What if we could see the climate in 2150?

Most climate models we use to imagine the future stop at 2100. It is like turning off the lights just when the party gets interesting. Yet, to truly grasp risks like irreversible ice melt or shifting ocean currents, we must look further. These are 'tipping points', thresholds that change everything for centuries.

Until now, running Integrated Assessment Models (IAMs) to 2150 was too slow and costly. Now, researchers have found a smart shortcut: an emulator. Imagine an autopilot learning to fly by watching a human pilot: the emulator 'learns' the behavior of complex models and then runs fast on its own, extending emission and temperature scenarios for another 50 years.

This framework does not replace original models but offers a valuable interim solution for international projects like ScenarioMIP, where many possible pathways need comparing. Thus, scientists and policymakers can test the consequences of today's choices on a much broader horizon. It is not a certain prediction, but a tool to rigorously explore 'what ifs', helping us avoid sailing blind into the long term.

The study (arXiv:2512.06026v1) addresses a structural limit of standard Integrated Assessment Models (IAMs): the fixed 2100 time horizon. IAMs couple economic, energy, and earth system models to generate emission pathways consistent with climate targets (e.g., SSP-RCP). However, their computational complexity makes simulating additional centuries prohibitive for every scenario, especially to explore parametric uncertainties or slow carbon cycle feedbacks (permafrost, methane hydrates) and tipping elements like AMOC or ice sheets.

The proposed approach uses a statistical emulator (often Gaussian Process or neural network based) trained on an ensemble of full IAM runs to 2100. The emulator learns the mapping from socio-economic drivers (population, GDP, energy intensity) to climate outputs (radiative forcing, temperature, concentrations). Once trained, it can rapidly project trajectories to 2150 for new inputs or SSP pathway extensions, assuming learned functional relationships remain valid (stationarity assumption).

This method provides an 'interim product' for ScenarioMIP (part of CMIP7), allowing the Earth System Model (ESM) community to have consistent post-2100 forcings without waiting for the next generation of long-term IAMs. Limitations include dependence on training data quality and inability to capture structural regime shifts in economy or technology not present in the training set.