Machine Learning models for subgrid scales in turbulent reacting flows

Despite their popularity, both approaches still present major challenges such as large amounts of high-resolution data (from direct numerical simulations or experiments) during the training and a lack of generalization capability. In addition, each method presents specific limitations. For example, CNNs lack the ability to accurately reconstruct high-frequency features for out-of-sample flows and need fully supervised training. On the other hand, GANs allow for semi-supervised and fully unsupervised training, but they are computationally more expensive, and there is still not a comprehensive understanding of the discriminator's contributions. Similarly, the role of the physics-informed loss function to improve the predictive capability is largely unknown.

The project's final outcome will be the realisation of reliable predictive models able to guide the design of future thermochemical energy conversion processes for hydrogen-based and sustainable fuels (carbon-neutral fuels). Such models will be able to overcome the limitation of current turbulence-combustion models in predicting multi-regime combustion and multi-scale phenomena (e.g. intrinsic instabilities, backscattering, for example).

The main tasks of the project are:

• collection and production of training databases (DNS) for different flame regimes and fuel
• optimization of network structure with respect to the computational efficiency and generalization capability
•  enhancement of physical-informed features of the network to improve generalization capability
• investigation of the optimal training procedure
•  integration into computational fluid dynamics code
•  a-posteriori validation of the models (LES)
•  enhancement of the scalability performance to Exa-scale computing