Construction of Theoretical Foundations for Deep Learning–Based Point Process Modeling of Earthquake Occurrence

Reference No.	2026a039
Type/Category	Grant for Project Research- Short-term Visiting Researcher
Title of Research Project	Construction of Theoretical Foundations for Deep Learning–Based Point Process Modeling of Earthquake Occurrence
Principal Investigator	Atsutomo Yara（Graduate School of Engineering Science, The University of Osaka・PhD student）
Research Period
Keyword(s) of Research Fields	deep learning, nonparametric method, convergence rate, curse of dimensionality, minimax optimality
Abstract for Research Report	The objective of this research is to develop a framework for modeling earthquake occurrence as a point process, in which a complex intensity function is estimated nonparametrically using deep learning, and to establish theoretical guarantees for the resulting estimation accuracy. Data on earthquake occurrence involve several challenging features, including spatiotemporal dependence, nonlinearity, and high-dimensional covariates, and conventional parametric modeling may lack sufficient expressive power to capture such complex structure. At the same time, flexible models such as deep learning require a solid theoretical foundation, particularly from the viewpoints of overfitting and interpretability. Accordingly, this research will evaluate the accuracy of estimates obtained when the intensity function of a point process model for earthquake occurrence is learned using deep learning, thereby clarifying under what circumstances deep learning–based modeling is effective for earthquake occurrence data. In addition, to improve interpretability, we will investigate a semiparametric modeling approach in which the intensity function is decomposed into an interpretable low-dimensional component and a flexible component modeled by deep learning, aiming to achieve both high estimation accuracy and interpretability. The expected outcomes of this research are as follows. First, we will assess the estimation accuracy of estimators for point processes with covariates and derive convergence rates for intensity function estimation. This will enable a quantitative evaluation of the performance of deep learning–based estimators. Second, we will theoretically characterize how estimation accuracy varies depending on the amount and nature of the data, thereby providing guidance for model selection and design in earthquake data analysis. Third, we will develop a highly interpretable estimation method based on semiparametric modeling, along with its theoretical guarantees, thereby advancing the analytical foundations for understanding seismic activity and applications to related fields.
Organizing Committee Members (Workshop) Participants (Short-term Joint Usage)	Atsutomo Yara（Graduate School of Engineering Science, The University of Osaka・PhD student）