0._prt};
     for (int ii = 0; ii < process_count; ii++) {
         auto const& scattering_process = scattering_processes[ii];
         coll_type[ii] = int(scattering_process.m_type);
diff --git a/Source/Particles/Collision/BinaryCollision/DSMC/DSMCFunc.H b/Source/Particles/Collision/BinaryCollision/DSMC/DSMCFunc.H
index a692d2cbb9e..5a3c925e9bd 100644
--- a/Source/Particles/Collision/BinaryCollision/DSMC/DSMCFunc.H
+++ b/Source/Particles/Collision/BinaryCollision/DSMC/DSMCFunc.H
@@ -163,7 +163,11 @@ public:
                     u1y[I1[i1]] = u1xbuf*std::sin(theta) + u1y[I1[i1]]*std::cos(theta);
 #endif
 
-                    CollisionPairFilter(
+                    const int max_process_count = 4; // Pre-defined value, for performance reasons
+                    AMREX_ALWAYS_ASSERT_WITH_MESSAGE(
+                        (m_process_count < max_process_count), "Too many scattering processes in DSMC routine (hardcoded to only allow 4). Update the max_process_count value in source code to allow more scattering processes."
+                    );
+                    CollisionPairFilter<max_process_count>(
                         u1x[ I1[i1] ], u1y[ I1[i1] ], u1z[ I1[i1] ],
                         u2x[ I2[i2] ], u2y[ I2[i2] ], u2z[ I2[i2] ],
                         m1, m2, w1[ I1[i1] ], w2[ I2[i2] ],