Global Crustal Heat Flow Using Random Decision Forest Prediction

Abstract:

We have applied supervised learning with random decision forests (RDF) to estimate, or predict, a global, densely spaced grid of crustal heat flow. The results are similar to global heat flow predictions that have been previously published but are more accurate and offer higher resolution. The training inputs are measurement values and uncertainties of existing sparsely sampled, (~8,000 locations), geographically biased, yet globally extensive, datasets of crustal heat flow. The RDF estimate is a highly non-linear empirical relationship between crustal heat flow and dozens of other parameters (attributes) that we have densely sampled, global, estimates of (e.g., crustal age, water depth, crustal thickness, seismic sound speed, seafloor temperature, sediment thickness, and sediment grain type). Synthetic attributes were key to obtaining good results using the RDF method. We created synthetic attributes by applying physical intuition and statistical analyses to the fundamental attributes. Statistics include median, kurtosis, and dozens of other functions, all calculated at every node and averaged over a variety of ranges from 5 to 500km. Other synthetic attributes are simply plausible, (e.g., distance from volcanoes, seafloor porosity, mean grain size). More than 600 densely sampled attributes are used in our prediction, and for each we estimated their relative importance. The important attributes included all those expected from geophysics, (e.g., inverse square root of age, gradient of depth, crustal thickness, crustal density, sediment thickness, distance from trenches), and some unexpected but plausible attributes, (e.g., seafloor temperature), but none that were unphysical. The simplicity of the RDF technique may also be of great interest beyond the discipline of crustal heat flow as it allows for more geologically intelligent predictions, decreasing the effect of sampling bias, and improving predictions in regions with little or no data, while rigorously estimating the uncertainty of the result.