Fig. 6: Prediction accuracy of DL models for R-tipping under out-of-sample dynamical systems.

a–c, DL models were trained on time series with a specific forcing rate of ϵ = 1.25 in the saddle-node system, and subsequently used to predict R-tipping cases with previously unseen forcing rates in the saddle-node system (a), Bautin system (b) and compost-bomb system (c). The prediction accuracy is presented as a function of forcing rate and forecast lead time. d–f, DL models trained on data with a forcing rate r = 0.1 in the Bautin system were used to predict R-tipping cases with previously unseen forcing rates in the Saddle-node system (d), Bautin system (e) and Compost-bomb system (f), respectively. g–i, DL models trained on data with a forcing rate v = 0.1 in the compost-bomb system were used to predict R-tipping cases with previously unseen forcing rates in the Saddle-node system (g), Bautin system (h) and Compost-bomb system (i), respectively. j–l, For comparison, DL models were trained on a combined dataset that integrated equal proportions of the above three training datasets, and subsequently used to predict R-tipping cases with previously unseen forcing rates in the Saddle-node system (j), Bautin system (k) and Compost-bomb system (l), respectively.