Accumulation of error in Abaqus Explicit?

[jball1]

I am performing an analysis of an event that takes between 0.1 seconds and 1 second. My stable time increment is usually between 2E-7 and 5E-6. In the past, I have used Abaqus Implicit to solve this type of problem. I believe I was told at one point that Implicit methods are generally better for simulating long duration events, since the error accumulates in Explicit due to the much larger number of steps relative to Implicit.

For example, I can usually capture the results that I care about with a time step size of 5E-5. So for a solution I am running out to 0.1 sec, that would be 2,000 time steps in Implicit, and 20,000 time steps in Explicit with a time increment size of 5E-6, and 500,000 time steps in Explicit with a time increment size of 2E-7.

I am curious - has anyone actually found this to be the case? Have you run a simulation for a long duration event and found that the error in the explicit solution builds over time?

Of course, I can run my own studies to compare answers (and I plan to!). But these simulations are pretty expensive, and so if I can benefit here from anyone else's experience, it would be much appreciated!

 

[FEA way]

The main problem is that Abaqus/Explicit analyses can take a lot of time to solve when the step time is higher (because of the growing number of increments that need to be calculated using the same stable time increment). In the case of quasi-static analyses, you can use mass scaling to achieve a significant speed-up.

When it comes to accuracy, it is recommended to enable double precision for the Abaqus/Explicit solver to reduce round-off errors when the number of increments is large (Abaqus warns about this when the estimated number of increments is higher than 300 000 and throws an error when it's higher than 20 000 000 but single precision is still used). Of course, this also increases the computational cost of the analysis. According to the Abaqus documentation, single precision typically provides 20-30% CPU savings and is sufficiently accurate unless the number of increments is large (above 300 000 as stated before), nodal displacement increments are less than 10^-6 times the corresponding nodal coordinates, the analysis involves hyperelastic materials or multiple revolutions of deformable parts.

 

[jball1]

Thanks, I have never noticed those specific numbers of increments in the documentation. That is very helpful.

 

[jball1]

Actually would you mind pointing me to those numbers in the Abaqus documentation?

 

[FEA way]

Actually would you mind pointing me to those numbers in the Abaqus documentation?

It’s in the chapter Analysis —> Analysis Procedures —> Introduction —> Defining an Analysis (paragraph "Precision Level of the Abaqus/Explicit Executable").