El Niño forecasts extended to 18 months with physics-based model
phys.org
phys.org
[1] Farmers frustrated after destocking following BOM's incorrect El Niño forecast: https://www.abc.net.au/news/rural/2024-01-22/farmers-frustra...
Predicting the past isn't the same thing as predicting the future. You can train a model or compose an algorithm that fits existing data but that doesn't necessarily mean it has predictive power far into the future.
In times well past there was a certain degree of divination that looked at accumulated patterns and made guesses that worked more often than not in the short term (less than a year in advance.
Then came "black box" non physics based methods:
Recent advancements in artificial intelligence (AI) have pushed these boundaries, achieving accurate predictions up to 16–18 months in advance. However, the "black box" nature of AI models has precluded attribution of this accuracy to specific physical processes.
Not being able to explain the source of the predictability in the AI models results in low confidence that these predictions will be successful for future events as the Earth continues to warm, changing the currents in the oceans and atmosphere.
"Unlike the 'black box' nature of AI models, our XRO model offers a transparent view into the mechanisms of the equatorial Pacific recharge-discharge physics and its interactions with other climate patterns outside of tropical Pacific," explained Fei-Fei Jin, the corresponding author and professor of atmospheric sciences in SOEST.
"The initial states of the extratropical Pacific, tropical Indian Ocean, and Atlantic enhance ENSO predictability in distinct seasons. For the first time, we are able to robustly quantify their impact on ENSO predictability, thus deepening our knowledge of ENSO physics and its sources of predictability."
Not every horseshoe map is folded to infinity and beyond. Creeks flow predictably enough and yet still contain turbulent vortexes.
Dynamic systems aren't an all or nothing proposition wrt chaos, there are often contained embeddings an portions that are difficult to predict but not impossible subject to error bounds.
This isn't how this works, at all.
We are talking about large scale systems on long timelines that have massive momentum and energy. Predicting something 18 months out isn't hard (Florida will have a hurricane in the next 18 months).
Predicting "will crops get enough water in the mid west in any given year" is impossible because of "chaos" because of complexity.
This result is not surprising to any one who pays attention to how the math for these systems works.
A long-standing challenge has been for these models to exhibit reliable skill at lead-times beyond ~6-12 months, especially when forecasts are initialized in the early Spring. Data-driven models ("AI" or "machine learning") seemed to break through this barrier a bit, but what's notable about this paper is that a physical model with a new underlying mechanism seems to catch up a bit.