Embedded Files
Interests
Interests
I am interested in probabilistic modeling for the physical sciences, particularly the integration of statistical inference and machine learning with physics-based models. My research has focused on developing new data assimilation algorithms that incorporate machine learning within physically grounded inference frameworks.
PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation
PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation
work at RIKEN
TL;DR: Scalable non-Gaussian state smoothing without true-state supervision using normalizing flow and ELBO training.
Abstract: Many physical data assimilation (DA) workflows require smoothing methods that represent non-Gaussian posteriors over physical state variables, scale to high- dimensional simulators, train from observation windows alone, and remain compatible with calibration of the prescribed simulator. We introduce PR-Smoother, a simulator-preserving amortized smoother designed for this prescribed-simulator DA regime. Its key design principle is to keep the prescribed simulator explicit in both the evidence lower bound and the variational family: rather than learning replacement dynamics or a learned trajectory prior, PR-Smoother learns only future-conditioned corrections around the prescribed rollout. This yields an explicit non-Gaussian smoothing distribution over physical trajectories and supports joint state, parameter, and sensor-bias learning from observations alone. The variational family contains the exact smoother in deterministic and linear-Gaussian limits. Empirically, PR-Smoother captures multimodal posteriors in 4-dimensional Lorenz–96, remains accurate under ambiguous nonlinear observations and process noise in 40-dimensional Lorenz–96, and scales to joint state-parameter-bias inference in 16,384-dimensional Kolmogorov flow.
To appear in Neurips 2026
Kolmogorov flow field reconstruction examples (Fig.4 of the paper).
Deep Bayesian Filter for Bayes-Faithful data assimilation
Deep Bayesian Filter for Bayes-Faithful data assimilation
work at PFN with Keisuke Fukuda and Shin-ichi Maeda (PFN)
TL;DR: Dynamical VAE-based latent DA, based on Koopman operator theory.
abstract: State estimation for nonlinear state space models is a challenging task. Existing assimilation methodologies predominantly assume Gaussian posteriors on physical space, where true posteriors become inevitably non-Gaussian. We propose Deep Bayesian Filtering (DBF) for data assimilation on nonlinear state space models (SSMs). DBF constructs new latent variables ht on a new latent (“fancy”) space and assimilates observations ot. By (i) constraining the state transition on fancy space to be linear and (ii) learning a Gaussian inverse observation operator q(ht|ot), posteriors always remain Gaussian for DBF. Quite distinctively, the structured design of posteriors provides an analytic formula for the recursive computation of posteriors without accumulating Monte-Carlo sampling errors over time steps. DBF seeks the Gaussian inverse observation operators q(ht|ot) and other latent SSM parameters (e.g., dynamics matrix) by maximizing the evidence lower bound. Experiments show that DBF outperforms model-based approaches and latent assimilation methods in various tasks and conditions.
Obtained from our poster presented at ICML 2025.
State space model for multiphase radar observations
State space model for multiphase radar observations
work at PFN with Keisuke Fukuda and Shin-ichi maeda (PFN)
TL;DR: Raindrop size distribution estimate directly from Zh, ZDR, and PhiDP based on variational inference with mean-field approximation.
Abstract: In Dual-polarization radar observations, we can utilize radar observables, namely differential reflectivity (Zdr) and differential phase shift (PhiDP), that allow us to infer the size of raindrops through their deformations. In particular, the specific differential phase shift (kDP), the derivative of the differential phase shift PhiDP, is unaffected by rain attenuation and is a moment with an exponent close to the precipitation rate (approximately the 4.78th power), allowing for better estimation than the Z-R relationship. However, observations of PhiDP are noisy in general. A stable estimation of the true kDP already poses a challenge. Here, we propose to build a state space model that takes raindrop size distribution (DSD) parameters as the latent variables ht and radar observables as the observation variables ot. This formulation allows us to use different relations between observables and the rain rate, thanks to the degrees of freedom in DSD parameters. We prepare a test distribution q(h1:T|o1:T), where the means and standard deviations of the distribution are yielded by neural networks (NN). The NN is trained by maximizing the evidence lower bound.
This work was presented in the fall meeting of Meteorological Society of Japan, 2024.
Page updated
Google Sites
Report abuse