18-1. Application of machine learning to ground motion forecasting for earthquake early warning

The desire for earthquake early warning (EEW) systems to provide rapid forecasts of ground shaking has been growing around the world. Although in countries with operational EEW systems these forecasts are viewed favorably, questions remain about the timeliness and accuracy of the shaking forecasts (e.g., Minson et al., 2018; 2019). In particular, the accuracy of ground motion forecasts issued by EEW systems needs further improvement to facilitate decision making by end-users to optimize their benefits (e.g. Cochran et al., 2018, and references therein).

Machine learning, specifically that based on deep neural networks, has shown outstanding performance on image and speech recognition tasks, leading to major advances in self-driving vehicles, medical diagnostics, virtual reality and augmented reality applications, etc. It is expected that these methods may be able to capture complex physical phenomenon using functions that are fast to compute in real time, when large enough datasets are available for training. In seismology, machine learning techniques can offer significant advantages over more conventional methods for tasks such as signal versus noise discrimination as well as seismic phase identification and classification (e.g., Ross et al., 2018). The unique features of modern machine learning methods fit the needs of real-time ground motion forecasting for application to earthquake early warning, where the tradeoff between timeliness and prediction accuracy is essential in assessing the utility of EEW for a given application.

This research opportunity seeks a candidate to broadly consider how machine learning techniques can be applied to improve rapid ground motion forecasts for application to EEW systems. Most EEW algorithms determine source information such as origin time, location, magnitude, and sometimes finite fault extent; thus, they are essentially an ‘accelerated’ version of real-time traditional earthquake detection methods. Shaking estimates, to determine areas that should be alerted, are then forecast using a ground motion prediction equation (GMPE) to estimate average expected shaking intensities as a function of source distance and magnitude. All GMPEs have the same function form, and no other approaches currently exist to reliably forecast ground motion amplitudes. The ground motions estimated from source information have large uncertainties that translate into inaccuracies in specifying which regions should be alerted for a given ground shaking threshold.  

With the availability of large seismic archives and widely available machine learning software libraries there is now an unprecedented opportunity to apply deep learning techniques and develop a new approach to EEW ground motion forecasting. Forecasting ground motions rather than source information is the objective, so it may be advantageous to use rapid observations of near-source ground motions to directly forecast the expected shaking distribution across a potentially impacted region without first estimating source distance and magnitude. Those forecasts would then be updated along with the ground motion observations as the rupture evolves.  

Applicants to this research opportunity are encouraged to use machine learning to develop novel techniques for rapidly identifying regions that should be alerted for impending shaking, with accuracies greater than is attainable using existing EEW methods. Because this is an unexplored field, there are many potential avenues of groundbreaking research that the candidate could undertake, including using convolutional neural networks (CNNs) or other machine learning approaches to forecast earthquake shaking, exploring the time dependence of earthquake shaking (i.e., how is our expectation of future shaking updated by the current state of earthquake shaking), or quantifying uncertainties in shaking intensities for individual earthquakes.  The candidate could also explore forecasts for shaking attributes that are not currently available in EEW, such as shaking duration or frequency content.

Interested applicants are strongly encouraged to contact the Research Advisor(s) early in the application process to discuss project ideas.

Proposed Duty Station: Pasadena, CA or Moffett Field, CA

Areas of PhD: Geophysics, seismology, or related fields (candidates holding a Ph.D. in other disciplines, but with extensive knowledge and skills relevant to the Research Opportunity may be considered).

Qualifications: Applicants must meet the qualifications for: Research Geophysicist  

(This type of research is performed by those who have backgrounds for the occupations stated above.  However, other titles may be applicable depending on the applicant's background, education, and research proposal. The final classification of the position will be made by the Human Resources specialist.)

Human Resources Office Contact: Audrey Tsujita, 916-278-9395, atsujita@usgs.gov