Wildfire spread prediction with deep learning surrogates
Physics-based wildfire simulators are interpretable, but each additional ensemble member adds cost. We trained deep learning surrogates on 10,584 fire spread simulations at 2 m resolution for the Rectoret region in Catalonia and compared a patch-based U-Net, a transfer-learned ResNet-50, a physics-informed network and a Swin-Unet transformer. Surface fuel load was the only terrain or vegetation variable with real predictive strength and reduced prediction error by 21%. Applied without retraining to a second region, Pedriza, the convolutional models still predicted fire spread with only a small loss of accuracy.
Scaling agent-based simulation of human migration
The Flee code forecasts population displacement in armed conflicts. We parallelised it and analysed its complexity: the code scaled to more than 16,384 cores with 100 million agents on the Hawk pre-exascale system, and to 8,192 cores on AMD EPYC Rome and Intel Xeon systems, with very detailed location graphs adding initialisation overhead.
Coupling models with data, and performance at scale
Coupling simulation models with data sources improves the accuracy of global-challenge simulations on data-centre infrastructures. The performance side, including CFD for urban air pollution on AMD EPYC clusters, is described on performance and energy efficiency of CFD.
AGCA workshop series
With Łukasz Szustak I co-organise the Workshop on Advancements of Global Challenges Applications (AGCA) at the PPAM conference, a forum on high-scale HPC, HPDA and AI simulations for global challenges (2024, Ostrava; 2026, Poznań).
Publications
- Interpretable Patch-Based Deep Learning for Wildfire Spread Prediction from Ensemble Simulations (arXiv preprint, 2026)
- Evaluating AMD EPYC CPU architectures on CFD applications (urban air pollution model) (Future Generation Computer Systems, 2026)
- Large-scale parallelization of Human Migration Simulation (IEEE Transactions on Computational Social Systems, 2024)
- P-FLEE: An Efficient Parallel Algorithm for Simulating Human Migration (IEEE IPDPS Workshops, 2021)
- Towards Accurate Simulation of Global Challenges on Data Centers Infrastructures via Coupling of Models and Data Sources (ICCS 2020, Lecture Notes in Computer Science)
Projects: HiDALGO2 (project coordinator), HiDALGO, CoeGSS. Project deliverables: HiDALGO2, HiDALGO, CoeGSS.
Other research topics: Performance and energy of CFD on AMD EPYC · Divisible load scheduling · IT for natural history collections · Virtual Laboratory and remote instrumentation · all research themes.
Go to top of the page.