A move 37 for the ocean

What AI could do for marine Carbon Dioxide Removal

If you saw the 2016 documentary AlphaGo, you remember Move 37. DeepMind’s fantastically sophisticated AI algorithm is playing against Lee Sedol, South Korea’s top Go player. The game looks “normal” until the machine does something that looks completely insane. You can hear the collective intake of breath from the professional commentators, the initial confusion after Move 37 —that can’t be right!— followed by slow, dawning recognition that something unprecedented is happening. Move 37 isn’t just bizarre, it’s superhuman. It turns the whole game. Lee Sedol eventually resigns.

Move 37 obsesses a certain breed of techie because it shows AI coming up with an idea no human being would even have considered. An idea so obviously “bad”, it merits no serious interest from a professional.

Go is hugely complex. On each turn, a player has hundreds of legal moves to choose from, and the opponent has hundreds of possible moves in reply. The decision trees that result are insane: just three moves out, there are hundreds of millions of possible board configurations. No one can hold that many possibilities in their mind at once, so human champions adapt by focusing their attention on just a handful of candidate moves that, after years of practice, they’ve learned to intuit as good. The vast bulk of possible moves, they discard without thinking about them.

This is how all human expertise works. To be an expert means to have a gut level feel for what’s likely to work. It means developing sophisticated pattern recognition that instantly eliminates the vast majority of possibilities, leaving a manageable handful open to consideration.

These pruning heuristics are what make human expertise possible.

But they also make certain solutions invisible.

AlphaGo beat Lee Sedol because it pruned the game tree differently. Not because it searched the entire thing — not even a computer can do that. Its neural networks, trained on millions of games and refined through self-play, had developed pattern recognition from different foundations. It hadn’t inherited human heuristics about shape and territory and influence. It learned directly from outcomes: positions that lead to winning.

When ocean biogeochemists think about marine Carbon Dioxide Removal, they face a pruning problem that makes Go look like child’s play.

Consider what’s actually involved in evaluating a potential ocean carbon removal intervention. Three-dimensional circulation patterns that vary seasonally and interannually. Biogeochemical cycles—carbon, nitrogen, phosphorus, iron, silica, others—each with their own dynamics and limitations. Microbial community responses. Temperature and pH effects on biological processes. Viruses. Potential downstream impacts on fisheries, on oxygen minimum zones, on atmospheric feedback loops.

The combinatorial space of possible interventions—location × timing × method × scale × duration × combination strategies—is utterly overwhelming.

Human ocean scientists have developed powerful pruning heuristics to navigate it. They focus on interventions they can measure exploiting mechanisms that fit their understanding of existing biogeochemical models on scales that match realistic funding constraints and available logistics. They mind their professional reputations, so they propose approaches that will seem “reasonable” to peer reviewers who’ve spent careers studying these systems.

These heuristics are earned. Built from decades of observations, failed experiments, hard-won theoretical understanding. They’re mostly correct. That’s why they persist, why they’re taught to graduate students, why they shape which proposals get funded.

“Mostly correct” means they also create systematic blind spots.

At Anthropocene, we support an approach to marine carbon dioxide removal that exploits a mechanism most research overlooks: extending the availability of bioavailable nitrogen in oligotrophic ocean systems by mimicking natural fertilization events. The specifics involve decision trees that make Go look like tic tac toe. How much of which nutrient do you consider, where exactly, in what form exactly, when exactly?

We have heuristics to answer these questions, sure, but they’re relatively blunt.

What would a Move 37 for nitrogen fixation for phytoplankton carbon look like? Impossible for humans to say. Maybe there’s some spatial patterns of intervention that exploits circulation in ways that seem geographically nonsensical until you see the global response. Maybe it’s a timing strategy that coordinate with seasonal dynamics in ways humans wouldn’t think to orchestrate, or some exotic cascade effect where small interventions in “unimportant” locations trigger disproportionate responses elsewhere, or some nutrient that bottlenecks a whole system that’s not on our radar. Who knows.

The best human Go player couldn’t have seen Move 37. The best human oceanographers probably can’t see the optimal approach to mCDR.

This is where physics-grounded AI becomes interesting.

I’m not talking about large language models that confabulate plausible-sounding nonsense, or pure machine learning approaches that need massive training datasets and can only interpolate within their training distribution.

I mean a world-model AI of the ocean. Imagine a neural networks training on synthetic data from GPU-accelerated ocean general circulation models with full biogeochemistry, constrained by conservation laws, grounded in thermodynamics. Training the neural network on physics model outputs could yield an emulator that runs orders of magnitude faster while still respecting thermodynamic constraints, making it possible to search intervention space at scale.

