The UrbanAIR data assimilation experts at the International EnKF workshop 2026
- 11 hours ago
- 5 min read
Data assimilation (DA)—the science of combining sparse observational data with advanced numerical models to produce more accurate forecasts—is at the very core of what we do at UrbanAIR. By integrating physical models with real-time data, our project builds high-fidelity digital twins for urban air quality and heat mitigation as part of the European Union's Destination Earth initiative. However, urban-scale atmospheric modeling deals with immense high-dimensional states, chaotic wind flows, and non-Gaussian statistics induced by interactions with building geometries.
Navigating these grand challenges is exactly why our team traveled to the picturesque town of Rosendal, Norway, for the EnKF (Ensemble Kalman Filter) Workshop 2026, which took place from June 15 to 18, 2026. This premier global gathering serves as a vital platform for cutting-edge research across atmospheric, oceanic, and environmental systems. Two of the workshop's five prestigious invited speakers are core partners in the UrbanAIR project, while the rest of our delegation made significant contributions across the entire program.
Following the chronological flow of the workshop program, here is a comprehensive look at how the UrbanAIR team showcased our latest breakthroughs, methodologies, and framework advancements in Rosendal.
Day 1: Interactive Poster Sessions (Optimizing Sensors & Resilient Layouts)
On June 15th, the workshop kicked off with vibrant, interactive poster sessions on the first day, providing an active space for our researchers, David Plazas and Zahra Mehraban, to present deep, localised technical solutions for microclimate tracking.
David Plazas
PhD Candidate in Geosciences and Engineering at TU Delft (Netherlands)
Poster Title:
Inflow Estimation and Sensor Placement in High-Resolution Urban Wind Flows
High-resolution Computational Fluid Dynamics (CFD) models provide invaluable data for pedestrian comfort and pollutant dispersion, but suffer from heavy input uncertainties. David presented work utilizing Ensemble Smoothing (ES) to infer unknown inflow wind angles. By conducting assimilation on a pre-computed database of steady Reynolds-averaged Navier–Stokes (RANS) simulations over the TU Delft campus spanning a full $360^\circ$ range, his results revealed that velocity magnitude carries directional information primarily in building wake regions. Observing velocity in open zones can lead to mathematically non-unique inversions, proving that local directional velocity components offer more robust, complementary information for optimizing future city sensor placement.
Zahra Mehraban
Research Associate at Ilmenau University of Technology (Germany)
Poster Title:
Parameter Estimation in Urban LES Using Ensemble Smoother with Multiple Data Assimilation
As a vital stepping stone toward creating real-world physics-based digital twins for urban planning and climate resilience, Zahra presented a data assimilation framework tailored for estimating uncertain parameters in Large-Eddy Simulations. Utilizing the Parallelized Large-Eddy Simulation Model (PALM), her work implements the Ensemble Smoother with Multiple Data Assimilation (ES-MDA) to approximate the Bayesian posterior distribution through iterative update steps. By isolating performance within an idealized urban geometry, her perfect-model twin experiments successfully validated how sparse spatial observation networks can recursively home in on hidden ambient conditions.
Day 2: Oral Presentations & Keynotes (Scaling Up & Evaluating "Success")
On the second day, the workshop transitioned to oral technical sessions and invited keynotes, with UrbanAIR being heavily represented. Two of the five workshop-wide invited keynote slots were delivered by our project partners, Geir Evensen and Jana de Wiljes, while Nikolaj Takata Mücke and our coordinator Femke Vossepoel complemented our team's contribution.
Prof. Geir Evensen (Invited Keynote Speaker)
Chief Scientist at NORCE (Norway) / Pioneer of the EnKF
Talk Title:
Controlling inflow boundary conditions in a lattice Boltzmann LES model for urban flows
In his invited lecture, Geir detailed how data assimilation can control and improve uncertain boundary conditions passed down when nesting localized microscale simulations inside coarser weather prediction models. Configuring the system as a standard parameter estimation problem, his team applied the Ensemble Smoother with Multiple Data Assimilation (ESMDA) over four update steps. Using local wind measurements within the simulation domain, they proved that a minimal sensor array can recursively correct time-varying boundary inflows, yielding stable, reliable urban flow structures over time.
Nikolaj Takata Mücke
PhD Researcher in Data Assimilation & Machine Learning at TU Delft (Netherlands)
Talk Title:
Ensemble Smoothing for Joint State and Parameter Estimation in Turbulent Urban Flows
Running data assimilation across completely different fluid dynamics software engines is traditionally a software engineering nightmare. Nikolaj introduced pyurbanair, a unified Python abstraction layer developed for our project that wraps over three major Fortran-based solvers: uDALES, PALM, and a custom LBM (Lattice Boltzmann Method) code.
Nikolaj's talk highlighted a critical project benchmark: by equipping ESMDA with specialized initial-condition updates or an SVD-subspace state reduction, joint state-and-parameter estimation successfully reconstructed complex wind geometries. Furthermore, he demonstrated an exciting pivot toward Neural Surrogates. Because running full Large-Eddy Simulations (LES) is computationally expensive, training a deep 3D ConvNeXt U-Net offline allows the neural network to step in as a drop-in forward model inside the ESMDA inner loop—yielding massive computational speedups with exceptional tracking accuracy.
Prof. Dr. Femke C. Vossepoel
Associate Professor / Principal Investigator at TU Delft (Netherlands)
Femke challenged the conventional standards used to measure data assimilation success. While many projects solely evaluate metrics like Root-Mean-Square Error (RMSE) against sparse observations or immediate forecasting skill, these standard measures do not always reflect how well a system captures the true underlying probability distribution of parameters. From a strict Bayesian perspective, she argued that performance should be assessed by how well the posterior distribution is reconstructed using information theory metrics like the Kullback–Leibler (KL) divergence and entropy. Her presentation underscored the importance of selecting proper performance criteria to ensure robust, validated urban digital twins.
Dr. Jana de Wiljes (Invited Keynote Speaker)
Junior Professor for Data Assimilation at Ilmenau University of Technology (Germany)
Talk Title: Sequential Learning Methods for High-Dimensional Data Assimilation
Focusing on the strict mathematical architecture of state tracking, Jana's invited lecture investigated continuous ensemble Kalman filtering under fixed, randomized, and adaptively varying partial observations. Her work establishes rigorous mathematical bounds for expected tracking errors in chaotic dynamical systems. She introduced an innovative sequential learning scheme that adaptively balances observation complexity with tracking accuracy, helping manage the "curse of dimensionality" when isolating the ideal filter-relevant state subspace.
All workshop presentations and materials are being progressively compiled on the official EnKF Workshop 2026 Portal.
As the workshop wrapped up last week, the UrbanAIR delegation returned from Rosendal with clear, collaborative objectives. Moving forward through our Horizon Europe timeline, the consortium will focus on:
Further refining localization strategies, ensemble sizing, and window lengths for high-dimensional states.
Strengthening the long-term rollout stability of our machine learning neural surrogate models.
Scaling these unified data assimilation tools up from idealized layouts to ultra-large, complex real urban geometries across our pilot cities of Antwerp, Barcelona, Paris, Bristol, and Rotterdam.
Many thanks to the organisers of the EnKF Workshop 2026 for putting together such an inspiring and engaging program and bringing the data assimilation community together in a spectacular Norwegian setting. We are looking forward to more knowledge-sharing beyond the workshop and future collaborations!
Stay tuned—more updates on our activities are coming up!
Comments