Pulling the climate switch

In October 1947, the citizens of Savannah watched in horror as a once inconsequential hurricane took a hairpin turn in the middle of the Atlantic. Chosen for its secluded trajectory and mild wind-speeds, Hurricane King was one of the first, high-profile attempts by scientists to play god with the weather. The result was an evacuated city with over $3 million in damages ($32 million today). At the time, many blamed Irving Langmuir's inauspicious experiment for the events that followed; however, further inspection by meteorologists raised other concerns.

It turns out that hurricanes tend to change directions all the time; their intensities have been known to wax and wane much like that of Hurricane King. So, what impact did Langmuir's pellets of dry ice really have? To answer this, we would have had to know where the hurricane would end up without the experiment and here lies the heart of the problem in geoengineering: the climate is not repeatable, so we can never truly compare scenarios. Similar projects over the years ran into similar headwinds. In Operation Popeye, for example, did the US military actually increase the level of rainfall on guerrilla fighters in Vietnam or would it have rained just as hard anyways?

Looking back, many question marks remain on the efficacy of such experiments. Looking forward, into a world of geoengineering, we lack the certainty that our interventions will have the results we expect them to achieve.

Tweaking the environment is very much like pulling the switch in the infamous trolley problem (as a side note: I particularly enjoyed flicking through trolley problem memes) — diverting the train to save some lives will have an adverse impact on (albeit fewer) others. Much like the trolley problem, geoengineering would have a massive accountability dilemma. Any positive action that still leads to calamities will come under heavy scrutiny, so each new disaster would easily be blamed on any prior meddling. And if the global superpowers (likely to be the US or China) invest most in the technology to build their own switches, they may prefer to only point the trolley in ways that they deem beneficial. Fortunately - or unfortunately - there isn't just one such switch. The climate version of the trolley problem is far more convoluted, with far more switches, people and directions involved.

A more 'accurate' depiction of the trolley problem when applied to geoengineering

Why we don’t know what’s going on

The reason why we struggle to forecast the weather was first uncovered by Lorenz after he pasted in rounded values into his climate model. When it finished computing, he noticed that the results differed wildly even for very small deviations in the inputs. This idea gave rise to 'chaos theory' which sat in direct contention with the Newtonian principle of cause and effect. The world is not a predictable sequence of events but instead it is one in which many complex systems interplay. Lorenz was the first to provocatively pose the question "does the flap of a butterfly’s wings in Brazil set off a tornado in Texas?”. Although slightly misleading, there are two important conclusions from his work for geoengineers:

1. Our measurements across the planet are both imprecise and incomplete

2. The delta between our measurements and reality can result in extremely divergent outcomes (such as an unaccounted butterfly triggering a hurricane)

Building a global thermostat

While I write this, I stare out of my window to watch the downpour of rain. I keep refreshing the weather forecast and it takes a few minutes to catch up, changing from sunshine to the more dire reality. One of the unexpected consequences of the recent pandemic has been a (slight) decline in weather forecast quality. Meteorologists take in billions of data-points, using everything from satellite imagery to remote weather stations to sensors on commercial planes, from around the world to build their forecasts. As hoards of flights have been grounded across the world, some of these data-points were lost and thus the forecasts became ever so slightly more imprecise. However, even in perfect conditions, there still remains enough whitespace such that forecasting even two weeks is a stretch.

In the world of the Internet of Things (IoT), there has been much hope for a world full of omnipresent sensors. The theory is that sensors will continue to become both smaller and cheaper, while 5g (or any subsequent technologies) would continue to become faster. These two factors create 'intelligent' specs of pixie dust that record every little detail and upload it to the 'cloud', ready for meteorologists to pump it into their ever growing models. Then, knowing what to do with this data will become the hard part.

'Learning' the weather

Companies like OpenAI and DeepMind have turned to reinforcement learning to solve problems where traditional cause and effect models have struggled. Instead of applying rules, reinforcement learning basically works by trying a bunch of actions and seeing which ones give it the most 'reward'. In a simple game like pong, this could look like randomly pressing the up & down arrow keys until the AI's score increases. Then over time, it learns when it should press which arrow key to maximise it's chances of achieving a high score.

This paper suggests that we could follow the same approach with Stratospheric Aerosol Injection (I touched upon this in my last article). However, instead of allowing the AI to play with arrow keys, it controls how much and where we emit reflective aerosols to cool the Earth. The input data then changes from pixels on the screen to all the meteorological sensor data we have across the world. The reward, however, is not as straight forward to set.

In a video game, the reward is often quite clear: the score is found at the top of the screen. In their model, the researchers were trying to minimise the number of extreme shifts in weather at a local level. I.e. the AI fails when the desert becomes a swampland or India doesn't get a monsoon. Unfortunately, if Bangladesh got a flood this year, the model will work towards ensuring that it gets another flood next year and so on for as long as it can be controlled.

Rather than just preserving the status quo, AI-controlled weather becomes exciting when we start to explore other reward functions. For example, it could potentially be tasked with maximising global GDP or improve global health or even increase happiness survey results. Imagine a controlled climate that can tailor the environment to maximise farm yields based on the produce of that region.

The road to a climate utopia

Unfortunately, many risks still remain. Even if we are able to capture all the data we need, there's no guarantee that pulling the switch will actually change the direction of the metaphorical trolley. If past experiments have shown anything, it's that complex systems respond in complex ways. This doesn't mean that there isn't any value in trying, but the outcome is likely to be less linear and often less pronounced than would have been expected. And, if effectual, it's unclear that these could ever be controlled with the granularity needed to optimise regional weather.

Furthermore, the core issue with reinforcement learning is the number of 'replays' required. For example, it took OpenAI half a billion attempts before it could solve hide and seek, and we just have one chance to fix one planet. The work-around is to use high fidelity simulations that can be run over and over again; however, it's proven very difficult to mimic the diversity of the real world in a computer, even for more basic tasks such as manipulating a rubiks cube. New modes of failure always pop up once the AI moves into the real world.

Until we build the prerequisite technology to measure, control and learn the climate, geoengineering will remain an untested band-aid that we can (at best) hope would buy us some more time and not leave too much destruction in its wake.