TREMR

Every few years, someone claims to have cracked it. A retired geophysicist with a proprietary algorithm. An app that tracks radon gas. A researcher who noticed that animals behave strangely before major quakes. Each time, the claim gets attention. Each time, it fails under scrutiny. Earthquake prediction — the ability to say where, when, and how large, with enough precision to be useful — remains one of the hardest unsolved problems in all of science.

That's not a failure of effort. Seismologists have been trying seriously since the 1960s. It's a failure of physics. The Earth's crust is a chaotic system, and faults are complex, heterogeneous, and partially hidden from observation. But understanding why prediction is so hard — and what we can actually do instead — matters for anyone who lives in an earthquake-prone region.

What Prediction Actually Means

The word "prediction" covers a wide range of claims, and precision matters enormously. A useful earthquake prediction must specify:

• Location — where the earthquake will occur, with enough geographic precision to be actionable
• Magnitude — how large it will be
• Time — when it will happen, with a narrow enough window to justify evacuation or other response

Probabilistic forecasting — saying that the southern San Andreas Fault has a 60% chance of rupturing in the next 30 years — is not prediction in this sense. It's hazard assessment. It's valuable for building codes and urban planning, but it doesn't tell anyone when to leave their house.

Why the Physics Is So Hard

Faults don't rupture when they reach a predictable stress threshold. They rupture when a small patch of rock somewhere along the fault loses friction and begins to slip — and whether that slip propagates into a minor tremor or a M8.0 megaquake depends on conditions that vary at scales too small to measure and too deep to observe directly.

This is the fundamental problem. A tiny earthquake and a devastating one begin identically: a small slip. The difference between them is determined by properties of the rock along the entire fault plane — how locked various segments are, how much stress has accumulated, where the fault geometry creates stress concentrations. We can't instrument the crust at the resolution required to know these things in advance.

Compounding this is the issue of chaotic amplification. Small differences in initial conditions lead to wildly different outcomes. A fault patch that might have produced a M5.0 in one stress state might produce a M7.5 under nearly identical-but-slightly-different conditions. The crust doesn't give us enough observable signal in advance to distinguish between these scenarios.

The Foreshock Problem

Many major earthquakes are preceded by smaller foreshocks — earthquakes that occur on the same fault in the days or hours before the main event. The 1975 Haicheng earthquake in China was successfully predicted partly on the basis of foreshock activity, and the evacuation of 90,000 people almost certainly saved thousands of lives. It remains one of the only successful operational earthquake predictions in history.

The problem is that most foreshocks are only recognisable as foreshocks after the main event. In real time, they look exactly like normal small earthquakes — because they are. Most small earthquakes are not followed by larger ones. If seismologists issued an evacuation order every time there was a cluster of small earthquakes near a fault, they would cry wolf thousands of times for every genuine warning. The social and economic cost would be enormous, and public trust in warnings would erode rapidly.

The Haicheng success was also followed, just a year later, by the Tangshan earthquake — a M7.8 event that killed at least 240,000 people with no warning whatsoever. It remains one of the deadliest earthquakes of the twentieth century.

Animal Behaviour and Other Popular Theories

Reports of unusual animal behaviour before earthquakes go back thousands of years — snakes emerging from hibernation, fish jumping, dogs howling. These accounts are culturally persistent and emotionally compelling. They are not scientifically reliable.

The problem is confirmation bias and the base rate of "unusual" animal behaviour. Animals behave strangely all the time, for reasons unrelated to seismicity. After an earthquake, people remember the strange behaviour they witnessed beforehand. They don't remember — and don't report — the dozens of times animals behaved strangely and no earthquake followed. Controlled studies have repeatedly failed to find a statistically significant correlation between animal behaviour and impending seismic events.

Similar issues apply to electromagnetic precursors, groundwater changes, radon gas emissions, and other proposed earthquake precursors. Some may correlate with seismic activity in specific conditions. None has been demonstrated to be reliable, specific, and consistent enough to serve as a basis for operational prediction.

What We Can Actually Do

The honest answer to "can we predict earthquakes?" is no — not in the sense of time, place, and magnitude with operational precision. But that doesn't mean we're helpless. There are three things science can deliver that are genuinely lifesaving:

• Probabilistic hazard assessment — identifying which faults are capable of large earthquakes and estimating the likelihood over decades. This drives building codes, insurance, and land-use planning that save lives passively, without any need for short-term prediction.
• Earthquake early warning — detecting the P-wave from an earthquake that has already started and sending alerts before the damaging S-waves arrive. Japan's system provides up to 90 seconds of warning. The US ShakeAlert system now covers California, Oregon, and Washington. This isn't prediction — the earthquake is already happening — but it's enough time to drop and cover, stop a train, or open a fire station door.
• Preparedness — understanding your risk, securing your space, knowing your evacuation routes, and practising what to do. No technology compensates for this. The countries with the lowest earthquake death tolls are the ones that have invested most heavily in building design and public education, not prediction research.

The Future of Prediction Research

Machine learning has revived interest in earthquake prediction over the past decade. Deep learning models trained on seismic catalogues have shown some ability to identify patterns associated with elevated seismic activity — but they remain far from operational prediction tools. The fundamental obstacle isn't computing power; it's the lack of observational data at the resolution the problem demands.

Some researchers focus on slow-slip events — episodes where fault segments creep silently over days or weeks, potentially loading adjacent locked segments toward failure. Others study seismic velocity changes in fault zones as stress accumulates. These are promising research directions, but the gap between laboratory results and operational prediction remains vast.

The most credible view among seismologists is that deterministic earthquake prediction — a specific time, place, and magnitude — may be fundamentally impossible, not just technically difficult. Faults may be intrinsically chaotic systems that cannot be predicted past a certain horizon, in the same way that weather cannot be forecast beyond about two weeks no matter how powerful the models become.

That's a humbling conclusion. But it has a practical upside: it refocuses attention on what actually works. Build better. Prepare more. Warn faster. Every dollar spent on earthquake-resistant construction saves more lives than any amount spent chasing prediction.