Should Integrator Validation Use Same Integrator as Production?

[EidRashed]

Should integrator validation use the same integrator as production runs, or is using different integrators the correct methodology?

My Current Setup

• Production simulations: LangevinIntegrator (NPT ensemble, thermostated)
• Integrator validation: VerletIntegrator (NVE ensemble, energy-conserving)
• Other tests (ensemble, kinetic energy, equipartition): Same LangevinIntegrator as production

The Dilemma

I'm getting integrator validation failures (ratios 1.1-13.2 instead of expected ~4.0), but all other tests pass beautifully:

• ✅ Ensemble validation: Excellent (1-2 quantiles)
• ✅ Kinetic energy: Perfect mean (300.16K), narrow variance (290K)
• ✅ Equipartition: All means ~300K

This makes me wonder: Am I testing the wrong thing by using a different integrator?

What I've Found So Far

Supporting different integrators:

• Shirts et al. (2013): "For symplectic integrators with NVE simulations, the RMS error will scale with the square of the step size"
• Physical validation docs: "For a symplectic integrator, the fluctuation is expected to be directly proportional to the square of the time step"
• All published examples seem to use NVE for integrator testing

My concern:

• If my production uses LangevinIntegrator, shouldn't I validate that specific integrator?
• Could VerletIntegrator be failing while LangevinIntegrator would pass?
• Are we testing system setup quality or actual integrator performance?

Specific Questions for Experts

1. Methodology: Is using different integrators for validation vs production the established best practice?

2. Interpretation: With my results pattern (integrator fails, everything else passes), what does this actually tell me about my simulation quality?

3. Practical: Should I be concerned about using these production parameters for research, given the integrator validation failure?

[mrshirts]

Integrator validation can only be used with integrators that give NVE simulations - the same relationships will not hold with stochastic integrators like Langevin. It's most useful for testing things like whether the energy function is continuous or not, and can also be affected by whether the initial configuration is strongly out of equilibrium.

Note that also the integrator validation only gives quadratic ratios for symplectic integrators - it could be the VerletIntegration, for whatever reason, is not symplectic?

Kinetic energy: Perfect mean (300.16K), narrow variance (290K)

You can run the ensemble check on potential and kinetic energy separately, and they should separately satisfy the ensemble validation checker. With a narrow KE variance, I would be interested in whether the kinetic energy separately obeys the right relationships.

This is a little narrow on the variance, I would be a little worried. What is the uncertainty on the variance (the uncertainty on the variance is almost alway significantly larger than the uncertainty of the mean)? What happens if you run with longer simulations, does it get closer to 290.

Ensemble validation: Excellent (1-2 quantiles)

Then in most cases, your thermodynamics are probably going to be OK. Randomizing velocities can absorb some lack of energy conservation and still give more or less quantitatively the right configuration space with the right weightings.