Predicting Wildfires with CAT Scans
CAT scanners are an imaging technique normally used in medicine to peer inside a patient’s body - helping in diagnosing them. Despite their traditional use in medicine, due to their ability to penetrate a material, they are also useful for studying the interior of wood – as established by Stanford engineers. The team of engineers have used the CAT scanners to investigate smouldering in depth which has allowed them to develop computer models capable of predicting where a wildfire may start next. The obvious benefits of this work include helping firefighters allocate their resources more efficiently, consequently reducing the loss of lives and properties due to wildfires.
Smouldering is the process of burning slowly where smoke is present without a flame, and the cause of consumption of more than half of plant matter which ends up burnt in wildfires [1]. Moreover, embers produced by smouldering logs can ignite fires in unsuspected areas when carried by hot updrafts, accelerating the spread of the fires, and have the potential to restart fires as smouldering can take place several days after flaming [2].
XCT (X-ray Computer Tomography) was used by Emeric Boigné to develop a new analytical method to study smouldering. Using this method provides a three-dimensional image of the material’s interior structure at a resolution of 135 μm [3, 4].
Through studying wood samples from around the country, the researchers aimed to compile a database of a range of materials and their burn rates. Collecting this data then allowed Matt Bonanni (Stanford graduate) to use it alongside land contour data (as well as other relevant data such as the air humidity and wind patterns) to build a model for predicting fires. The simulation runs through many permutations of different variables to predict how a fire is most likely to spread.
“Smoldering depends strongly on the heat exchange with the surrounding environment, ambient oxygen concentration, permeability, and density, as well as the content of moisture and inorganic material in the fuel" [4].
These simulations cover forest-level scales, therefore running them requires a great amount of processing power. For this reason, Tensor Flow, an open-source framework developed for use in machine learning, was adopted which vastly improved the performance of the model as many simulations were able to be completed simultaneously [3].
As the model was developed, simulations that took a day to complete before can now be run in close to real-time [3]. Improving the model to be able to predict the spread of fires in such a short time means that it can become fit for practical purposes and we can start to experience the benefits it brings to our society as it aids firefighters and influences their firefighting strategies.
Modelling Tsunamis with Supercomputers
Fugaku has been used to develop an AI computer model for predicting tsunami flooding. Being able to make such accurate predictions in real time is greatly beneficial in helping disaster management. It will aid residents to be evacuated in the most effective way as well as planning the effects on infrastructure [5].
Normally tsunami predictions have been made by comparing the tsunami's earthquake occurrence conditions with the data of the most similar observations from a compiled database of simulations run in advance [7]. Therefore, predicting the effects of a tsunami in this way relies makes it difficult to implement a feasible prediction system as it incorporates supercomputers and database searches on a large scale.
Anticipated flooding by Cabinet Office Japan compared to new AI predictions [7]
By putting the supercomputer’s power to use, the research team in Fujitsu, Tohoku University, and the University of Tokyo's Earthquake Research Institute - allowing to simulate 20,000 scenarios for training the AI model [6] which would be much more feasible to run.
Waveform data and data on the resulting flooding was used by a deep learning algorithm. This training allowed the team to produce an AI model capable of predicting tsunami flooding with high accuracy in near real-time. While the AI has to be trained on the supercomputer, the model can be run on a PC, allowing observational waveform data to be input into the model to produce predictions on the flooding that will take place [6].
References:
[1] G. Rein. SFPR Handbook of Fire Protection Engineering fifth ed.: Springer, 2016, pp. 581-603. Cited in: E. Boigné, N.R. Bennett, A. Wang, K. Mohri, M. Ihme. Simultaneous in-situ measurements of gas temperature and pyrolysis of biomass smoldering via X-ray computed tomography. Proceedings of the Combustion Institute 2020; 38: 3899 – 3907
[2] E.R.C Rabelo, C.A.G. Veras, J.A. Carvalho Jr., E.C. Alvarado, D.V. Sandberg, J.C. Santos. Log smoldering after an amazonian deforestation fire. Atmospheric Environment 2004; 38: 203 - 211. Cited in: E. Boigné, N.R. Bennett, A. Wang, K. Mohri, M. Ihme. Simultaneous in-situ measurements of gas temperature and pyrolysis of biomass smoldering via X-ray computed tomography. Proceedings of the Combustion Institute 2020; 38: 3899 – 3907
[3] A. Myres. Stanford researchers combine CAT scans and advanced computing t fight wildfires. Stanford Report, 2020. 29/10/2021
[4] E. Boigné, N.R. Bennett, A. Wang, K. Mohri, M. Ihme. Simultaneous in-situ measurements of gas temperature and pyrolysis of biomass smoldering via X-ray computed tomography. Proceedings of the Combustion Institute 2020; 38: 3899 – 3907
[5] L. Magloff. World’s Fastest Supercomputer Used to Model Tsunamis. Springwise, 2021. 03/11/2021
[6] N. Lavars. Supercomputer-developed AI predicts tsunami flooding in real-time on a PC. New Atlas, 2021. 03/11/2021
[7] International Research Institute of Disaster Science, Tohoku University, Earthquake Research Institute, The University of Tokyo, Fujitsu Laboratories Ltd. Fujitsu Leverages World’s Fastest Supercomputer ‘Fugaku’ and AI to Deliver Real-Time Tsunami Prediction in Joint Project. Fujitsu Limited, 2021. 03/11/2021