First guess is a numerical problem in one or more of
the element models, that has gotten the simulator
"out over its skis" to where it can't back down the
timestep enough to get convergence (because the last,
was insane and the insanity does not repeat neatly).
zero-capacitance nodes (as often are found in
behaviorals) are one way to get there. LT loved behaviorals
for a couple of reasons (one being, you don't get to
look inside their encrypted ones - so good luck with
the debug).
Try optional transient algorithms like Euler and higher
order TrapGear (if those are accessible - never spent
that much time with LTSpice options). Try setting a cmin
value to load down the matrix some, which could hold
things that blow up to NaN now, to a valid floating point
range.
Barf messages might point to a node that fails to converge
(if you let it sit long enough). If you set the sweep time to
barely before the error-time, you might be able to eyeball
some "well, that's not right..." that leads to the culprit.