I’m increasingly convinced this is what mCDR research will look like in the next decade.

An Ocean AI system still wouldn’t be able to search every possibility: it still has to prune the decision-tree. But it eliminates possibilities for different reasons than human experts do.

The system would simulate millions of intervention scenarios, evaluating each against physical constraints—energy balance, nutrient budgets, circulation dynamics—without inheriting human preconceptions about what’s “worth trying.” It won’t know that certain approaches are “obviously wrong” by expert consensus. It would only know what the physics and the biogeochemistry allows.

The interventions that survive AI pruning but fail human pruning—that’s where Move 37 lives for marine carbon dioxide removal.

Building a GPU-accelerated ocean model with sufficient resolution and biogeochemical complexity is technically challenging but conceptually straightforward. The computational infrastructure exists, the physics is just a wall of Navier-Stokes equations—Newtonian physics in fluid dynamics terms. It’s a fully solvable engineering problem.

The hard part is what comes after.

When the model proposes an intervention that human experts have learned to prune away, then what do we do? Move 37 looked like a mistake to every professional Go player watching in real time. Several thought AlphaGo had just malfunctioned.

The ocean’s Move 37 will trigger the same response from marine scientists. At first.

This creates a bootstrapping problem. A physics-based AI system for ocean carbon removal needs to build credibility before anyone implements its most counterintuitive suggestions. It might start by proposing relatively conservative interventions—things that look plausible to human experts but that no one had specifically thought to try. Demonstrate that those work on the basis of physical measurements and you can gradually expand the envelope of what seems reasonable to test.

Of course, if the system only ever proposes things that feel safe to human intuition, we’ve wasted our time. We’d build it precisely because human pruning has blind spots. The credibility-building phase can’t become a permanent constraint. At some point, we need institutional capacity to take seriously—and actually test—interventions that violate expert heuristics but respect physics.

Thing is, this isn’t a problem that gets solved with better AI or more computational power. It’s institutional and cultural. We need funding mechanisms that can evaluate proposals outside the pre-pruned solution space. Peer review processes that can distinguish “this violates what we expect” from “this violates what we know.” Experimental designs that can test counterintuitive proposals from physics-based models while maintaining appropriate safeguards.

None of this means dismissing human expertise. Ocean scientists have built extraordinary understanding of these systems through patient observation and theory-building. That knowledge is what constrains the physics models, what validates their outputs, what identifies when simulations are producing artifacts rather than insights.

But expertise creates structure in how we search for solutions. And that structure has gaps.

Human-led incremental research is systematically exploring a pre-pruned solution space. That’s fine for building fundamental understanding of ocean systems—possibly optimal, even. It may not be good enough for finding gigaton-scale carbon removal solutions on climate-relevant timescales.

We’re at an inflection point where computational power has finally caught up to the complexity of the problem. GPU acceleration makes it possible to produce high-resolution synthetic training data from global ocean models at speeds that enable the resulting emulator to systematically explore the intervention space. The physics is well enough understood to constrain these models meaningfully, we’re not short of observational data to validate their behavior in unperturbed conditions.

What we don’t yet have is institutional infrastructure to take seriously what these models might reveal.

Physics-based AI for ocean carbon removal would serve multiple functions. High-precision verification and monitoring of interventions—tracking carbon flux changes, attributing them to specific actions, quantifying permanence and downstream effects. This capability alone would be hugely valuable, helping distinguish effective interventions from failed ones and building the empirical foundation for scaling what works.

But the real value proposition is discovery. Building a platform that can show us interventions we’ve learned not to see, by pruning possibility space according to physics rather than human intuition.

And then—this is the hard part—building the institutional capacity to actually try what that different pruning reveals.

The solutions we need for gigaton-scale carbon removal are almost surely sitting in the 95% of possibilities that expert pruning has learned to eliminate. Not because ocean scientists are wrong about what works. They’re probably mostly right.

Being “mostly” right means systematically missing the exceptions.

We’re still designing research programs for a world where human intuition is our only guide. That world ended with AlphaGo. The climate timeline doesn’t give us time to exhaustively search the 5% of intervention space that feels safe to human experts.

We need to search the other 95%—systematically, guided by thermodynamics rather than familiarity. The computational tools exist. The institutional capacity to trust them does not.

I wouldn’t be surprised if building that capacity turns out to be harder than building the models.

It’s also more urgent.