Hi Vasco - sorry for the slow response! I don't think I have a particularly satisfactory answer for you on this topic, but I think my instinctive response is that the potential could be limited, depending on whether we think the patterns that underly the algorithms for forecasting now will be the same/ be able to be quickly adapted for these scenarios. My understanding is that forecasting depends on predicting the physics in the atmosphere (which may not change), as well as interactions with the land/oceans (which may be affected, e.g. changing landscapes due to death of flora/fauna). If this scenario is one that we haven't experienced before, maybe we could simply update the parameters based on new observations and the models would work, but maybe we would need to do research to improve the methods that are in use, which could take time. Another limitation could be that I'm not sure how well models could predict the evolution over several years, as you mention - our focus was on <6 months and we found limited accuracy.
Perhaps just one other thing to flag is that doing forecasting at somewhere like ECMWF uses a huge amount of energy, so we'd want to think about the trade-offs with other needs!
Hey Ben - thanks for your question. Unfortunately we didn't come across anything in our research that would quickly explain the decrease for 2004-2006 (or the increase in 2012-2016). It could have been differences in the atmospheric conditions at that time that the model was less able to handle, or changes to the ECMWF forecasting methods.
Re. machine learning, it does seem likely that there's scope for that here. While we didn't look into this in detail, we mention two ideas briefly in this spreadsheet of potential interventions.
Would be interested to hear if you have any further thoughts on either topic!