摘要：Models of magnetospheric ultralow‐frequency (ULF) waves can be used to study wave phenomena and to calculate the effect of these waves on the energization and transport of radiation belt electrons. We present a decision tree ensemble (i.e., a random forest) model of ground‐based ULF wave power spectral density driven by solar wind speed v s w , north‐south component of the interplanetary magnetic field B z , and variance of proton number density var( N p ) . This model corresponds to four radial locations in the magnetosphere (roughly L ∼ 4.21 to 7.94) and spans 1–15 mHz, with hourly magnetic local time resolution. The decision tree ensembles are easier to use than the previous model generation; they have better coverage, perform better at predicting power, and have reduced error due to intelligently chosen bins in parameter space. We outline the difficulties in extracting physics from parameterized models and demonstrate a hypothesis testing framework to iteratively explore finer driving processes. We confirm a regime change for ULF driving about B z = 0 . We posit that ULF wave power directly driven by magnetopause perturbations corresponds to a latitude‐dependent dawn‐dusk asymmetry, which we see with increasing speed. Model uncertainty identifies where the underlying physics is not fully captured; we find that power due to substorms is less well characterized by B z > 0 , with an effect that is seen globally and not just near midnight. The largest uncertainty is seen for the smallest amounts of solar wind driving, suggesting that internal magnetospheric properties may play a significant role in ULF wave occurrence. Plain Language Abstract To describe the energy and motion of electrons in the radiation belts for space weather modeling purposes, we need to be able to describe the large‐scale ultralow‐frequency plasma waves. These waves can energize and transport high‐energy electrons, which pose a hazard to spacecraft. Using ground‐based magnetometers whose observations directly correspond to waves in space, we construct a machine learning model that predicts power in those waves based on incoming solar wind properties. This performs better than the previous model and is more easily used to investigate the physics. We demonstrate how to use it to explore the physics by comparing it to existing wave phenomena. We find that the solar wind drives it in the ways expected but that we should not average over these processes as some of them have a significant on/off switch. We find that two aspects of magnetospheric physics are not fully captured in our model and contribute to the remaining uncertainty: the effect of substorms throughout the magnetosphere and the internal state of the magnetosphere. These are likely to have a significant additional impact on wave occurrence.
