Interaction without an interaction module

The prevailing way to model a many-body system is to encode every component and then let a module capture interaction. We show that the explicit interaction module is harmful when the number of components increases. We use a simple plastic encoder that compresses a component’s interaction contribution into its representation, and combine the representations by linear aggregation. This modelling approach is simple in form yet delivers unexpectedly excellent performance: it matches the prevailing explicit route under most conditions and exceeds it when a mixture carries more components than the training set covered. Moreover, an encoder produced this way has a clear advantage over per-component descriptors when it is transferred to a similar task: it greatly reduces the data requirements.