ATD: Resilient Dynamic Autoencoders for Modeling and Predicting Earthquake Threats

大地震产生强烈的地面运动和海啸，可能导致大量人员伤亡，并对包括美国西海岸在内的地震活跃地区的社会复原力造成严重影响。已经开发了早期预警系统，通过检测地震震中附近的第一批到达的地面运动并预测强烈破坏性地面运动的强度和时间来减轻直接威胁。为了进一步提高这些系统的有效性和准确性，深度学习方法具有很强的潜力，但显著扩展现有模型的预测视野是至关重要的。地震波传播的高度异质性给机器学习带来了根本性的挑战。该项目旨在通过开发弹性和可靠的深度学习方法来预测噪声和复杂的时空地面运动数据来应对这些挑战。由此产生的方法和开源软件工具将应用于地震学以外的领域，使包括地球、大气和气候科学在内的各种科学和工程领域受益。此外，该项目为本科生和研究生提供研究培训机会。该项目将主要侧重于促进时空数据处理的深度学习。它将为(I)学习连续动力学，(Ii)在空间和时间上联合建模多尺度结构，以及(Iii)提高神经网络对输入数据中自然扰动的稳健性而发展实用理论。计算成果将是神经网络结构，用于学习稳健的潜在空间嵌入和改进的预测，这将考虑到传感器噪声和地震记录稀缺造成的地面运动数据的不确定性。将使用观测和模拟的地面运动数据来演示所提出的方法的优点。这种技术方法结合了动力系统理论、地震学和深度学习等领域的想法。通过动态系统理论的视角观察时空数据处理和稳健性，该项目旨在建立一个将对广泛的科学问题产生重大影响的原则性框架。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Large earthquakes generate strong ground motions and tsunamis that may lead to a significant number of casualties and cause severe impacts on social resilience in seismically active regions including the West Coast of the United States. Early warning systems have been developed to mitigate immediate threats by detecting first-arriving ground motions near an earthquake epicenter and forecasting the intensity and timing of strong destructive ground motions. To further improve the efficacy and accuracy of these systems, deep learning methods have strong potential, but it is crucial to significantly extend the forecast horizons of existing models. The highly heterogeneous spatiotemporal nature of the seismic wave propagation poses a fundamental challenge to machine learning. This project aims to address these challenges by developing resilient and reliable deep learning methods for forecasting noisy and complex spatiotemporal ground motion data. The resulting methods and open-source software tools will have applications beyond seismology, benefiting diverse scientific and engineering domains including earth, atmospheric, and climate sciences. Furthermore, the project provides research training opportunities for undergraduate and graduate students.The project will primarily focus on advancing deep learning for spatiotemporal data processing. It will develop practical theory for (i) learning continuous dynamics, (ii) modeling multiscale structures jointly in space and time, and (iii) improving robustness of neural networks to natural perturbations in the input data. Computational deliverables will be neural network architectures for learning robust latent space embeddings and improved forecasting, which will account for uncertainties in ground motion data caused by sensor noise and scarcity in seismic recordings. Both observed and simulated ground motion data will be used for demonstrating the advantages of the proposed methods. The technical approach combines ideas from fields such as dynamical systems theory, seismology, and deep learning. By viewing spatiotemporal data processing and robustness through the lens of dynamical systems theory, the project aims to establish a principled framework that will significantly impact a broad range of scientific problems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.