How do you measure numerical convergence in a storm simulation?
Estimate convergence order by comparing a consistent error metric as grid spacing decreases. For successive grid halvings, take the base-two logarithm of the error ratio.
Start with How do you initialize a warm bubble in an idealized supercell model?.
| Relative spacing | Error | Order relative to previous |
|---|---|---|
| 1 | 0.8 | First run |
| 0.5 | 0.2 | 2 |
| 0.25 | 0.05 | 2 |
Worked example
If error behaves like CΔxᵖ, halving Δx reduces error by 2ᵖ. Errors 0.8, 0.2 and 0.05 each decrease by a factor of four, giving p = 2. Use a trusted reference and keep the comparison time and physical problem fixed. Other error sources can dominate at some resolutions.
Practice questions
- What order corresponds to error shrinking by a factor of two each time?
- What does an increasing error after refinement suggest?
Worked answers
Answer 1
log₂2 = 1.
Answer 2
Check whether the assumed convergence regime applies and whether other errors, settings or a mismatched reference dominate.
- log₂2 = 1.
- Check whether the assumed convergence regime applies and whether other errors, settings or a mismatched reference dominate.