Feasibility of synthetic, simulated predictions w/o matching against empirical physical measurement?

This has puzzled me lately.

So, while it's sometimes common to approach modelling and model development by making empirical observations and measurements first (e.g. small experiments). And then developing a model that "matches" against these observations.

However, is it possible to also take a "synthetic" approach to prediction in the sense of:

  • if the phenomenon is complex


  • why not create a model that measures general patterns that do not depend on "observation-level" phenomena? I.e. why not create a model that does not require collecting empirical observations.

However, as it sounds a bit "unorthodox", then are there perhaps resources that would discuss the feasibility of this idea?

An example:

  • Could one predict the growth of cancer by extrapolating from properties that increase the risk of cancer growth, instead of creating a model that matches against empirically observed cancer growth patterns?
  • $\begingroup$ These kinds of measurements are called "synthetic measurements". They are of course only useful if you have a model that accurately describes both the system you are interested in and the measurement uncertainty. $\endgroup$ Nov 18, 2021 at 18:25
  • $\begingroup$ @Wolfgang But don't physical laws also contain uncertainty? I'm not very experienced in this accuracy perspective, but I've perceived that e.g. fluid and mechanical laws are "rough approximations". E.g. because of things like: en.wikipedia.org/wiki/Factor_of_safety $\endgroup$
    – mavavilj
    Nov 18, 2021 at 18:57
  • $\begingroup$ The physical laws are best considered axiomatic within their realm of applicability, but any more involved model is just a reflection of the assumptions that went into its derivation. For example, linear elasticity is a good model as long as deformations are small; but don't deform that piece of metal too much, or it becomes nonlinear and then plastic! That's why I say that synthetic models are only useful if you have a good model. $\endgroup$ Nov 18, 2021 at 23:41


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.