Failure in capturing details of solution-structure using m3gnet model

[mm21d013]

The m3gnet model seems to work well for materials with periodicity in their arrangement. The moment, solutions are created by placing interstitial atoms at specific positions lets say octahedral positions, the m3gnet model seems to fail. The model doesn't seem to capture the interaction of interstitial atom with remaining atoms even though the distance between any adjacent N-interstitial is less than 5 angstrom (Two body cut off).

Please correct me if I am wrong...
The possible reason behind this might be when using a 3-body interaction model (where n=3), the value of n−2=1(referenced from the m3gnet paper "A universal graph deep learning interatomic potential for the periodic table" , Figure attached alongwith) means we only consider one additional neighbor along with the central atom (i) and its immediate neighbor (j). This restricts the model's ability to account for more complex local environments, such as those involving interstitial atoms or multiple neighboring atoms that significantly affect the local chemistry or physics.
In such cases, a 4 body interaction or an higher order interaction might be helpful. I tried everything with m3gnet, but that possibily works only for compounds and not solutions. Even the pretrained potential has been trained with materials project which barely has any solution.

I would be grateful if someone do approve or want to comment something on the above.