level: research
climate scientists often struggle to tell how much of a warming trend comes from greenhouse gases and how much from natural wobbles. most methods need many climate model runs to average out the noise. a fresh approach called pullbackdmdc works on a single observed record. it treats external forcing, like rising co2, as a driver inside a linear stochastic system. the idea comes from pullback attractor theory, a branch of non-autonomous dynamical systems.
the technique builds on dynamic mode decomposition with control, a data-driven method that finds patterns in complex flows. by adding forcing as an input, the algorithm learns how the climate responds over time. it splits the observed signal into a forced part and an internal part without needing a big ensemble. tests on synthetic and real data show it can recover the forced response even when natural variability is strong. the method also handles non-stationary forcing, like the ramp-up of greenhouse gases.
this matters for climate projections and detection-attribution studies. single-realization methods often use linear inverse models that ignore forcing, or plain regression that skips dynamics. pullbackdmdc blends both, offering a physically grounded decomposition. it could help researchers extract the human fingerprint from regional temperature records, improve near-term predictions, and better assess climate model fidelity. the code is open and ready for use on paleoclimate or future scenario data.
why it matters: it gives a practical tool to isolate the forced climate signal from a single observational record, aiding attribution and projection without costly large ensembles.