Deep learning in the field of thermodynamics

Thesis prepared by Jingang Qu: "Acceleration of Numerical Simulations with Deep Learning: Application to Thermodynamic Equilibrium Calculations"

Reactive fluid transport simulation has multiple applications - flows in porous media, combustion, process engineering - and requires thermodynamic equilibrium calculations (also knows as “flash” calculations). However, these calculations can take a long time and, as they are involved in large numbers in the simulations carried out, in practice they limit the latter to systems containing few chemical species or to restricted time and space scales.

The research conducted for this thesis consisted in accelerating the two-phase flash calculations, integrating deep learning models in the existing flash algorithm [1]. The idea was to accelerate the convergence of the algorithm, thereby reducing the calculation time, without compromising the accuracy of the solution.

The objective was met, with a six-fold acceleration on existing single-core computing architectures [2, 3]. Moreover, the provision of a parallel algorithm able to function on GPU¹ -type hardware architectures led to a 100-fold time saving (figure 1).

There are numerous prospects for further research since implementation in industrial reactive transport simulators requires significant developments to adapt codes to the hardware architectures used. Moreover, extending this research to three-phase flash calculations, as well as to more complex equation of state models, is under consideration. This will make it possible to model a broader range of chemical interactions, particularly concerning aqueous systems.

Figure 1: Comparison of calculation times between CARNOT² and the PTFlash algorithm incorporating learning models for a mixture of nine chemical species.

^1- Graphics Processing Unit
^2-IFPEN library implementing the conventional algorithm

References

https://www.ifpenergiesnouvelles.com/brief/faster-flash-calculations-thanks-deep-learning
Jingang Qu, Thibault Faney, Jean-Charles de Hemptinne, Soleiman Yousef, Patrick Gallinari, PTFlash: A vectorized and parallel deep learning framework for two-phase flash calculation, Fuel, Volume 331, Part 1, 2023.
>> https://doi.org/10.1016/j.fuel.2022.125603
Jingang Qu, Thibault Faney, Jean-Charles de Hemptinne, Soleiman Yousef, Patrick Gallinari, HMOE: Hypernetwork-based Mixture of Experts for Domain Generalization, arXiv, submitted to ICCV 2023.
>> https://doi.org/10.48550/arXiv.2211.08253

Scientific contact: thibault.faney@ifpen.fr

>> ISSUE 53 OF SCIENCE@IFPEN