TipPFN and TipBox: Early tipping point detection using in-context learning

Climate tipping points emerge from nonlinear feedbacks that can trigger abrupt and potentially irreversible transitions in the Earth system, with far-reaching societal and environmental consequences. Anticipating such critical transitions in complex, high-dimensional systems remains a central challenge. While traditional early-warning indicators rely on assumptions of stationarity, long time series, and simple bifurcation structures. Because these assumptions rarely hold in real data, machine learning approaches have emerged as an alternative, but typically require training data from the specific systems under study, limiting their generalizability.

Here, we present TipBox and TipPFN. TipBox is an open-source, JAX-based repository containing a collection of simple dynamical systems and box models designed for accelerated generation of synthetic data. It enables efficient simulation of deterministic and stochastic systems exhibiting a wide range of bifurcation behaviour such as fold, Hopf, rate- and noise-induced tipping. Since TipBox is differentiable out-of-the-box, it enables easy parameter sensitivity tests for tipping point studies especially when different box models are coupled together.

Building on this synthetic data foundation, we develop TipPFN, a Prior-Data Fitted Network (PFN) approach based on a transformer machine learning architecture that performs approximate Bayesian inference via in-context learning. Trained on carefully selected synthetic dynamical systems, during inference it conditions on a short context of noisy observed data and produces a probabilistic forecast in a single forward pass based on synthetic priors generated from TipBox. This enables fast and computationally cheap probabilistic prediction on systems not seen during training, including time-to-tip as well as the type of tipping point. 

We validate our approach on three systems spanning different domains: an AMOC box model representing climate tipping elements, a predator-prey system from ecology, and a simplified power-grid model from infrastructure research. Preliminary results indicate that our PFN-based predictor generalizes to these complex test cases despite being trained exclusively on the simpler systems in TipBox. Benchmarking against state-of-the-art machine learning approaches shows promising results. We observe improved performance over traditional variance- and autocorrelation-based EWS, particularly under noisy conditions. Ongoing work evaluates conditional probabilistic predictions of the effects of changes in forcing on tipping dynamics.

Overall, we show that TipBox and TipPFN enable robust inference of tipping points on previously unseen systems with models trained purely on synthetic data without the need for additional retraining. This capability is especially powerful for the climate system where direct real-world observations of crucial tipping elements are unavailable but their prior proxies are.