Łukasz Zając
Thermal ML Engineer · Thermal Engineering
Warsaw, Poland
Thermal ML Engineer at the intersection of computational engineering and machine learning. Develops surrogate models that replace expensive CHT simulations with fast ML predictions for design exploration and optimization. Works with thermal engineers to formulate engineering problems as ML tasks and with the Data & AI Platform team to deploy surrogates in production engineering workflows.
Expertise
- ML surrogate models for thermal analysis
- physics-informed neural networks (PINN)
- Gaussian process regression for engineering data
- active learning for simulation data sampling
- integration of ML models into engineering workflows
Technologies
Work History
2025-01
Surrogate model deployment pipeline with MLflow — packaged 4 thermal surrogate models (blade temperature, cooling effectiveness, heat flux, life estimate) as versioned MLflow models with automated retraining triggers when new CHT data is added.
Challenge: Surrogate retraining triggered by new CHT data could silently degrade model performance if the new data came from a different operating regime. Implemented automated comparison tests: new model must match or exceed previous model on a held-out reference set before promotion.
Learned: Surrogate model retraining pipelines must include model regression testing before deployment. Continuous learning without regression testing leads to silent model degradation — adding new data does not always improve the model, especially if it comes from an unbalanced part of the design space.
2024-05
Physics-informed neural network (PINN) for turbine blade thermal conduction — embedded the heat equation as a soft constraint in the neural network training loss, reducing training data requirements further and improving extrapolation behavior.
Challenge: PINN training convergence was unstable — the PDE residual loss term dominated the data loss term for certain initial weight configurations, causing the optimizer to minimize the physics constraint while ignoring the training data. Required careful loss weighting annealing during training.
Learned: PINN loss weighting between data and PDE residual terms is critical and problem-specific. Start with high data loss weight and gradually increase PDE weight during training — this avoids the optimizer getting trapped in PDE-satisfying but data-ignoring solutions.
2023-10
Active learning sampling strategy for CHT training data — used query-by-committee with 3 GP models to identify the most informative next simulation points, reducing the training data requirement from 300 to 120 cases for the same accuracy.
Challenge: Active learning query selection required running the GP ensemble at every candidate design point to compute disagreement — with 5,000 candidate points and 3 GP models, this was computationally expensive. Optimized using batch query selection with expected improvement computed on a sparse candidate grid.
Learned: Active learning for engineering simulation surrogates saves significant HPC cost but the query selection computation itself needs to be efficient. A two-stage approach — coarse screening on a sparse grid, then fine-grained optimization in the promising region — balances query quality and selection speed.
2023-04
Gaussian process surrogate model for HPT blade metal temperature prediction — trained on 150 STAR-CCM+ CHT results with 8 design variables (cooling channel dimensions and blowing ratios). Deployed for use in early design exploration.
Challenge: GP model prediction variance exploded outside the training data convex hull — designers were querying the surrogate at infeasible design points and getting highly uncertain predictions without realizing it. Implemented a convex hull check that flags out-of-bounds queries with an explicit warning.
Learned: GP surrogates for engineering design need explicit domain boundary enforcement. Uncertainty inflation outside the training domain is a feature, but designers need a clear signal that they are extrapolating — not just high uncertainty values that may be ignored.