Continuous satellite measurements of the Earth's magnetic field have advanced the characterization of spatial‐temporal variations of the main field over the past two decades. To comprehend the underlying mechanism responsible for the geomagnetic field variations, we develop a novel core surface flow inversion scheme based on physics‐informed neural networks. The inversion method can account for the secular variation contributed by the interaction between the core flow and undetectable small‐scale magnetic fields. Based on the novel inversion framework, we derive a time‐dependent core surface flow model between 2000 and 2022 from the CHAOS‐7 core field model. The inverted core flow is then analyzed using the dynamic mode decomposition to extract wave‐like fluid motions. By calculating the magnetic secular acceleration contributed by each dynamic mode, we identify that the dynamic modes with period of about 10 and 7 years are responsible for geomagnetic jerks in the Atlantic and Pacific equatorial regions.