Neural Network Plus Orbital Physics Predicts Satellite Falls to Within About 13 Minutes
A BRIN-led team tested the method on a Soviet Venus lander, a NASA telescope and a Chinese rocket stage, with all three reported fall times landing inside the predicted window.
On May 10, 2025, a spacecraft built in the Soviet Union to land on Venus finally came home. Cosmos-482, launched in 1972, had been stranded in orbit by a failed upper stage and circled Earth for 53 years. Roscosmos later reported that it came down in the Indian Ocean, west of Jakarta. A new study in Acta Astronautica shows how a neural network working alongside Newtonian orbital mechanics could pin down such falls far more tightly.
The problem is growing. By the end of 2025, low Earth orbit held 25,060 cataloged objects, nearly twice as many as a decade earlier, according to the European Space Agency. Many old satellites and spent rocket stages will fall back without control and could come down anywhere. Timing is everything: an object near the end of its life moves at about 7.8 kilometers per second, so an error of one minute shifts the fall location by about 460 kilometers along its path, and an error of one hour shifts it by about 28,000 kilometers.
The work was led by M. Khawariz Andaristiyan of the Research Center for Satellite Technology at Indonesia's National Research and Innovation Agency (BRIN), with Affiani Machmudah and Wahyudi Hasbi of the same center and Ridanto Eko Poetro of the Institut Teknologi Bandung. Denys Dutykh described the study.
The key variable is drag from the thin upper atmosphere, which swells when the sun is active and brakes satellites faster. How strongly the air slows an object is captured by its ballistic coefficient, which depends on its shape, area and mass. That coefficient is hard to pin down for a tumbling rocket stage or a satellite with wide solar panels, and fresh tracking data become scarce just when braking accelerates. A common shortcut holds the coefficient fixed.
The team divided the work. Newton's laws, Earth's gravity and flattening, and a standard atmosphere model driven by solar and geomagnetic activity handle the orbit. A neural network with a memory learns how each object's coefficient will change, using 11 years of public tracking data, a full solar cycle. Every three hours the network receives the new orbit and returns an updated coefficient. The calculation stops at 80 kilometers, the height taken as the threshold of reentry. Running the network 100 times with random parts switched off gives a window of uncertainty.
The authors tested it after the fact on three objects. For Cosmos-482 they predicted 06:22 UTC, give or take 11.4 minutes, against a reported 06:24. For NASA's NEOWISE telescope they predicted 00:54 UTC on Nov. 2, 2024, give or take 7.8 minutes, against a reported 00:49. For a Chinese Long March 4B upper stage launched in 2006, they predicted 22:45 UTC on March 26, 2023, give or take 13.2 minutes; its debris was seen over southeastern Spain between 22:33 and 22:36. In every case the reported time fell inside the window.
Freezing the coefficient showed what the learning adds. For compact Cosmos-482 the predicted fall moved only five minutes. For NEOWISE and the rocket stage it came 31 and 36 minutes later, outside the windows and farther from the reported times.
These were hindcasts, made after the fact from tracking data collected up to shortly before each fall. The real test will be forecasts made ahead of objects that have not yet fallen, which is where minutes of accuracy translate into hundreds of kilometers of ground track.
