Computational Volcanology: How AI and Machine Learning Are Changing Eruption Forecasting in 2026

Volcanology used to be a science of boots and notebooks. Today it is a science of sensor arrays, satellite passes, and inference pipelines. A single well-instrumented volcano now produces more data in a week than a field campaign of the 1980s produced in a decade, and no human analyst can read it in real time. That shift has quietly turned volcano monitoring into one of the most interesting applied machine learning problems in the earth sciences: high-frequency multimodal time series, extreme class imbalance, weak labels, and decisions where a false negative can cost lives.

This guide explains what computational volcanology actually involves in 2026, which techniques are being deployed, what the engineering stack looks like, and where the hard problems still sit.

Why Volcanology Became a Data Problem

There are roughly 1,500 potentially active volcanoes on Earth, and around 800 million people live close enough to one to be affected by an eruption. Only a fraction of those volcanoes are instrumented well enough to give meaningful warning. The ones that are instrumented, however, are monitored continuously and densely.

A modern volcano observatory operates seismometers sampling at 100 Hz or higher across a network of stations, continuous GNSS receivers logging position, satellite interferometry acquisitions every few days, thermal and gas sensors, infrasound arrays, and webcams. All of it runs twenty-four hours a day, whether the volcano is restless or asleep.

The consequence is a familiar engineering situation. Data acquisition has outpaced data interpretation. Observatories in Italy, Japan, Iceland, Indonesia, and the United States all face the same bottleneck: continuous streams arriving faster than trained analysts can review them, with the genuinely interesting signals buried in months of ordinary background noise.

That is precisely the shape of problem machine learning is good at.

The Four Primary Data Streams

Data Characteristics That Shape the Modelling Problem

Stream	Sampling	Volume	Primary challenge
Seismic	100–200 Hz continuous	Terabytes per year per network	Class imbalance, event overlap, station noise
InSAR	Every 6–12 days	Gigabytes per scene	Atmospheric artefacts, decorrelation on vegetation
GNSS	1 Hz to daily solutions	Modest	Reference frame drift, seasonal signals
Gas	Intermittent, campaign-based	Small	Sparse, weather-dependent, hard to automate
Thermal satellite	Multiple passes daily	Moderate	Cloud cover, low spatial resolution

Notice the asymmetry. Seismic data is abundant and continuous but noisy and hard to label. Gas data is diagnostically powerful but sparse. Fusing streams with radically different sampling rates and reliability is one of the central technical problems in the field.

Machine Learning for Seismic Signal Classification

This is the most mature application of AI in volcanology, and the one with the clearest operational payoff.

The Problem

Volcano seismicity comes in distinct families. Volcano-tectonic events look like small brittle-failure earthquakes with sharp onsets and broadband frequency content. Long-period events have emergent onsets and narrow low-frequency peaks, generally attributed to fluid movement in cracks and conduits. Hybrid events show both characters. Volcanic tremor is sustained, sometimes lasting hours or days, and its appearance or intensification often precedes eruptions.

Traditionally an analyst reviews recordings and classifies events by hand. On a restless volcano producing thousands of events per day, this is impossible to sustain, and the backlog means the analysis that matters most arrives too late to be operationally useful.

How It Is Being Solved

The dominant approaches convert waveform segments into time-frequency representations, typically spectrograms or scalograms, and treat classification as an image problem. Convolutional networks trained on catalogued events from a given volcano reach high accuracy on that volcano’s event families.

More recent work has moved toward:

• Self-supervised pretraining on unlabelled continuous data, followed by fine-tuning on small labelled catalogues, which addresses the chronic label scarcity problem
• Transformer architectures applied directly to raw waveforms, borrowing from developments in tectonic earthquake phase picking
• Unsupervised clustering to discover event families without a predefined taxonomy, which is valuable at poorly studied volcanoes where no labelled catalogue exists
• Transfer learning across volcanoes, which remains difficult because each volcano’s signals are shaped by its own structure and plumbing

The Honest Limitation

A model trained on Mount Etna does not transfer cleanly to Campi Flegrei. Volcanic seismic signatures are strongly site-dependent, shaped by conduit geometry, edifice structure, and near-surface velocity. Generalisation across volcanoes is an open research problem, not a solved one. Most operational systems today are volcano-specific models retrained on local catalogues.

