When doing molecular dynamics simulations, there are countless sources of error (i.e. approximations and numerics).

Molecular dynamics can be used to determine intrinsic thermal properties such as Melting Point/Freezing Point/Boiling Point. What tools do give me a quantified error on these properties, and what is the underlying thechnique to quantify all these different error terms?

  • $\begingroup$ DO you mean to quantify the uncertainty due to errors other than the statistical uncertainty associated with MD? $\endgroup$
    – jheindel
    Oct 26 '20 at 17:38
  • $\begingroup$ Yes, ideally I would like to have all possible error sources accounted for. If an error (such as it is the case for e.g. the numerical truncation error) is limited by a constant, I am ok with that. $\endgroup$ Oct 27 '20 at 12:04

The error of a MD simulation can be roughly partitioned into four contributions:

  1. Error due to short-period statistical fluctuations;
  2. Error due to poor- or non-ergodicity (in other words, extremely long-period statistical fluctuations);
  3. Error due to numerical round-off and the use of a finite time step;
  4. Error due to the force field itself.

Quantification of 1 is trivial, for example you can compute your desired property using a series of short MD trajectories, and compute the standard deviation of their mean. 2 is very hard and basically impossible to reliably estimate, because it requires one to sample all low-barrier paths that leads the simulation to a basin with low free energy, which should be NP-hard. 3 is probably predominantly reflected in the conservation of energy, and if in doubt, you can always estimate it by using higher numerical precision (double precision instead of single precision) and a smaller time step.

4 is the most interesting (and arguably the most underrated) contribution. There are some methods for estimating it, for example https://aip.scitation.org/doi/10.1063/1.3545069. However this requires that you go into the training set of the force field, and also have knowledge of the experimental uncertainties of the training data, which can be a non-trivial amount of work. You may search for articles that cite this paper for more examples of error quantification of the force field itself.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.