|
|
| Fig. 1: Visualization of the penetrative capabilities of three types of ionizing radiation: α and β particles are stopped by paper and aluminum, respectively, while γ particles are only attenuated by lead. (Source: Wikimedia Commons) |
Theoretically, nuclear energy has significant benefits in terms of power output and CO2 emissions. However, this energy source has the potential to be extremely damaging when things go awry. The risks associated with nuclear energy are made infamous by specific catastrophic events at power plants in places like Chernobyl and Fukushima.
When discussing how to improve nuclear energy from a public safety perspective, most of the focus is on prevention. It's definitely important to determine potential failure points of power plants and high-risk events like natural disasters and figure out ways we can engineer structures to be resilient to these factors. However, it's impossible to guarantee 100% safety. Therefore, we also need to improve how we can respond to nuclear disaster.
One of the biggest concerns is the material released from a nuclear meltdown. Radioactive material can emit several types of radiation, including α particles, β particles and γ rays. Each type is more energetic and penetrative than the former, illustrated by Fig. 1. Small doses of exposure can cause minor tissue damage while larger doses can cause severe damage to major organs, leading to death. [1] In the aftermath of the Fukushima nuclear meltdown, radioactive material spread across a wide range via wind and water currents. Predicting the dispersion of radioactive material is a crucial task to give accurate evacuation warnings and minimize the number of people exposed to radiation. However, predictions made about the dispersion of radioactive material in the immediate aftermath of Fukushima were not utilized due to their unreliability. This naive response is a main source of motivation for Yoshikane et al. in developing machine learning models that can more accurately predict the short-term dispersion of radioactive material. [2]
 |
| Fig. 2: Example SVM visualization for binary classification in two dimensions. w is the normal (perpendicular) vector to the hyperplane found by the model with a magnitude of ||w||. The paramter b/||w|| is the amount that the hyperplane is offset from the origin. The margins are defined as w · x - b = 1 (any point on this margin belongs to class 1) and w · x - b = -1 (any point on this margin belongs to class -1. The maximum-margin hyperplane itself is defined as w · x - b = 0, in between both class margins. The overall distance between both margins is 2/||w||. (Source: Wikimedia Commons) |
The near-surface winds were identified as the most significant weather-phenomenon that affects the dispersion of radioactive material. So, Yoshikane et al. used two types of near-surface wind data as input data to the machine learning model. [2] The input data consisted of near-surface wind fields and area-averaged near-surface wind fields in the form of weather maps generated from Grid Point Value data provided by the Japan Meteorological Agency. To generate ground truth dispersion directions, the researchers hindcast simulations on data from March 11 - 31, 2011. Long-term simulations were conducted on data from January, March, April, July and October from 2009 to 2013. The dataset generated includes the simulated dispersion direction ground truth for given near-surface wind fields at intervals of 3 hours. [2]
The goal of the machine learning model is to classify the dispersion direction of radioactive material from data on near-surface wind fields. The particular model used was a support-vector machine (SVM). SVM is a supervised learning method commonly used for classification tasks, which involves using decision planes to calculate optimal decision boundaries. Fig. 2 shows an example two-dimensional SVM model for binary classification. Here we have 2 classes, 1 (blue) and -1 (green). Each point on the graph is a training example that has 2 vectors (x1, x2) of real numbers, and a label 1 or -1 indicated by color. We want to find a hyperplane that maximizes the distance from the nearest points in both classes. In this case, the hyperplane is the red line in the graph that separates the two classes of data. The margins are the dotted lines that run through the points closest to the red line, and points that lie on the dotted line are called support vectors. Once the SVM has been fed the training examples and finds the "maximum-margin" hyperlane (i.e. the red line that maximizes the distance from the margins of each class), it can be given a completely new example and predict what class it should be in based on which side of the red line the point lies. SVMs are favored for their ability to produce significant accuracy with less computational demand relative to other statistical models. [3]
The trained SVM model was tested on a held out set of data from months of years it was not trained on. The average accuracy obtained on unseen data was 85%. When evaluated on unseen data from March 2011 (when the Fukushima meltdown occurred) the accuracy rate was 93%. [2]
The researchers also evaluated the model in regard to lead time, or how far out the model can accurately predict dispersion direction. They found the model was able to predict dispersion direction for lead times up to 33 hours in advance with an accuracy exceeding 0.77 in all months. [2]
 |
| Fig. 3: International Atomic Energy Agency experts at the Fukushima nuclear power plant after the disaster. (Source: Wikimedia Commons) |
The significance of this research cannot be understated. Yoshikane et al. have shown that near-surface wind field data can be used with machine learning models to predict dispersion direction with significant accuracy. [2] Previous modeling attempts proved to be too unreliable and common sense doesn't carry the statistical power necessary when making decisions critical to saving lives. Additionally, predictions can be made with reliable accuracy up to 33 hours in advance. This information can be tremendously useful in future emergency situations, where knowing how radioactive material is likely to disperse can help prioritize which regions to evacuate. While it's still important to assess and respond to the longstanding aftermath of nuclear disaster, as the IAEA experts pictured in Fig. 3 are at Fukushima, existing residents in the surrounding area likely had little warning to evacuate before being exposed to radioactive material in the first place. Governments and emergency responders can be armed with information that helps them more effectively protect the population earlier, while civilians are no longer kept guessing on whether to evacuate or not.
The disaster relief industry is trending towards increased utilization of machine learning to help save lives. For example, One Concern, a startup in Palo Alto, CA has deployed machine learning systems capable of making accurate predictions about how earthquakes and floods can affect major cities. [4] The research performed by Yoshikane et al. is a promising step towards predicting the effects of non-natural disasters and has clear potential to save lives in the face of nuclear meltdown. [2]
However, I would caution to fully embrace this research outright. The model was trained and tested on ground truth data generated from simulations. Ideal next steps would be to collect real-world measurements on dispersion patterns (of non-radioactive material) and near-surface wind fields to further experiment with.
© Cameron Cruz. The author warrants that the work is the author's own and that Stanford University provided no input other than typesetting and referencing guidelines. The author grants permission to copy, distribute and display this work in unaltered form, with attribution to the author, for noncommercial purposes only. All other rights, including commercial rights, are reserved to the author.
[1] Adverse Reproductive Outcomes in Families of Atomic Veterans: The Feasibility of Epidemiologic Studies (National Academies Press, 1995), Ch. 4.
[2] T. Yoshikane and K. Yoshimura, "Dispersion Characteristics of Radioactive Materials Estimated by Wind Patterns," Sci. Rep. 8, 9926 (2018).
[3] J. Gareth et al., An Introduction to Statistical Learning: with Applications in R (Springer, 2013), Section 9.
[4] J. Palca, "Betting On Artificial Intelligence To Guide Earthquake Response," NPR, 20 Apr 18.
.jpg){kind=link}