Detailed chemical kinetics remains a principal bottleneck in reactive computational fluid dynamics (CFD), as integrating the stiff system of ordinary differential equations for many species at every grid point and time step often dominates the total CPU-time. To alleviate this cost, we present a data-driven surrogate modeling approach that learns the time-advancement operator mapping the current thermochemical state to its next state after a time step. An unsupervised Self-Organizing Map (SOM) is used to automatically partition the high-dimensional composition space into N topologically ordered regions. Within each region, a feed-forward neural network is trained to predict all species mass fractions at the next time step, effectively specializing on a particular subset of the compositional space. We evaluate the proposed surrogate on two canonical reactive systems. (i) In a zero-dimensional perfectly stirred reactor (PSR) with a detailed 19-species hydrogen–NOx mechanism; its performance is also compared against a physics-informed neural network (PINNs). (ii) A two-dimensional laminar hydrogen–air flame with an 8-species hydrogen combustion kinetic mechanism. In the 0D case, the SOM-partitioned neural ensemble accurately predicts the next-step thermochemical state in agreement with direct integration of the detailed kinetics. In the 2D case, accuracy is limited by rollout length: consistent up to about 200–300 steps, then deteriorating past 500 steps, predominantly downstream of the flame front.
La cinetica chimica dettagliata resta un collo di bottiglia primario nella CFD reattiva: l’avanzamento, in ogni punto di griglia e a ogni passo temporale, di sistemi multi-specie con comportamento rigido spesso domina il tempo computazionale complessivo. Per ridurre il tempo computazionale associato all’avanzamento chimico, proponiamo un modello surrogato guidato dai dati che apprende l’operatore di avanzamento temporale in grado di mappare lo stato termo-chimico corrente nello stato dopo un intervallo di tempo. Una Self-Organising Map (SOM) partiziona automaticamente lo spazio composizionale ad alta dimensionalità in N regioni ordinate topologicamente; all’interno di ciascuna regione, una rete neurale è allenata a predire le frazioni di massa di tutte le specie al passo successivo, specializzandosi su un sottoinsieme dello spazio termochimico. Valutiamo il surrogato su due sistemi reattivi canonici: (i) un reattore perfettamente miscelato zero-dimensionale (PSR) con un meccanismo dettagliato a 19 specie per idrogeno–NOx, confrontato anche con una physics-informed neural network (PINN); e (ii) una fiamma laminare bidimensionale idrogeno–aria con un meccanismo cinetico a 8 specie per l’idrogeno. Nel caso 0D, l’ensemble partizionato tramite SOM predice accuratamente lo stato termochimico al passo successivo in accordo con l’integrazione diretta. Nel caso 2D, l’accuratezza è limitata dalla lunghezza del rollout: il comportamento rimane fisicamente consistente fino a circa 200–300 passi, per poi degradare oltre i 500 passi, prevalentemente a valle del fronte di fiamma. Pur non costituendo una soluzione definitiva, il metodo offre una via competitiva per l’accelerazione della CFD reattiva, in particolare con meccanismi dettagliati estesi.
A neural network based chemistry integrator for reactive flow simulations
Forte, Riccardo
2024/2025
Abstract
Detailed chemical kinetics remains a principal bottleneck in reactive computational fluid dynamics (CFD), as integrating the stiff system of ordinary differential equations for many species at every grid point and time step often dominates the total CPU-time. To alleviate this cost, we present a data-driven surrogate modeling approach that learns the time-advancement operator mapping the current thermochemical state to its next state after a time step. An unsupervised Self-Organizing Map (SOM) is used to automatically partition the high-dimensional composition space into N topologically ordered regions. Within each region, a feed-forward neural network is trained to predict all species mass fractions at the next time step, effectively specializing on a particular subset of the compositional space. We evaluate the proposed surrogate on two canonical reactive systems. (i) In a zero-dimensional perfectly stirred reactor (PSR) with a detailed 19-species hydrogen–NOx mechanism; its performance is also compared against a physics-informed neural network (PINNs). (ii) A two-dimensional laminar hydrogen–air flame with an 8-species hydrogen combustion kinetic mechanism. In the 0D case, the SOM-partitioned neural ensemble accurately predicts the next-step thermochemical state in agreement with direct integration of the detailed kinetics. In the 2D case, accuracy is limited by rollout length: consistent up to about 200–300 steps, then deteriorating past 500 steps, predominantly downstream of the flame front.| File | Dimensione | Formato | |
|---|---|---|---|
|
Executive_Summary___Riccardo_Forte_POLIMI.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
2.05 MB
Formato
Adobe PDF
|
2.05 MB | Adobe PDF | Visualizza/Apri |
|
Thesis_Manuscript_Riccardo_Forte.pdf
accessibile in internet per tutti
Descrizione: Thesis Manuscript
Dimensione
31.54 MB
Formato
Adobe PDF
|
31.54 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243869