Nikolaj Takata Mücke
Postdoc · Delft University of Technology · Associated member, CWI
I am an applied mathematician working at the intersection of scientific computing, machine learning, and Bayesian inference. I am a postdoc at Delft University of Technology, the Netherlands, and an associated member of the Scientific Computing group at CWI, Amsterdam.
I develop fast surrogate models, physics-consistent generative methods, and data assimilation techniques for systems governed by partial differential equations, particularly turbulent flows, atmospheric dynamics, subsurface flow, and pipe flow. My research combines numerical simulation, scientific machine learning, and uncertainty quantification to extract as much information as possible from sparse observations. The goal is to strengthen the mathematical, computational, and statistical foundations of digital twins that forecast efficiently, adapt to real-world measurements, and provide reliable uncertainty estimates.
News: Together with Syver Døving Agdestein, I won 1st place in the NeurIPS 2026 RealPDE Competition (Track 2: Long-Term Test-Time Adaptation).
Publications
Generative Super-Resolution of Turbulent Flows via Stochastic Interpolants
AI Enhanced Data Assimilation and Uncertainty Quantification Applied to Geological Carbon Storage
Markov Chain Generative Adversarial Neural Networks for Solving Bayesian Inverse Problems in Physics Applications
Reduced Order Modeling for Parameterized Time-Dependent PDEs Using Spatially and Memory Aware Deep Learning
Xaurora: Finetuning an Atmospheric Foundation Model for Generative Weather Forecasting
Advancing Data Assimilation and Uncertainty Quantification for CO2 Sequestration through AI-Hybrid Methods
Machine Learning-Based Digital Twin for Water Distribution Network Anomaly Detection and Localization
Reduced Order Modelling for Wave-Structure Modelling
Reduced Order Modelling for Dispersive and Nonlinear Water Wave Modelling
Reduced Order Modeling for Nonlinear PDE-Constrained Optimization Using Neural Networks
EC-EarthFlow: Probabilistic Emulation of Daily Transient Global Climate Model Simulations with Flow Matching
Xaurora: Generative Weather Forecasting with Denoising Stochastic Interpolants from a Foundation Model Prior
A Unified Framework for Bayesian Data Assimilation with Generative Models and Observation Interpolants
Leveraging Crowdsourced Weather Data across Europe with Scalable Multi-Fidelity Gaussian Process Regression
Experience
Postdoc
Working with Femke Vossepoel in the Horizon Europe project UrbanAIR, where I am also scientific co-coordinator. I develop data assimilation methods for three-dimensional urban atmospheric large-eddy simulations (LES), including state and parameter estimation under uncertain boundary conditions, and transformer-based neural surrogates that generalise across domains and geometries. I designed and built pyurbanair, which integrates Fortran LES solvers (uDALES, PALM, and LBM), ensemble simulation, and learned surrogates into a common framework. I collaborate with TU Ilmenau, NORCE, Barcelona Supercomputing Center, and Imperial College London.
Associated Member
Continuing joint work with the group.
Postdoc
Working with Benjamin Sanderse, funded by the NGF AiNed XS Europe grant we co-wrote. I developed energy-aware stochastic-interpolant generative models for probabilistic turbulence forecasting, and methods for conditioning generative models on observations. I collaborated on probabilistic climate and weather modelling with KNMI and Vrije Universiteit Amsterdam, and with DTU on super-resolution for turbulent flows.
AI Lead (part-time)
I developed deep learning methods for soil organic carbon estimation from remote sensing data and the deep learning infrastructure on Databricks and Google Cloud, and led a small machine learning team. I led the machine learning contribution to a successful €350,000 ESA demonstration project proposal and to a €350,000 investment from UNIIQ, and was main applicant for two EuroCC computing allocations on Snellius.
PhD Candidate
Research on neural network surrogate models for PDEs, generative priors, Bayesian inverse problems, and uncertainty quantification, aimed at real-time data assimilation for digital twins. Supervised by Cornelis W. Oosterlee and Sander M. Bohté.
Research Assistant
Working with Allan Peter Engsig-Karup and Mads Peter Sørensen, I developed a GPU-accelerated C++ simulator for the generalised nonlinear Schrödinger equation, with Monte Carlo and polynomial-chaos uncertainty quantification. The project dealt with low-noise supercontinuum sources for ultra-high resolution optical coherence tomography.
Student Assistant
I developed PDE-based buckling analysis for offshore-wind foundations and computational tools for wind-farm construction planning.
Science Communicator
Experimentarium is a science museum mainly aimed at children and families. I hosted exhibitions, demonstrations, and live science shows for public audiences and school groups, and developed experiences explaining unfamiliar physical phenomena accessibly.
