Handling variable numbers of features per data point

In my field (particle physics), we process data from particle collisions that create new particles. It’s a natural consequence that the data we may want to feed into a neural network has a ‘jagged’ structure, where one entry has the data from two particles, and another may have the data from 5 particles, etc., which will have a different number of features due to there being more particles.

I’m not aware of a natural way to handle this kind of data beyond imposing some kind of reduce() paradigm. Is there any literature or nice way of handling this where there could be a variable number of features for each input?

LSTM could work well here?