Anomaly Detection of Unmodeled Forces in Chaotic Three-Body Systems via Neural ODE Latent Representations

Detecting unmodeled or 'hidden' forces in chaotic dynamical systems is a fundamental challenge in physics, particularly for the three-body problem where sensitivity to initial conditions makes long-term prediction intractable. We present a proof-of-concept AI framework based on a Neural ODE Autoencoder that learns compact latent representations of three-body trajectories and uses them to detect the presence of five categories of hidden perturbations. On a balanced held-out validation set, our model achieves an F1-score of 0.74 and AUROC of 0.69, outperforming three baselines: a physics-based energy drift detector (AUROC 0.61), a flat MLP classifier (AUROC 0.51), and a standard autoencoder without the Neural ODE component (AUROC 0.45). The gap between the Neural ODE model and the standard autoencoder demonstrates that continuous latent dynamics modelled by the ODE contribute meaningfully to anomaly detection beyond reconstruction alone. All three learned latent dimensions exhibit statistically significant differences between clean and perturbed trajectories (p < 0.05), suggesting that the latent space encodes physically interpretable structure. We discuss limitations of the current simulation-based evaluation and outline directions for extending the framework to real observational data. Code available at: https://github.com/AkshaanAngral/Three-Body-Causal-AI