I was recently looking in the library for Process Noise options and found a couple of options.
As far as I see, there is a ConstantProcessNoise class, which, from what I could understand, is a State Noise Compensation algorithm.
In addition, I see the UnivariateProcessNoise class that can be used to have a Process Noise matrix that changes with time. However, that requires a certain degree of knowledge regarding the variational behaviour of the un-modelled dynamics within the filter.
What if I would like to estimate these un-modelled accelerations?
Supposing I have real data from an observation system, I would like to understand which un-modelled perturbations could be worthy of being included in the filter or understand and quantify the difference between real and modelled perturbations.
Does Orekit have any class/function for Dynamic Model Compensation?

Thank you in advance for any clarification,
Edoardo.

Not in the ProcessNoise classes no.
Orekit has a forces.empirical package where you will find acceleration models that can be used to model an unknown process.
The ParametricAcceleration takes a direction and either a HarmonicAccelerationModel or a PolynomialAccelerationModel.
During an orbit determination process (least-squares or Kalman) the coefficients of the acceleration models can be estimated.
There is an example in the tutorials of a polynomial acceleration usage : NumericalOrbitDetermination (this is a real-life satellite in GTO that experienced a thruster leak).

Is it close to what youâ€™re looking for or am I off-topic?

Thank you for the answer and the welcome.
I am not sure whether this would work for us. I need to think about it and discuss it with my colleague.
I am not sure because, okay, letâ€™s suppose that I want to estimate the atmospheric drag acting on my spacecraft in LEO. I could use a HarmonicAccelerationModel to estimate the acceleration modelâ€™s parameters of the overall perturbation, but will this eventually be able to track spikes in magnitude or non-periodic behaviours?
At the same time, a first (or second) order Gauss-Markov processes could hopefully track them.

I probably had to ask how and if you can model a Gauss-Markov process in Orekit.

We do have Gauss-Markov processes but only at Hipparchus level, it was added this year. I typically uses it to model some types of errors in measurements simulation. You can use it for your own models, but you will have to built it by yourself.

That is more the direction that I am interested in.
The best solution would be to have something like the DMC described in Statistical Orbit Determination which uses a first-order Gauss-markov process to represent the unknown acceleration (or a second order Gauss-Markov process like the one described here).

I wonder whether, in order to do this, one would need to significantly modify Orekitâ€™s code in order to expand the state vector to take the un-modelled accelerations and accordingly update the STM and covariance during the estimation process.
Is this something that could be done in Python or would this require larger changes to the Java codebase?
Alternatively, it might be quicker to add options for different parametric acceleration modelsâ€¦

It is good to see that the GM processes have been added at the Hipparchus level.
Does this mean that the Orekit team is also thinking of developing in a similar direction?

I guess we could add something along these lines in the Orekit library at the Java level.
I am not sure yet if it should be another extension of the ParametricAcceleration base class or something entirely different.
Could you open a feature request in the forge with what you have in mind?

The Gauss-Markov process has indeed been added for the sake of space applications, but these are not used in the Orekit library itself, it is something I needed in my day work in an application suite that relies on Orekit.