tridiagonal coefficient inversion lacks singularity guard

[asalmgren]

Location

• Source/TimeIntegration/ERF_MakeFastCoeffs.cpp:243-247
• Source/TimeIntegration/ERF_MakeFastCoeffs.cpp:293-296

Problem

The Thomas forward elimination computes bet and stores it in coeffB_a(i,j,k), then later inverts it unconditiona
lly:

coeffB_a(i,j,k) = 1.0 / coeffB_a(i,j,k);

No guard is present for near-zero diagonal values.

Why this is a bug

If the fast linear system becomes locally singular/ill-conditioned, this creates infinities/NaNs that then pollute m
omentum and conserved updates.

Suggested patch

Add a safety check before inversion (with abort for now; alternative is clamping with diagnostics).

--- a/Source/TimeIntegration/ERF_MakeFastCoeffs.cpp
+++ b/Source/TimeIntegration/ERF_MakeFastCoeffs.cpp
@@
-            ParallelFor(bx_shrunk_in_k, [=] AMREX_GPU_DEVICE (int i, int j, int k)
-            {
-                coeffB_a(i,j,k) = 1.0 / coeffB_a(i,j,k);
-            });
+            ParallelFor(bx_shrunk_in_k, [=] AMREX_GPU_DEVICE (int i, int j, int k)
+            {
+                Real b = coeffB_a(i,j,k);
+                if (amrex::Math::abs(b) <= 1.e-14_rt) {
+                    amrex::Abort("make_fast_coeffs: singular tridiagonal diagonal entry");
+                }
+                coeffB_a(i,j,k) = 1.0 / b;
+            });

[asalmgren]

We construct the coefficient in such as a way that it is always non-zero so I will close this issue.