Application and testing of the <i>L</i> ^* neural network with the self-consistent magnetic field model of RAM-SCB

We expanded our previous work on L* neural networks that used empirical magnetic field models as the underlying models by applying and extending our technique to drift shells calculated from a physics-based magnetic field model. While empirical magnetic field models represent an average, statistical magnetospheric state, the RAM-SCB model, a first-principles magnetically self-consistent code, computes magnetic fields based on fundamental equations of plasma physics. Unlike the previous L* neural networks that include McIlwain L and mirror point magnetic field as part of the inputs, the new L* neural network only requires solar wind conditions and the Dst index, allowing for an easier preparation of input parameters. This new neural network is compared against those previously trained networks and validated by the tracing method in the International Radiation Belt Environment Modeling (IRBEM) library. The accuracy of all L* neural networks with different underlying magnetic field models is evaluated by applying the electron phase space density (PSD)-matching technique derived from the Liouville\textquoterights theorem to the Van Allen Probes observations. Results indicate that the uncertainty in the predicted L* is statistically (75\%) below 0.7 with a median value mostly below 0.2 and the median absolute deviation around 0.15, regardless of the underlying magnetic field model. We found that such an uncertainty in the calculated L* value can shift the peak location of electron phase space density (PSD) profile by 0.2 RE radially but with its shape nearly preserved.