I was reading this post on Reddit here thinking about building a predictive model for an extensive system. I know it is widely known that a neural network might be a better choice to apply the knowledge learned from a small system to large systems. But I was wondering if we can use kernel regression modules（KRR, and GP）to predict properties of large systems by learning small systems?
One of the ideas I have is that for the prediction of extensive property, we can build multiple modules, and the predicted property is the sum of the outputs from the modules. This method requires collective training of the model. I believe there must be a better idea than this.