AI turns planetary forecasts into equations scientists can read

Researchers have converted an AI’s planetary-instability predictions into algebraic expressions, using a dataset of 113,543 simulations of three-planet systems. Their 26 October preprint examines when planets approach closely enough to destabilise their orbits.[1]

The network first learns combinations of orbital properties. Symbolic regression then searches for formulae reproducing its predictions. The inputs still require a short simulation covering 10,000 orbits of the innermost planet.[1]

On a separate three-planet grid, for systems unstable before one billion inner-planet orbits, the equations’ root-mean-square error was 1.52 dex, against 1.74 for the network. This measures error in logarithmic instability time: one dex corresponds to roughly a factor-of-ten error.[1]

The readable formulae expose sensitivity to orbital resonances, where periods form simple ratios. Some predictions exaggerate unstable bands; the formulae remain fitted approximations.[1]

The authors’ public repository offers a small predictor requiring NumPy, the REBOUND orbital simulator and Matplotlib, alongside the complete training workflow. That gives researchers a route to examining the equations without first rebuilding the neural network.[2]

The reproduction guide separates plotting stored results from rerunning experiments. It estimates about 15 hours for the main training and distillation pipeline: seven for training and pruning the network, then eight for extracting equations. Its minimal examples accept either prepared test cases or a planetary simulation supplied by the researcher.[2]

Report an error or suggest a correction