How Long Should You Run an MD Simulation? (And How to Know It Has Equilibrated)

There is no single correct length. Run production long enough that the properties you care about stop drifting and fluctuate around a stable average. For a small, stable protein that often means 100 ns or more, but the honest answer is set by convergence, not a round number. Judge it from RMSD, energy, temperature and pressure plateaus.

Almost every beginner asks the same question before their first real project: how long is long enough? The number you pick has to survive a thesis committee, so guessing “100 ns because everyone uses it” is not a defense. This guide from the StemSkills Lab team (10+ years in sequence and structural bioinformatics, drug discovery and design, and multiscale molecular modeling) shows you how to choose a length you can justify and how to read the plots that tell you the system has settled. It sits inside our pillar guide on learning molecular dynamics with GROMACS.

How long should you run an MD simulation?

Long enough that the observable you are measuring reaches a stable plateau and stays there, with time left over to sample fluctuations around it. Length is a means to an end, and the end is sampling. A 100 ns run that never settles is worse than a 50 ns run that reaches a clean plateau and holds it.

The reason there is no universal number is that biological motions span an enormous range of timescales. Bond vibrations happen in femtoseconds. Side-chain rotations take picoseconds. Loop rearrangements and small conformational changes take nanoseconds to microseconds. Full protein folding can take microseconds to seconds. Your simulation only samples motions faster than its total length, so the process you want to observe sets the floor for how long you run.

A useful reference point comes from the extreme end of the field. In their 2011 study “How Fast-Folding Proteins Fold” (Lindorff-Larsen and colleagues, Science, 2011), a group at D. E. Shaw Research simulated twelve fast-folding proteins on the special-purpose Anton machine, reaching timescales of hundreds of microseconds to about a millisecond to watch complete folding and unfolding. That is far beyond a typical desktop or GPU run, and it is a reminder that a standard 100 ns simulation samples local relaxation and stability, not large-scale folding.

For orientation, here is what different lengths realistically buy you on commodity hardware. Treat these as starting points to test, not promises.

The canonical GROMACS lysozyme tutorial by Justin Lemkul runs a 1 ns production simulation, which is deliberately short so it finishes on a laptop and teaches the workflow. Do not copy that 1 ns as a research length. It is a teaching size, not a scientific one. If you are still assembling that workflow, our walkthrough of your first GROMACS MD simulation covers the full pipeline first.

What does equilibration actually mean?

Equilibration is the phase where your system relaxes from an artificial starting point to a physically reasonable state that no longer depends on where it began. When you build a system, you place the protein, add a solvent box, insert ions and minimize energy. That configuration is arbitrary. The water is not packed the way it would be around the real protein, the temperature is not evenly distributed, and the pressure is off. Equilibration lets those quantities settle before you start collecting data you intend to analyze.

In practice a GROMACS workflow splits this into two short phases before production. First an NVT phase (constant number, volume and temperature) brings the system to your target temperature, usually 300 to 310 K. Then an NPT phase (constant number, pressure and temperature) lets the density and pressure settle to their correct values. Both typically use position restraints on the heavy atoms so the solvent equilibrates around a fixed solute. The settings that control all of this live in the run-parameter file, which we break down in the GROMACS .mdp file explained.

The key idea is that equilibration is not a fixed duration you can look up. It is a state you reach and then verify. You confirm it by plotting the right quantities and checking that they have stopped drifting.

How do you tell if your simulation has equilibrated?

You plot four quantities over time and check that each has reached a plateau: temperature, pressure, potential energy and structural RMSD. Temperature and pressure tell you the thermostat and barostat are working. Potential energy tells you the system has relaxed energetically. RMSD tells you the structure has stopped drifting away from its starting point. GROMACS writes these into the energy file (.edr) during your NPT and production runs, and you extract them with gmx energy.

Run these on your NPT equilibration output first, then on the start of your production trajectory:

• Temperature: gmx energy -f npt.edr -o temperature.xvg, then type the number next to “Temperature” and press Enter twice. It should fluctuate tightly around your setpoint, for example 300 K, with no upward or downward trend.
• Pressure: gmx energy -f npt.edr -o pressure.xvg, then select “Pressure”. Pressure is noisy by nature, so expect large instantaneous swings; what matters is that the running average sits near 1 bar and does not drift.
• Potential energy: gmx energy -f npt.edr -o potential.xvg, then select “Potential”. It should fall during minimization and early equilibration, then level off around a stable value.
• Backbone RMSD: gmx rms -s md.tpr -f md.xtc -o rmsd.xvg and choose the “Backbone” group for both the reference and the calculation. A curve that rises and then flattens signals a relaxed, stable structure.

The table below is the diagnostic checklist you can keep next to your screen. “Equilibrated” always means the same thing: a plateau, not a trend.

