Can Machines Discover Relationships We Never Encoded?

DISCOVERY

A planetary dataset can contain weather, vegetation, water, movement, prices, infrastructure and human activity across decades. The number of possible relationships among those variables is far larger than any research community can inspect by forming one hypothesis at a time.

Machine learning can search that space cheaply. This sounds like automatic discovery until the first statistical problem arrives: with enough variables, the machine will discover relationships that are perfectly real in the dataset and completely unhelpful in the world.

The distinction matters because prediction and explanation reward different things. A variable can improve a forecast without causing the outcome. A model can exploit that variable productively for years and fail the moment the relationship changes. Discovery asks for more than usefulness. It asks whether the relationship survives an attempt to explain it.

THE FALSE-DISCOVERY PROBLEM

Earth is exceptionally good at producing convincing coincidences

Seasonality is the first trap. Temperature, vegetation, electricity demand, animal activity, river flow and many economic behaviours all change through the year. Two variables can follow each other beautifully because both have a calendar, not because either affects the other.

Spatial autocorrelation creates another. Nearby pixels, stations and communities are not independent samples. A map with a million cells does not necessarily contain a million independent pieces of evidence. If neighbouring observations share climate, soils, infrastructure and history, the effective sample size can be far smaller than the row count suggests.

Common drivers create a third trap. Two distant variables can move together because both respond to a large-scale climate mode. Lag makes the relationship look even more persuasive: one changes first, the other follows. Time order is necessary for causality and nowhere near sufficient.

Then there is simple arithmetic. Search enough candidate relationships and some will look strong by chance. Statistical corrections can control part of this multiple-testing problem, but planetary data adds structure that the textbook version does not: autocorrelation, hidden common causes, irregular sampling and changing regimes. More data does not abolish false discovery. It gives false discovery better infrastructure.

PCMCI

The Walker circulation is a useful test of what conditioning changes

A 2019 Nature Communications paper by Jakob Runge and colleagues showed the problem with a compact climate example. The researchers analysed pressure and temperature variables associated with the Walker circulation, the large east-west atmospheric circulation over the tropical Pacific.

Simple correlation and bivariate Granger-causality tests produced dense networks with many links. That is not surprising. Climate variables share persistence and common drivers, so almost everything can predict something else if examined pair by pair. The resulting graph was statistically busy and physically difficult to defend.

Their PCMCI method uses a two-stage strategy. It first selects candidate parents for each variable, then tests whether a proposed relationship remains after conditioning on relevant past variables. In the Walker-circulation example, the method recovered a much sparser network more consistent with the known physical structure.

That is the useful role of causal-discovery machinery. It does not turn observational data into an oracle. It asks whether an apparent link survives after obvious alternatives are accounted for. A machine can search thousands of possibilities. Conditioning is how it learns that most of them were introduced by the rest of the system.

HIDDEN CAUSES AND CHANGING RULES

Causal structure has assumptions too

Earth-system causality is difficult because hidden common causes are normal rather than exceptional. No observing network measures every relevant process. Atmospheric variables are coupled to oceans and land. Ecological variables respond to weather, competition and human activity. Economic outcomes respond to all of those plus institutions nobody has encoded as a continuous physical field.

Contemporaneous effects are another problem. Many processes interact faster than the sampling interval. If two variables are observed daily but influence each other within hours, time ordering at daily resolution cannot reliably decide which direction the interaction ran. The clock can be too coarse for the causal question.

Non-stationarity makes yesterday's graph provisional. Climate trends, land-use change, infrastructure and policy can alter relationships through time. A relationship that was stable for thirty years can weaken because a reservoir was built or irrigation expanded. Causal inference is not exempt from the planet changing underneath the dataset.

This is why physical knowledge remains useful even in data-driven discovery. Conservation laws, known time scales, impossible directions and domain constraints can eliminate candidate relationships before statistics are asked to judge them. The model searches. Physics helps decide which parts of the search space deserve respect.

THE DISCOVERY LOOP

The useful machine does not stop at finding a pattern

The realistic form of machine discovery is iterative. A model surfaces an unexpected relationship. Statistical tests remove obvious artefacts. Domain knowledge removes physically implausible ones. What remains becomes a hypothesis rather than a conclusion.

The next question is then unusually concrete: what evidence would distinguish the remaining explanations? A different variable may need to be measured. A location with contrasting conditions may be more informative. A transient event may provide a natural experiment. Discovery becomes a problem of choosing the next observation, not merely analysing the existing ones.

This is where planetary AI crosses an architectural boundary. So far the model has consumed whatever the observing system happened to provide. A system that can identify which missing evidence would reduce uncertainty can begin to influence the observing system itself.

That possibility is no longer theoretical. Satellites have already demonstrated onboard event detection and autonomous retasking. The next chapter follows the moment the analysis pipeline stopped waiting passively for data and began asking the instruments to look again.

Finding an unexplained relationship is not the end of discovery. It is the moment the system earns the right to ask a better question.