Facilitating hydrocarbon exploration from earth system models

A method comprises access training data 106 of a modern feature of interest from direct observations, remotely determined data, or a combination thereof. The modern feature may be chlorophyll concentration in seawater or total organic carbon in sediments. Parameter data is compiled from at least one model simulation 108 (e.g. hydrosphere, atmospheric, heliosphere, geosphere, cryosphere, biosphere model) that impacts the modern feature of interest. A machine-learning model is trained to generate a predictive model 104 that matches the training data of the modern feature of interest using the compiled parameter data as input. A feature of interest is predicted 112 in a past time period using the predictive model and at least one historical model simulation that impacts the feature of interest. Hydrocarbon exploration may be facilitated based on the predicted feature of interest from the predictive model.