I think the answer is "with great difficulty". When you're developing something like this, you might have an issue with:
- Experiment
- Model
- Approximation to the model
- Numerics
- Code
(- Compiler & hardware)
So, your answer might not match experiment. Maybe the experiment is wrong. Maybe it's an expected effect due to an approximation you used. Maybe your integration grid isn't sufficiently converged. You just have to systematically eliminate these, often by comparing to more expensive but verified simulations, or analytic solutions on simplified models.
- Experiment
- Model
- Approximation to the model
- Numerics
- Code
(- Compiler & hardware)
So, your answer might not match experiment. Maybe the experiment is wrong. Maybe it's an expected effect due to an approximation you used. Maybe your integration grid isn't sufficiently converged. You just have to systematically eliminate these, often by comparing to more expensive but verified simulations, or analytic solutions on simplified models.