Physical-Process Foccused (2PF) MLWP Seminar Series

The Physical-Process MLWP Seminar Series is intended as a lightweight forum for scientific exchange on machine-learning weather prediction through the lens of atmospheric processes and physical consistency.

The aim is not to promote one model architecture or software framework. The aim is to ask what current MLWP systems can and cannot represent physically, and how we can learn from successes, failures, and targeted diagnostics.

I have also written an introductory blog post about the motivation for the series here.

If you want to be informed about upcoming talks you can join the 2pf-mlwp mailing list by sending email to 2pf-mlwp-request@freelists.org with ‘subscribe’ in the Subject field OR by visiting the mailing list page.

Talk Series

The general plan is that talks will be on the third Thursday of every month at 11:00 Copenhagen time (CET in winter, CEST in summer), with a duration of 1 hour. It is planned that talks will be recorded and made available on the seminar series website.

Speaker (Institution)	Time and date	Title	Recording and slides
Gergely Bölöni (DWD)	24 Sep 2026, 11:00-12:00 CEST	Towards a machine-learning enhanced nowcasting tool for thunderstorm severity analysis and prediction	Recording | Slides (PDF)
Marvin Kähnert (MetNo)	15 Oct 2026, 11:00-12:00 CEST	A first assessment of process-representation in Bris	–
Hauke Schulz (DMI)	19 Nov 2026, 11:00-12:00 CET	Graph Neural Networks as Large-Eddy Simulators: Modelling Cold-Pool Dynamics at Hectometre Scales	–
Francesco Pasquini (KNMI)	17 Dec 2026, 11:00-12:00 CET	Assessing the ability of a stretched-grid deep-learning weather prediction model to capture physical balances	–
Gabriel Moldovan (ECMWF)	21 Jan 2027, 11:00-12:00 CET	Improving the representation of light precipitation in AIFS-Single v1 through non-negativity constraints	–

Have an idea for a presentation? Get in touch with me to discuss your idea and schedule a seminar.

Gergely Bölöni (DWD)

Title: Towards a machine-learning enhanced nowcasting tool for thunderstorm severity analysis and prediction

Authors: Gergely Bölöni, Paul James, Michael Debertshäuser, Susanne Theis

Recording and slides: Watch the recording on YouTube | Download the slides (PDF)

Abstract:

At the German Weather Service (DWD), the NowCastMIX system (James et al., 2018) is an important tool for issuing severe weather warnings on the next-hour timescale using radar, lightning, NWP, and other data. In its analysis step, NowCastMIX detects thunderstorm cells and assigns a severity level based on remote sensing and numerical model data at their locations. In the prediction step, the thunderstorm cells are clustered and then linearly extrapolated according to estimated motion vectors, with the extrapolated clusters retaining the severity assigned in the analysis step.

In order to improve the prediction step, we present a machine learning (ML) approach that aims at non-linear propagation and a representation of the convective life cycle of thunderstorm cell clusters. Our ML-approach is inspired by 2D fluid dynamics, i. e. convective storm development is treated as a forced 2D advection problem. A proof of concept and stages of the ML-model development will be shown through idealized experiments with synthetic data, followed by a demonstration with real data.

Back to top

Marvin Kähnert (MetNo)

Title: A first assessment of process-representation in Bris

Abstract:

This talk presents a process-based evaluation of the ML weather prediction model Bris. Instead of focusing on conventional verification metrics, selected case studies of cold-air outbreaks (CAOs), stable boundary layers (SBLs), and persistent fog are analyzed. A deeper analysis of a CAO case investigates whether Bris reproduces the spatial coherence and organised structures of the event, providing first insights into how well the model represents atmospheric processes beyond pointwise forecast skill.

Back to top

Hauke Schulz (DMI)

Title: Graph Neural Networks as Large-Eddy Simulators: Modelling Cold-Pool Dynamics at Hectometre Scales

Abstract:

Data-driven weather prediction has recently matched or surpassed traditional physics-based models at regional and global scales, yet progress towards models operating at hectometre (sub-kilometre) resolution remains limited. This scale is particularly challenging: training data are expensive to generate, and the dominant physical processes such as shallow convection and the cold pools it produces are highly nonlinear and spatially intermittent.

