Research

Interpretable AI-Analogs for Seasonal Prediction

Deep learning models can make skillful climate forecasts, but it is often hard to explain why they make the choices they do. We developed a hybrid method that keeps the forecast interpretable. It starts from model-analogs forecasting: find past model states that look like today's observed state and follow how they evolved. A convolutional neural network then learns which regions matter most when choosing those analogs. Every forecast is a set of real model trajectories, so the physics stays visible.

We first applied this method to the El Niño–Southern Oscillation (ENSO). The network learns to weight the initial ocean state so that the selected analogs lead to better forecasts, and it reveals where initial errors matter most for ENSO prediction. Read more in Toride et al. (2025, AIES) or try the code on GitHub.

We are now extending the approach to global seasonal precipitation. The network learns spatially varying weights for the ocean initial state, and the resulting forecasts outperform unweighted analogs and dynamical models at long leads, especially in regions where ENSO teleconnections drive rainfall. Results have been presented at the CESM Workshop 2026 and the NOAA AI Workshop 2026, and a paper is in preparation.

Subseasonal Predictability of Atmospheric Rivers

Atmospheric rivers bring much of the flood-producing rain to the West Coast of Northern America. Predicting them two to five weeks ahead remains difficult. The Madden–Julian oscillation (MJO) is known to modulate atmospheric river activity on this time scale, but the actual forecast signal is often weak because many other factors are involved.

In Toride, Hakim, and Hoell (2026, J. Climate), we used a linear inverse model (LIM) to ask which subseasonal conditions lead to landfalling atmospheric rivers over Alaska, the Pacific Northwest, and California in winter. We separated the dynamics into three kinds of modes: modes tied to tropical sea surface temperature, such as ENSO; modes coupled with tropical heating, such as the MJO; and modes that are only weakly coupled between the tropics and the extratropics. The result was a surprise. Prolonged atmospheric river activity in all three regions is driven mainly by the weakly coupled modes. The MJO's contribution is relatively minor: its only clear signal is subtropical vapor transport toward Alaska during MJO phases 6–7. The MJO does help initial conditions grow into atmospheric river patterns, but the weakly coupled modes remain the main driver. These weakly coupled processes decay faster than MJO signals, which makes them harder to predict, but resolving them well is key to better subseasonal atmospheric river forecasts.

This work builds on two earlier studies. Toride and Hakim (2021, GRL) showed that the MJO's influence on North American atmospheric rivers depends strongly on the low-frequency state of the Pacific/North American (PNA) pattern, which helps explain why MJO-based forecasts succeed in some years and fail in others. Toride and Hakim (2022, J. Climate) then asked what distinguishes the MJO events that actually produce atmospheric rivers from those that do not.

Water Vapor Isotopes for Weather Forecasting

Water isotopes such as HDO and H₂¹⁸O carry a record of every evaporation, condensation, and mixing step that an air mass has gone through. That makes them a tracer of moist processes and latent heating that ordinary temperature and humidity observations cannot see. Satellite instruments such as IASI now provide global isotope observations in the mid-troposphere, which raises a simple question: can they improve weather forecasts?

In Toride et al. (2026, Commun. Earth Environ.), we gave the first answer with real satellite data. We assimilated bias-corrected mid-tropospheric δD from IASI into the isotope-enabled global model IsoGSM using an ensemble Kalman filter. Compared with assimilating temperature and humidity alone at the same coverage, adding δD improved forecasts of wind, temperature, humidity, and geopotential height out to five days, with the largest gains in the midlatitudes. Heavy precipitation forecasts improved as well. The isotope information helps most through its constraint on transport, which is a signal that humidity alone does not provide.

This result follows a line of earlier work. Toride et al. (2021, GRL) used idealized experiments to show that mid-tropospheric isotopes could improve large-scale circulation and predictability, and the paper was highlighted as a Science Editors' Choice. Tada et al. (2021, Sci. Rep.) demonstrated forecast gains with a simpler setup. Schneider et al. (2024, AMT) examined how well satellite isotope observations can constrain analyses of convective events. With international collaborators, we are now extending isotope data assimilation to the global cloud-resolving model NICAM-WISO.

Extreme Precipitation and Hydrologic Design

Probable maximum precipitation (PMP) is the theoretical upper limit of rainfall over a region. It is widely used to design and assess dams and other water infrastructure. We estimated PMP for the Pacific Northwest by realistically maximizing storms in an atmospheric model. Instead of scaling moisture everywhere, we perturbed moisture only along the path of the atmospheric river, guided by vertically integrated water vapor flux. This produced more realistic atmospheric fields and more severe precipitation than earlier methods. Read more in Toride et al. (2019, J. Hydrometeorology).

We also reconstructed 160 years of precipitation over two US West Coast watersheds with a regional atmospheric model. The reconstructions show increasing extreme precipitation, a sharp rise in year-to-year variability, and trends that differ from the regional-scale picture. Read more in Toride et al. (2018, Sci. Total Environ.) and Toride et al. (2019, Sci. Total Environ.).

Reconstructing Historical Weather from old diaries

Long before instruments, people wrote down the weather. Diaries around the world hold daily records of sky conditions going back centuries. We built a system that assimilates this qualitative information into a numerical weather model. Using visually observed cloud cover over Japan, we showed that it is possible to reconstruct daily weather from such records (Toride et al. 2017, MWR). Later work with Xiaoyu Wang and Kei Yoshimura introduced a Gaussian transformation for assimilating cloud cover and weather categories, which improves the reconstruction (Wang et al. 2022, JSCE; Wang et al. 2023, MWR).

Land Surface Hydrology

The land surface is the most heterogeneous part of the hydrologic cycle. Land use, soil, vegetation, and topography all vary over short distances. We proposed a way to observe soil moisture at high space and time resolution by combining two microwave sensors in a land data assimilation system: a passive sensor that observes often but coarsely, and an active sensor that observes finely but rarely. The system captured soil moisture at high spatiotemporal resolution at two lightly vegetated sites (Toride et al. 2019, Sensors). Related work retrieved soil moisture, vegetation water content, and surface roughness at the same time from microwave and optical observations (Sawada et al. 2017, IEEE TGRS). I also developed a theoretical treatment of spatial heterogeneity in soil water flow using a stochastic differential equation, the Fokker–Planck equation.