[1P5-OS-16] OS-16

[1P5-OS-16-03] Parameter Estimation of Known Parts in a Target Mechanism Using Independent Component Analysis

〇Daisuke Azuma², TAKASHI WASHIO¹ (1. Kansai University, 2. Osaka University)

Keywords:Data Assimilation, Independent Component Analysis, Model Parameter Estimation

In a wide range of industrial and scientific fields, computer simulations using mathematical models that reflect known physical mechanisms or structures of the target system are employed to understand and estimate target behaviors. In recent years, data assimilation methods, which learn mathematical model parameters to align simulation results with observational data, have been widely utilized. However, in many cases, our knowledge is incomplete, and there are numerous targets that include unknown mechanisms not reflected in the mathematical models. In such cases, applying data assimilation to the known mathematical model may introduce bias in model parameter estimation due to the unknown mechanisms. General robust estimation methods that assume random disturbances struggle to accurately estimate parameters without such bias. To address this issue, this study proposes a novel data assimilation principle that enables unbiased model parameter estimation when unknown mechanisms independent of the known ones exist in the target system. This is achieved by incorporating the Hilbert-Schmidt Independence Criterion (HSIC), a widely applicable measure of independence, into the loss function to ensure that distributions of the input variables and the estimated behavior residuals of the mathematical model are mutually independent.