Xenia Wróbelska
Digital Twin Engineer — Physics-Based Models · MLOps & Digital Engineering
Warsaw, Poland
Digital Twin Engineer specializing in physics-based modeling for aerospace system digital twins. Uses Modelica/Dymola to build first-principles models of thermal, fluid, and mechanical systems, and converts them to FMUs (Functional Mock-up Units) for deployment in co-simulation and real-time digital twin environments. Works closely with domain engineers (thermal, propulsion, ECS) to ensure physics models are valid and appropriately simplified for real-time use.
Expertise
- Modelica and Dymola physics-based modeling
- physics-informed digital twin development
- model order reduction for real-time simulation
- co-simulation (FMI/FMU)
- thermal and fluid system digital twins
Technologies
Work History
2025-01
Physics-informed surrogate training data generation — used the validated Dymola FMU to generate 10,000 thermal management system operating points as training data for neural network surrogates. Designed sampling strategy to cover the operating envelope.
Challenge: FMU batch simulation with 10,000 parameter combinations at 1 Hz from Python was rate-limited by FMU initialization overhead — each simulation call included a costly FMU reset. Optimized by running continuous parameter sweep simulations with step changes rather than individual FMU calls, reducing total simulation time by 70%.
Learned: FMU batch simulation for surrogate data generation is much faster with continuous parameter sweeps than with individual initialization calls. Design the sampling strategy as a continuous trajectory through the parameter space where the FMU can run continuously, rather than as independent point samples requiring FMU re-initialization.
2024-05
ECS digital twin physics model — built a Dymola model of the aircraft Environmental Control System (bleed air, pack, mixing manifold, cabin) for integration into the fleet digital twin. Validated against test rig data.
Challenge: The ECS Dymola model had 380 states — too many for real-time FMU execution at 1 Hz update rate. Applied singular perturbation model order reduction: algebraic substitution of fast pneumatic states reduced model to 45 states while maintaining <3°C cabin temperature prediction error.
Learned: Modelica/Dymola models for real-time FMU deployment must undergo model order reduction targeting the specific real-time update rate. Singular perturbation is the standard approach for pneumatic/thermal systems — document the reduced model validation against the full-order model as part of the FMU acceptance criteria.
2023-10
FMU export and co-simulation integration — exported the thermal management Dymola model as an FMU (FMI 2.0 Model Exchange) and integrated into an Azure Digital Twins co-simulation framework with the flight performance model.
Challenge: The FMU exported from Dymola had a stiff ODE (stiffness ratio ~10^5) that made the FMI master algorithm's default explicit Euler solver unstable. Required switching to an implicit Euler solver in the FMI master, which was not supported by the Azure Digital Twins FMI integration — had to implement a custom FMI master wrapper.
Learned: FMU stiffness must be assessed before committing to a co-simulation master algorithm. Stiff Modelica/Dymola models require implicit solvers that many standard FMI masters do not provide. Request a model index and stiffness estimate from the Dymola engineer before designing the co-simulation integration architecture.
2023-03
Dymola thermal management system model for turbofan lube system — translated the AMESim model from thermal engineering team into Modelica/Dymola for FMU export and co-simulation. Validated equivalence between the Dymola and AMESim implementations.
Challenge: AMESim and Dymola use different formulations for hydraulic orifice flow — the AMESim turbulent flow coefficient at low Reynolds number diverged from the Dymola MSL Orifice model. Required re-parameterizing the Dymola orifice using AMESim's discharge coefficient definition to achieve equivalent low-flow behavior.
Learned: Translating system models between AMESim and Dymola requires careful component-by-component equivalence verification, not just top-level output matching. The underlying mathematical formulations for even simple components (orifices, heat exchangers) differ between tools — match physics at the component level, not just at the system output.