ecciv/thesis

On strain-rate dependent properties of model ice for simulating ice-structure interaction

thesis's Issues

storing c2nonlin as an improved creep iteration start point

Nonlinear creep is solved by multiplying the creep velocity into the creep coefficient, giving us a velocity-specific creep coefficient, and iterating on this value until the creep velocity*velocity-specific creep coefficient match the force over the elastic and viscoelastic components to a specified tolerance. This iteration can be fed with an initial guess to the velocity-specific creep coefficient, which ideally reduces the number of necessary iterations.

Right now, that guess is always c2_initialguess. However by storing the value of c2nonlin (velocity-specific creep coefficient) from a previous timestep, and assuming that subsequent timesteps have more similarity than the current timestep compared to an average timestep (which is how c2_initalguess is calculated), we can achieve a significant speedup in cases where c2exp is far from 1. It would also decrease the performance sensitivity to poorly chosen c2_initialguess.

this has been implemented in another environment and resulted in a 2-3x speedup for c2exp 1/3.

variable timesteps

variable timestep length would eliminate the need for complicated sub-timestep corrections.

The sub-timestep corrections are working well and there are certain benefits to keeping every timestep equally long, but adding variable timesteps (specifically to break the timestep when an ice element fails) would be a simpler solution and thus preferable.

reimplementation in julia

matlab is a great tool for writing code quickly, but matlab also has many drawbacks. For example, I don't have a matlab license.

Octave is fine but doesn't have support for parallel-computing, which is necessary for acceptable performance. Also Matlab is fairly slow, and Octave is even slower + less stable. I haven't measured Octaves speed in comparison to Matlab, but tests indicate that a near-straight julia port would run 50x faster (single threaded) than the equivalent Octave code.

The drawbacks with julia include

• more verbose than Matlab
• less intuitive for debugging
• worse documentation compared to Octave and especially Matlab

however the speed of julia is hard to beat