Hybrid modelling for hydrogen combustion: Accelerating LES with adaptive uncertainty-aware LEM and neural ROMs

Descrizione del progetto

Simulazione della combustione dell’idrogeno per raggiungere la decarbonizzazione

La combustione dell’idrogeno può ridurre le emissioni di carbonio nei trasporti, nell’industria e in altri settori ad alto consumo energetico, ma una simulazione accurata rimane costosa poiché le fiamme di idrogeno comportano fenomeni complessi di turbolenza e diffusione differenziale. Con il sostegno del programma di azioni Marie Skłodowska-Curie, il progetto H2SAFIRE sta sviluppando un quadro di modellizzazione ibrido che combina la simulazione dei grandi vortici (LES, Large Eddy Simulation), un modello dei vortici lineare adattivo (LEM, Linear Eddy Model) e un modello di ordine ridotto basato su una rete neurale profonda (ROM-DNN, Reduced Order Model Based on a Deep Neural Network). Il surrogato ROM-DNN sostituisce i costosi calcoli LEM quando sono affidabili, mentre un gate sensibile all’incertezza ricorre al LEM quando necessario per preservare l’accuratezza. Il quadro di riferimento sarà convalidato su fiamme canoniche e simili a quelle industriali. H2SAFIRE si propone di fornire simulazioni più rapide e affidabili della combustione dell’idrogeno, favorendo così soluzioni energetiche più pulite, riducendo i costi e rafforzando la leadership europea nelle tecnologie dell’idrogeno.

Obiettivo

Reaching climate neutrality by 2050 requires deep decarbonisation of transport, industry and other energy-intensive sectors where fossil fuels still dominate. Hydrogen is a zero-carbon energy carrier whose safe, efficient use depends on fast and trustworthy predictions of turbulent flames. For large-scale devices, however, high-fidelity Large Eddy Simulation (LES) is costly because hydrogen combustion couples multiscale turbulence with differential diffusion and intermittent local ignition and extinction. H2SAFIRE proposes a hybrid modelling framework that couples LES with a Linear Eddy Model (LEM) for micro scale chemistry and a Reduced Order Model based Deep Neural Network (ROM-DNN) surrogate trained on LEM data. The surrogate replaces the costly chemistry step when reliable; a calibrated uncertainty gate monitors confidence at run time and reverts to LEM whenever uncertainty exceeds a threshold, preserving physical fidelity.
The project pursues three objectives. First, design and validate an adaptive LES-LEM solver tailored to hydrogen combustion. Second, construct a physics-guided LES-LEM-ROM-DNN framework in which ROMs compress LEM outputs and a DNN predicts modal dynamics in loop to supply sub-grid source terms. Third, evaluate accuracy, robustness, generalisation and efficiency on canonical and semi-industrial flames, targeting at least a ten-fold speed-up at matched fidelity. Validation uses a jet-in-hot-coflow flame and an industrial-like furnace.
H2SAFIRE will release an open solver, FAIR datasets, benchmark cases and practical guidelines for hydrogen combustion simulations. By enabling faster and reliable predictions, the project can lower modelling cost and energy use, support cleaner industrial heat and transport, and strengthen European capability for hydrogen technologies. Training at UPM and a secondment at ULB ensure rigorous validation and knowledge transfer. The outcomes align with the European Green Deal and the UN Sustainable Development Goal.