Machine Learning Predicts Earthquakes with New Accuracy
Technical Analysis | 11-12-2025 | By Liam Critchley
Key Things to Know:
- Earthquake prediction remains a major challenge because seismogenic zones are hard to access and provide limited, noisy data, making it difficult to identify reliable precursors.
- Machine learning has already shown it can predict the timing of cm-scale laboratory quakes by analysing acoustic emissions and the evolution of shear stress in rock samples.
- New research demonstrates that ML can also predict metre-scale laboratory earthquakes, suggesting that, when scaled, similar approaches could offer early warning times from weeks to decades for large natural quakes.
- Real-world deployment will depend on better fault-zone data and models that capture off-fault seismicity, fault geometry and fluid pressure effects, but the work highlights ML as a promising tool for future earthquake early warning strategies.
Earthquake forecasting and prediction are still seismological challenges. This is because seismogenic zones are very inaccessible, leading to a lack of data that makes it difficult to predict earthquakes. It has been shown that it’s almost impossible to distinguish whether an earthquake will be large or small until the rupture begins, which means that there are limited precursor events in the Earth’s crust to predict earthquakes.
Machine learning (ML) has been applied to many fields and is now being used to provide advanced earthquake forecasting capabilities by predicting laboratory earthquakes in cm-scale rock friction experiments known as shear-slip failures. However, until recently, it was still debated as to whether ML could be applied to predicting larger-scale laboratory quakes and, eventually, natural earthquakes. Recent research has now used machine learning on the metre-scale, showcasing that ML can be used on larger-scale quakes, and provides some optimism that it could eventually be used for natural earthquakes.
Machine Learning Has Potential in Earthquake Prediction
ML has provided some optimism in the earthquake prediction field due to its ability to uncover hidden patterns in complex, multidimensional data. The advanced prediction capabilities can also help to ‘fill in the gaps’ when data is missing from datasets. ML can be applied to laboratory-scale experiments that simulate the natural faulting of rocks through rock shearing. In these experiments, ML has successfully been able to predict the timing of these quakes by analysing the precursory acoustic emissions.
These ML-enhanced experiments have shown that the evolution of nominal shear stress is associated with the preparatory phase of earthquakes, which is shown by an increased variance in the acoustic emission signal amplitude. Therefore, ML models that have been trained on ultrasonic pulse data can be used to predict both the evolution of shear stress in a fault and the time to failure of the fault that generates the earthquake.
However, even though the models have shown a lot of promise, the usage of these models has primarily been realised on cm-scale rock samples or sub-metre-scale gel models. The frictional behaviour of rocks at metre scales and above is different to cm-scale rocks as there are a lot more (and larger) forces that are exerted on the rocks. So, to fully understand natural earthquakes, larger-scale rock failures need to be predicted with high accuracy before being translated to natural scales. Beyond this, there have still been some questions about how ML predicts these earthquakes, especially for understanding the underlying physical processes and evolution of fault conditions.
New Machine Learning Approach Predicts Metre-Scale Earthquakes
Researchers have now advanced beyond the small cm-scale earth prediction experiments with ML and have successfully used ML for metre-scale laboratory quake tests and predicting the evolution of on-fault shear stresses. The researchers used a random forest algorithm with datasets from a large-scale, rock-friction laboratory experiment. By using tiny acoustic emission events as foreshock detection, the algorithm was able to accurately predict the time-to-failure of metre-scale mainshocks. The ML algorithm predicted the foreshocks from tens of seconds to milliseconds before the main earthquake occurred.
While this doesn’t sound like a long time before the main earthquake hits as an early warning system, the scales on which natural earthquakes occur mean that the relative time is much higher than is generated in laboratory experiments. Taking the tens of seconds to milliseconds range from the metre scale and translating it to the context of large-scale natural earthquakes, would correspond to pre-warning timescales from decades to weeks before a big earthquake. This means that machine learning could theoretically be applied for the prediction of natural large-scale earthquakes with some more work, as the faults were idealised in the experiments compared to natural fault lines.
The researchers compared the results of the model with a dynamic model of shear failures, and it was found that tracking the evolution of shear stress on creeping fault areas instead of nominal shear stress through acoustic emissions provides much more accurate predictions that are suitable for short-term forecasting.
By investigating both the laboratory and numerical results, the researchers were able to come up with a reason why the ML algorithm can predict laboratory earthquakes, and it is based on some patterns that precede an earthquake. Before a main earthquake hits, foreshocks occur sporadically. As the fault becomes more stressed to critical levels, there is a positive correlation between the nucleation of foreshocks and subsequent prolonged creep that accelerates towards a laboratory quake. This means that one of the nucleating ruptures is more likely to grow into a larger quake as the foreshock zone increases. ML can track these processes to predict if a major earthquake is likely to occur.
However, while the study shows promise, the researchers did state that the model doesn’t yet account for off-fault seismicity that occurs in natural quakes, which would add a lot more complexity to the prediction. It’s also thought that other natural occurrences, such as fault geometrical complexity and pore fluid pressure evolution, that were not present in the laboratory experiments, could have an impact on natural world seismic events. Finally, like any ML algorithm, the quality of the output relies heavily on the quality of the input the datasets. While not much data is available, the researchers have stated that transfer learning techniques could take synthetic data from experiments to train the ML model alongside the natural data to help cover areas of data sparsity and improve the prediction capabilities of the ML algorithm. More natural data from fault zones would also be highly useful for training the algorithms, especially data from slowly slipping fault regions such as local fault slip velocities, as that could help to better understand the predictability of large earthquakes.
Nevertheless, the research shows promise for scaling up ML algorithms as an early warning detection method for earthquakes and goes beyond current cm-scale capabilities. ML can’t predict the size of the earthquake before rupture begins, but offers a way to predict the timing of them, which could help with future earthquake prevention strategies.
Conclusion
This research demonstrates that machine learning can identify foreshock patterns and stress evolution processes that precede laboratory earthquakes on metre scales. By scaling these relationships to natural fault systems, ML-based prediction could potentially offer early warning times ranging from weeks to decades for large seismic events. Although more complex natural factors, data scarcity and off-fault seismicity remain challenges, the findings point to a promising path toward reliable short-term earthquake forecasting using ML powered acoustic and stress evolution analysis.
Reference:
