Predicting local magnetic perturbations has been one of the major goals of space weather modeling for decades. The traditional approach is to solve the MHD equations in a large 3D domain driven by the solar wind inputs measured at L1 point. This method provides a prediction window of 30 to 90 minutes, which is the time it takes the solar wind to propagate from L1 to Earth for solar wind speeds varying from 900 to 300 km/s. These first-principles models, including the Michigan Geospace model running at the Space Weather Prediction Center of the National Oceanic and Atmospheric Administration (SWPC/NOAA) have been improved over the years by adding more sophisticated physics, improved parameterization and improved numerical methods. The predictive skill of the Geospace model measured by the Heidke Skill Score (HSS) improved from about 0.54-0.66 (depending on dB thresholds)  to 0.57-0.7 range in about 7 years. In 2024 we have started developing a machine learning (ML) model driven by the same inputs plus prior storm time disturbance (Dst) index observations. The new GeoDGP (Geomagnetic Deep Gaussian Process) model outperforms the first-principles model significantly: it can improve HSS values by about 0.2, and it can provide an extra hour of forecast window without much loss in skill. This dramatic improvement can be attributed to the flexibility of ML and the physical intuition used to formulate the problem in a way that ML can optimally use. In fact, previous ML models using a more straightforward parameterization were underperforming both the GeoDGP and the Geospace models. I will show how combining physical insight played a crucial role in the success of GeoDGP. The new ML model also provides the basis for a new aurora forecast app, AuroraTonight.