Deep Learning on Ground Deformation

Ground deformation is the second major front, and satellite interferometry has made it a genuinely global dataset.

Why InSAR Matters

Interferometric Synthetic Aperture Radar measures surface displacement by comparing the phase of radar returns from repeat satellite passes. It can resolve centimetre and even millimetre-scale motion over an entire volcanic edifice, including volcanoes with no ground instrumentation at all. With Sentinel-1 providing regular free acquisitions, systematic monitoring of hundreds of volcanoes became possible for the first time.

The catch is that most interferograms contain no volcanic signal, and many contain atmospheric artefacts that look superficially like deformation. Water vapour variation in the troposphere produces phase delays that mimic ground motion. Finding real deformation means separating a rare true signal from a common convincing false one.

The Machine Learning Approach

Convolutional networks trained to classify interferograms as deforming or non-deforming have been applied to large archives, flagging candidate scenes for expert review. The training data problem is severe, because genuine deforming interferograms are rare, so synthetic deformation patterns generated from analytical source models are commonly injected into real atmospheric noise to build balanced training sets.

Beyond binary detection, active research directions include:

• Estimating source parameters directly from interferograms, effectively learning an inverse model
• Time-series analysis across long interferogram stacks to detect slow accelerating trends
• Combining InSAR with GNSS in joint models that exploit the complementary spatial and temporal resolution of each

Campi Flegrei as a Case Study

The caldera west of Naples is one of the most closely watched volcanic systems in the world, and it illustrates why deformation monitoring matters. The caldera undergoes bradyseism, slow ground uplift and subsidence associated with pressure changes in the shallow hydrothermal and magmatic system. Uplift episodes have displaced the ground by metres over decades, accompanied by seismic swarms.

Because roughly half a million people live within the caldera, the interpretation of every centimetre of uplift has direct civil protection consequences. This is a setting where the difference between a model that produces a calibrated probability and one that produces an unexplained score is not academic.

Probabilistic Eruption Forecasting

Detection tells you what is happening. Forecasting attempts to say what happens next, and it is a fundamentally probabilistic problem.

Bayesian Event Trees

The dominant operational framework is the Bayesian event tree. The structure decomposes the forecasting question into a sequence of conditional nodes:

1. Is there unrest?
2. Is the unrest caused by magma?
3. Will magma reach the surface?
4. If it erupts, where will the vent open?
5. What size will the eruption be?
6. What hazardous phenomena will result?

At each node, prior probabilities derived from the volcano’s geological history are updated with current monitoring data. The output is a probability distribution over outcomes, with uncertainty explicitly propagated, rather than a single deterministic prediction.

This framework is used operationally in Italy and elsewhere. Its strength is that it produces calibrated, auditable probabilities that civil protection authorities can act on. Its weakness is that node probabilities often rest on sparse historical records and expert elicitation, both of which carry substantial uncertainty.

Where Machine Learning Fits

Machine learning does not replace this framework. It feeds it. Automated event classification supplies the monitoring parameters that update the tree in near real time. Anomaly detection on multivariate streams flags state changes that trigger reassessment. Pattern recognition across analogue volcanoes helps constrain priors where local history is thin.

The most defensible architectures being built now are hybrid: machine learning for detection and feature extraction, physics-informed and Bayesian models for forecasting and uncertainty quantification.

Physics-Based Simulation and Surrogate Models

Hazard assessment requires simulating what an eruption would do. Established codes model ash dispersal in the atmosphere, pyroclastic density current propagation, lava flow paths, and lahar routing over digital elevation models.

These simulations are computationally expensive. A single ash dispersal run under one wind scenario can take substantial compute, and probabilistic hazard assessment requires thousands of runs sampling across eruption sizes, vent locations, and meteorological conditions.