Teaching & Supervision
Wind Forecasting for Renewable Energy
Turbulence Closure Modelling Using Stochastic Interpolants
Diffusion Models for Time-Series Denoising
Guidance in Using Robotic-Arm-Assisted Surgical Systems for Knee Arthroplasty
Stock-Price Simulation under Jump-Diffusion Dynamics: A WGAN-Based Framework with Anomaly Detection
Hamiltonian Neural Networks for Fluid Flows
Traditional and ML Approaches to Implied-Volatility Surfaces
Data Assimilation for the Geosciences
Neural Networks in Finance
Computational Imaging Masterclass
Scientific Computing for Differential Equations 2 (02687)
Advanced Engineering Mathematics 2 (01025)
Scientific Computing for Differential Equations (02685)
Conferences, Workshops & Seminars Organised
Data Assimilation for Fluid Dynamics
Co-Organisers: G. Hastermann (TU Ilmenau) and K. Takeda (Nagoya University)
Theory-Guided Design of Deep-Learning Surrogates
Co-Organisers: N. R. Franco, S. Brivio, and A. Manzoni (Politecnico di Milano)
Physics-Consistent Generative Modelling
Co-Organiser: Benjamin Sanderse
Deep-Learning Latent-Space Models for Scientific Computing
Co-Organiser: Wouter Edeling
Machine Learning and Stochastic Modelling for Dynamical Systems
Co-Organiser: Wouter Edeling
AIMET-NL Workshop: Scientific Machine Learning & Uncertainty Quantification
Co-Organiser: Daan Crommelin
Digital Twins for Pipe-Transport Networks
Co-Organisers: Prerna Pandey, Shashi Jain, Cornelis W. Oosterlee, and Sander M. Bohté
Machine Learning for Physics-Based Modelling
Co-Organisers: Prerna Pandey, Shashi Jain, Cornelis W. Oosterlee, and Sander M. Bohté
AI and IoT for Flow Modelling
Co-Organisers: Shashi Jain, Cornelis W. Oosterlee, and Sander M. Bohté
Science Park Informal Probability Seminars
Talks
NeurIPS RealPDE Competition, Track 2: Long-Term Test-Time Adaptation
Presentation of our winning entry, together with Syver Døving Agdestein.
Energy-Consistent Generative Modelling for Forecasting and Data Assimilation via Stochastic Interpolants
Energy-Consistent Generative Models for Physics Simulations
Probabilistic Leak Localisation in Pipe Flows
Soil Organic Carbon Monitoring with Deep Learning
Probabilistic Digital Twins for Leak Localisation
ISDA 2026 (Nanjing) · EnKF Workshop 2026 (Oslo) · ENUMATH 2025 (Heidelberg) · SIAM CSE 2025 (Fort Worth) · ECCOMAS 2024 (Lisbon) · CFC 2023 (Cannes) · SIAM CSE 2023 (Amsterdam) · SIAM UQ 2022 (Atlanta) · MMLDT 2021 (San Diego) · ECMI 2021 (Wuppertal, online) · ENUMATH 2019 (Egmond aan Zee)
Service & Leadership
Scientific Co-Coordinator and Executive Committee Member, UrbanAIR
Writing and reviewing deliverables, coordinating scientific collaboration across partners, organising workshops, representing the consortium at conferences, and engaging with stakeholders.
Rotterdam Municipality Liaison, UrbanAIR
PhD Hiring Committee Member, CWI
AIMET-NL
Reviewing
Software
pyurbanair
Open-source simulation and data assimilation framework for the Horizon Europe project UrbanAIR. It runs three CFD solvers (Lattice Boltzmann, uDALES, PALM) and neural surrogates in ensembles on HPC through a common interface, with ESMDA and ensemble Kalman filters (ETKF, LETKF) in JAX for estimating inflow parameters and flow states from sensor data. Includes a browser app that builds city cases from OpenStreetMap and an MCP server exposing the workflow to AI agents.
scientific-stochastic-interpolants
Framework for generative modelling of spatiotemporal PDE dynamics with stochastic interpolants, flow matching, and diffusion models. Applied to turbulent Navier–Stokes flow, urban LES, and KNMI climate simulations, with posterior sampling for Bayesian data assimilation from sparse observations and benchmark implementations of competing methods.
non_gaussian_data_assim
Library and experiment harness for ensemble data assimilation with non-Gaussian filters (EnKF, adaptive Gaussian mixture filter, particle flow filter, bootstrap particle filter), benchmarked on Lorenz 63, Lorenz 96, and Kuramoto–Sivashinsky systems.
SCIFiR
Scalable Crowdsource data Integration with multi-Fidelity Regression: maps daily temperature across Europe by combining sparse official stations with dense crowdsourced weather stations using a multi-fidelity Gaussian process. Developed in a collaboration between CWI and KNMI.
atlas
Local-first research library and project workspace built on Zotero, with Markdown/LaTeX notes, BibTeX export, literature feeds from arXiv and OpenAlex, and a local MCP server giving AI assistants full-text retrieval over the library.
Education
PhD in Mathematics
Thesis: Deep Learning for Real-Time Inverse Problems and Data Assimilation with Uncertainty Quantification for Digital Twins. Supervisors: Cornelis W. Oosterlee and Sander M. Bohté.
MSc, Mathematical Modelling and Computation
Thesis: Optimal Control of Nonlinear PDE Systems Using Reduced Order Modeling and Neural Networks.
BSc, Mathematics and Technology
Thesis: Dynamical Systems Approach to Optimal Tuna Trajectories.