Density is worth a fifth check after NPT: it should settle close to about 1000 kg per cubic metre for a water-dominated system. Once temperature, pressure, energy and density are stable, your equilibration phase is done and production can begin. RMSD is then the main gauge you watch during production itself. For the full analysis toolkit around RMSD, see our guide on analyzing RMSD and RMSF from a GROMACS trajectory.

Is equilibration the same as convergence?

No, and confusing the two is the most common mistake in beginner MD write-ups. Equilibration means the system has relaxed to a stable state, which you can reach in a few hundred picoseconds. Convergence means your production run has sampled the relevant conformations enough that the property you report would not change if you ran longer. A simulation can be fully equilibrated and still badly under-sampled.

The uncomfortable truth is that you cannot prove convergence. You can only fail to disprove it. A flat RMSD plateau shows the structure is stable, but it does not guarantee you have sampled every important state. The molecule could sit in one basin for the whole run while a second, relevant conformation is never visited. Grossfield and Zuckerman, in their standard reference on sampling quality (Annual Reports in Computational Chemistry, 2009), make exactly this point: apparent stability in a single trajectory can hide poor sampling, so convergence has to be tested with independent measures such as block averaging and multiple runs rather than assumed from one flat line.

This is why length alone is a weak defense. “I ran 100 ns” says nothing about whether 100 ns was enough for your system. What you actually report is that the observable of interest, whether that is RMSD, a distance, an angle or a binding contact, stabilized and stayed stable across replicas.

So how do you defend “why 100 ns” in a thesis?

You defend it with evidence, not with the number itself. A committee will accept a length that is justified by the data far more readily than a length justified by convention. Build your argument from these steps:

1. State the process you are studying and its expected timescale. If you are measuring the stability of a folded protein or the persistence of a bound ligand, hundreds of nanoseconds is a defensible window because those motions fall inside it. Say so explicitly.
2. Show the equilibration plots. Include temperature, pressure, energy and RMSD reaching plateaus. This proves the production window is measuring a relaxed system, not a relaxing one.
3. Show that your target property has stabilized. Compute your key observable over time and demonstrate it stops drifting. Report the average over the stable window, not the whole trajectory.
4. Run replicas. Three independent runs from different starting velocities that agree are much stronger than one long run. Agreement across replicas is the most convincing convergence evidence you can present.
5. Report uncertainty. Use block averaging or the spread across replicas to give an error bar. A number with an honest error bar reads as science; a bare number reads as a guess.

Framed this way, the answer to “why 100 ns” becomes “because the system equilibrated within the first few nanoseconds, the RMSD and my target property held a stable plateau for the remaining time, and three replicas agreed within the reported uncertainty.” That is a defense a committee cannot easily reject.

One long run or several short replicas?

For most projects, several shorter replicas beat one long run of the same total cost. Three runs of 100 ns started from different random velocity seeds explore three independent trajectories, so agreement between them is real evidence that your result does not depend on one lucky path. A single 300 ns run gives you one path and no internal check. Independent replicas also let you compute a straightforward error bar from the spread of results.

The exception is when you are chasing a single slow event, such as a large conformational transition, that will not happen inside a short window. There, one long trajectory (or an enhanced-sampling method) may be the only way to see the event at all. For the typical stability or protein-ligand question at MSc level, prefer replicas. Choosing the right force field matters just as much as length, which we cover in how to choose a force field for your GROMACS simulation.

Frequently asked questions

Is 100 ns enough for a protein-ligand simulation?

Often yes for judging whether a ligand stays bound and the pose is stable, provided the RMSD of the protein and the ligand both plateau and hold. It is usually not enough for a converged binding free energy, which needs dedicated free-energy methods and much more sampling. Confirm with replicas rather than assuming.

How long should the equilibration phase be?

Common practice is a short NVT phase followed by a short NPT phase, on the order of a few hundred picoseconds each with position restraints. The exact length is less important than the check: extend it until temperature, pressure, potential energy and density are all stable before you start production.

My RMSD keeps rising and never flattens. What does that mean?

A continuously climbing RMSD means the structure has not stabilized. Possible causes include a genuinely flexible or unfolding region, a too-short run, or a preparation problem such as a bad starting structure or force-field choice. Plot RMSD per residue (RMSF) to see whether the drift comes from one loop or the whole protein before concluding the system is unstable.

How many replicas should I run?

Three is a widely used minimum because it lets you see whether independent runs agree and gives a basic estimate of spread. More is better when the compute budget allows, especially for properties that are sensitive to sampling.

Does a longer simulation always give a better result?

No. A longer run samples more, but if the system is poorly prepared or the force field is wrong, more time only produces a more confident wrong answer. Correct setup and convergence checks matter more than raw length.

New to computational biology and not sure where MD fits in the bigger picture? Start with our computational biology skills roadmap to see how simulation sits alongside the other skills grad programs look for.