This is where surrogate modelling has become valuable. A neural network trained on a large ensemble of physics simulations can approximate the simulator at a fraction of the runtime, enabling:

• Real-time hazard maps updated as wind fields change
• Rapid what-if analysis during an unfolding crisis
• Monte Carlo sampling over parameter spaces that would be infeasible with the full simulator

The engineering discipline here matters. A surrogate that is accurate in the bulk of the distribution but wrong in the tails is dangerous, because hazard assessment is precisely about the tails.

The Technical Stack

For engineers considering this field, the working toolset is recognisable.

Layer	Typical tools
Seismic data handling	ObsPy, SeisComP, SAC
Data access	FDSN web services, IRIS, EIDA nodes
InSAR processing	ISCE2, SNAP, GMTSAR, LiCSBAS
Geospatial	GDAL, rasterio, GeoPandas, QGIS
Machine learning	PyTorch, scikit-learn, TensorFlow
Numerical and array	NumPy, SciPy, xarray, Dask
Physical modelling	FALL3D, Titan2D, VolcFlow
Visualisation	Matplotlib, PyGMT, Plotly
Infrastructure	Docker, Kubernetes, cloud object storage

The pipeline discipline is standard MLOps applied to a scientific domain: continuous ingestion, versioned datasets, reproducible preprocessing, model registries, drift monitoring, and alerting. What differs is the evaluation philosophy. In a support chatbot, a false positive is an annoyance. In a volcano monitoring system, a false positive erodes the credibility that makes evacuation orders effective, and a false negative can be fatal.

The Hard Problems

The interpretability problem deserves particular emphasis. When an observatory raises an alert level, the decision is reviewed by scientists, communicated to authorities, and potentially used to justify evacuating a population. A model that cannot explain which features drove its output is difficult to use in that chain of responsibility, regardless of its accuracy on a held-out test set. This is one domain where the pressure toward interpretable and physics-informed models is not academic preference but operational necessity.

How to Enter the Field

The demand is real, and the supply of people who can work on both sides is thin. Most volcanologists are not machine learning engineers, and most machine learning engineers know nothing about volcanic processes. The intersection is small and growing.

Coming from software or AI: the mathematics transfers directly. Signal processing, time series modelling, Bayesian inference, and geospatial data handling are the core skills, and they are the same skills whether the sensor is measuring seismic velocity or something else. What you need to acquire is the domain physics: how magma moves, why long-period events occur, what deformation implies about a pressure source. Formal study is the efficient route, and specialised programmes exist. The Master of Science in Volcanology at the University of Naples Federico II, taught in English in partnership with INGV and the Osservatorio Vesuviano, includes statistics and mathematics for geosciences, geophysical exploration, GIS and remote sensing, and volcano hazard and risk assessment, alongside the classical volcanology.

Coming from geosciences: the fastest gains come from Python fluency, then ObsPy and the standard geophysical stack, then statistics and machine learning fundamentals in that order. Resist the temptation to start with deep learning; most operational value in this field still comes from careful signal processing and well-specified statistical models.

For everyone: the data is largely open. Seismic archives are publicly accessible through FDSN services, Sentinel-1 imagery is free, and several volcano observatories publish catalogues. There is nothing stopping you from building something real this month.

Summary

Computational volcanology sits at an unusual intersection: a domain with genuinely open scientific questions, freely available high-volume data, direct humanitarian stakes, and a shortage of people equipped to work across the boundary between geophysics and machine learning.

The maturity varies by application. Seismic event classification is operational and reliable within a single volcano. InSAR deformation detection is working at archive scale but still fighting atmospheric artefacts and label scarcity. Probabilistic eruption forecasting remains fundamentally a Bayesian and expert-driven framework that machine learning feeds rather than replaces. Surrogate modelling of physical simulations is promising and increasingly deployed.

What has not changed is the responsibility attached to the output. Every model in this chain eventually informs a decision about whether people should leave their homes. That constraint shapes everything: it favours calibrated probabilities over point predictions, interpretable models over opaque ones, and conservative deployment over rapid iteration.

For an engineer who wants their work to matter in a measurable way, there are worse places to spend a career.