Albeit the choice of variables differs from that of the paper for the non-greedy four feature selector, it is not worth mentioning:
The R2-score of the two models barely differs at all (see multivariate_polynomial_fit_v2.ipynb):
rh, clw, clw, T: 0.5545711625803007
rh, dzRH, cli, T: 0.5511722237713514