The 1 thing I noticed:
Matching the boundary condition in the Fern et al paper is a bit tricky. In the paper, they are preemptively placing material points along the entire base of the model:
The columns rested on a thin layer called the base layer which provided the friction necessary for the deposition. It is modelled as a stiff elastic body. According to the experimental data, the friction of the base layer plays a small role in the column collapse  and was confirmed numerically when some realistic friction angles were applied .
I notice you are imposing a zero-velocity constraint in the x-dir along the model base. I don’t think a frictional layer and zero-velocity condition are equivalent in this case. It might be tough to replicate results with this approach.
Without changing any boundary conditions, I reduced the time step to
dt=1e-5 which appears to improve stability. Even if
dt is below something like the CFL condition, in my experience “shooting material points” often means dt is still too big. With the smaller
dt, I can run your model until around t=1.0s when material points begin to reach the right mesh boundary and displace upwards due to the nodal velocity constraints.
Time Step + Friction Constraint
Next, I changed the bottom boundary condition to be a friction constraint and kept the reduced time step of
dt=1e-5. I could get this to run until t=2.5s. Changing the friction coef. along the base might help make this one look a bit more like the paper.