We present a graph neural network (GNN) trained to emulate Large-Eddy Simulation (LES) at 100 m horizontal resolution and sub-hourly time steps. The model is trained on 42 days of high-resolution, realistic simulation output over the winter-time trade-wind regime in the western Atlantic, a region where shallow convective clouds strongly influence the planetary energy budget through their effect on reflected solar radiation.

A key focus is cold-pool dynamics: the cool, dense air masses that spread near the surface when rain evaporates beneath clouds. Cold pools drive the formation of new clouds, organise cloudiness at mesoscales, and require the model to correctly couple processes in the cloud layer with the surface, making them a stringent physical consistency test. Through an ablation study, we quantify which input variables matter most for accurately reproducing the surface temperature perturbations cold pools produce, offering practical guidance for future parameterisation schemes and data selection.

Finally, we demonstrate that the GNN can deterministically predict the lifecycle of cold pools across multiple successive generations, suggesting that graph-based LES emulators can robustly capture the nonlinear feedbacks that govern mesoscale organisation in shallow convective regimes.

Back to top

Francesco Pasquini (KNMI)

Title: Assessing the ability of a stretched-grid deep-learning weather prediction model to capture physical balances

Abstract:

Weather forecasting has traditionally relied on Numerical Weather Prediction (NWP) models, which simulate weather by solving the governing fluid equations. Recently, the emergence of Deep Learning Weather Prediction (DLWP) models has opened a new era in weather forecasting, offering a data-driven alternative to classical NWP approaches. Regional DLWP models such as the stretched-grid model Bris developed by Met Norway have demonstrated performance on par with, or even slightly better than, regional NWP models across a range of standard forecast metrics. By overcoming the coarse horizontal resolution that constrained earlier global data-driven models, the operational use of regional DLWP systems now appears increasingly promising.

Nevertheless, the performance of such models during extreme events is generally inferior to that of regional NWP models, and comprehensive evaluations of their ability to generate physically realistic forecasts are still lacking. Here, we present a study comparing the physical consistency of the deterministic version of Bris with the control run of the operational MetCoOp Ensemble Prediction System (MEPS) in forecasting the severe extratropical cyclone Poly, which hit the Netherlands on 5 July 2023. We examine whether Bris accurately represents deviations from key atmospheric balances and whether it reproduces expected dynamics of the storm.

We show that, despite its relatively good performance in terms of RMSE, Bris struggles to capture important mesoscale features of the event and that it significantly disrupts several atmospheric balances. This unrealistic disruption is mainly linked to the fine-scale noise evidenced in its output fields, which leads to incorrect and unrealistic spatial gradients. These results raise critical questions for improving AI-based models to better represent extreme events and how to ensure physical consistency in their predictions.

Back to top

Gabriel Moldovan (ECMWF)

Title: Improving the representation of light precipitation in AIFS-Single v1 through non-negativity constraints

Abstract:

Graph neural network-based models such as AIFS and GraphCast have been shown to represent precipitation with an intense drizzle bias. In this work, we show that this behaviour is related to gradient ambiguity arising at the zero-precipitation boundary when training with MSE. By simply enforcing non-negativity on the total precipitation head, we show that the model learns a substantially different representation of precipitation. More specifically, the resulting geometrical boundary at the zero-precipitation threshold allows the model to use the negative-output region as a proxy for dryness, while eliminating the gradient ambiguity at zero precipitation.

Back to top

Guiding Questions

Each seminar should try to answer:

• What physical process is being studied?
• What model, dataset, architecture, training setup, diagnostic, or constraint is being tested?
• What was the hypothesis for improving or evaluating representation of the process?
• What were the findings?
• What might transfer to other MLWP settings?
• What remains unresolved?

Candidate Themes

• Convective initiation and precipitation organisation.
• Cloud and precipitation physics.
• Physical consistency and conservation constraints.
• Geostrophic balance, stability, waves, and dynamical consistency.
• Extremes and rare events beyond pointwise scores.
• Spatial structure and object-based verification.
• Generative and probabilistic models: are sampled structures physically meaningful?
• Physics-based constraints and regime separation, for example dry versus slight-rain conditions.
• Data assimilation and latent-state representations.
• Architecture choices from a physical-process perspective.

Participation

The series is still being formed. Suggestions for speakers, papers, topics, or examples are welcome. See Contact.