A Variance-Reduction Framework for Practical Event-Based Monte Carlo PSHA

Earthquakes are a matter of chance, and predicting their impact on a specific location requires navigating a vast landscape of uncertainty. Scientists use a method called probabilistic seismic hazard analysis to estimate how often the ground might shake with enough force to damage buildings. This process involves imagining thousands of possible earthquakes, each with different sizes, locations, and depths, and then calculating how the shaking would travel to a specific site. Because the ground does not shake in a perfectly predictable way even for identical earthquakes, the models must also account for random variations in the shaking intensity. Traditionally, to get a reliable answer, researchers have relied on running massive computer simulations that generate millions of these random scenarios, hoping that the sheer volume of data will smooth out the noise and reveal the true risk.

A new approach developed by Rajesh Rupakhety at the University of Iceland challenges the way these random variations are handled. The study suggests that for calculating the overall risk of shaking at a single site, it is unnecessary to randomly generate a specific shaking intensity for every single imagined earthquake. Instead, the researcher found that by using a mathematical technique to calculate the average shaking probability for each scenario directly, the results become far more stable and accurate. This method, which the paper calls a variance-reduction framework, effectively removes a layer of random noise that has been built into these simulations for decades. The result is a way to produce hazard maps and safety guidelines that are smoother, more reliable, and require significantly less computer time to generate, particularly when looking at rare, high-magnitude events that are critical for engineering design.

For decades, the standard way to estimate earthquake risk has been to simulate a long timeline of history, filling it with thousands of random earthquakes. In these simulations, every time an earthquake is generated, the computer also picks a random number to represent how much the ground shakes. This random number accounts for the fact that two earthquakes of the same size and distance can produce different levels of shaking. The computer then checks if that specific, randomly generated shake is strong enough to exceed a safety threshold. If it is, it counts as a hit; if not, it counts as a miss. By repeating this process millions of times, the computer builds a picture of how often the threshold is likely to be crossed. This method works well when the risk is high and many earthquakes exceed the limit, but it struggles when looking at very rare, dangerous events. In those cases, the random nature of the shaking means that a simulation might accidentally miss a dangerous event entirely, or count one that is unlikely, leading to jagged, unreliable results that look like a staircase rather than a smooth curve.

Rupakhety's work identifies that this randomness is the culprit. He realized that for the specific goal of calculating the overall probability of shaking, the exact random value for each earthquake is not actually needed. Instead of picking a random number and seeing if it crosses the line, the computer can simply calculate the exact probability that the shaking would cross the line for that specific earthquake scenario. This is a subtle but powerful shift. It is like trying to guess the average height of a crowd. One way is to measure every single person and average the numbers. Another way, which is what this new method does, is to look at the group's known characteristics and calculate the average directly without measuring every individual. By doing this calculation for every simulated earthquake, the method removes the "noise" caused by the random selection of shaking values. The paper demonstrates that this approach, known as Rao-Blackwellization, produces a much smoother and more accurate hazard curve, especially for the long return periods that engineers use to design critical infrastructure.

The researcher tested this idea using two different approaches. First, he used a simplified, controlled model to see how the two methods compared. In these tests, the traditional method produced hazard curves that jumped up and down erratically as the return period increased, sometimes stopping completely because no random earthquakes happened to exceed the threshold in that specific run. The new method, however, produced a smooth, continuous line that stayed close to the true answer every time. The study showed that the traditional method's errors were not due to a lack of data, but rather the unnecessary introduction of random noise. By integrating out the random shaking component, the new method reduced the uncertainty in the results by a factor of twenty-six in one specific test case, and by a factor of nearly nine hundred when combined with another advanced sampling technique.

To ensure this was not just a theoretical exercise, the study applied the method to a realistic model of earthquake sources in South Iceland. This region is complex, with multiple fault lines and different types of tectonic activity. The researchers simulated the hazard for a specific rock site in the area, comparing the traditional random-sampling approach with the new conditional-probability approach. The results mirrored the controlled tests. The traditional simulations produced hazard curves that varied wildly from one run to the next, making it difficult to trust the numbers for rare events. The new method produced consistent, stable curves that matched a high-precision reference solution. Furthermore, the study looked at how these methods perform when calculating the "uniform hazard spectrum," which describes the shaking intensity across different vibration frequencies. The traditional method often produced jagged, broken spectra that were hard to use for engineering, while the new method preserved a smooth, logical shape across all frequencies.

The paper also explored how this new way of thinking changes the way scientists break down the sources of risk, a process called deaggregation. In the traditional method, scientists look at which specific earthquakes actually caused the shaking to exceed the limit in their random simulation. This can be misleading if the simulation happened to miss a few key events. The new method assigns a fractional weight to every single simulated earthquake based on its probability of causing a hazard, rather than just counting the ones that happened to cross the line. This provides a much clearer and more stable picture of which earthquake scenarios are actually driving the risk. In the South Iceland example, this approach reduced the uncertainty in identifying the most dangerous earthquake magnitudes and distances by a significant margin, giving engineers a more reliable basis for design.

Beyond just improving accuracy, the study demonstrates that this method is also much faster. Because the new approach removes a major source of error, it requires far fewer simulated earthquakes to reach a reliable result. In one test, the researchers found that to achieve a specific level of precision for a rare, ten-thousand-year event, the traditional method needed to simulate over one hundred million years of earthquake history. The new method, when combined with adaptive sampling techniques that focus computer power on the most dangerous scenarios, achieved the same precision with only one hundred thousand years of simulation. This represents a reduction in computer time by a factor of over two hundred, turning a calculation that might take minutes into one that takes less than a second.

The implications of this work extend beyond just running simulations faster. The paper argues that the way we treat randomness in these models should depend on what we are trying to measure. For calculating the overall risk at a single site, the random variations in shaking are a nuisance that can be calculated away. However, the study notes that in more complex situations, such as calculating the risk for a whole city where the shaking at different locations is linked, some of that randomness must be kept to preserve the relationship between sites. The new framework allows scientists to choose exactly which parts of the randomness to keep and which to integrate out, offering a flexible tool for future hazard analysis.

Ultimately, this research provides a unifying perspective on how we calculate earthquake risk. It shows that the classical methods used for decades and the modern simulation-based approaches are not fundamentally different in what they are trying to calculate, but rather in how they handle the random elements of the problem. By recognizing that some random variables can be solved analytically rather than sampled, the study offers a path toward more efficient, stable, and accurate seismic hazard assessments. The findings suggest that the future of earthquake risk modeling lies not just in generating more data, but in generating smarter data, using mathematical insight to strip away unnecessary noise and focus computational effort on the variables that truly matter.