A neural network model for the microlayer evaporation in wall boiling flows

Microlayers, the thin layers of liquid forming at the underside of a steam bubble growing at a heated wall, have shown to contribute significantly to the bubble growth under certain wall boiling conditions.
The measurement of their shape and thickness remains an experimental and the simulation of their formation via CFD a computational challenge.
Thus, it is difficult to develop microlayer evaporation models covering a wide enough parameter space, which would allow their inclusion into advanced Euler-Euler wall boiling models.
In this work we present a feed-forward neural network (NN), which was trained by a small set of direct numerical simulation (DNS) data with the aim to predict microlayer profiles and volumes under different wall boiling conditions. Various configurations of such machine learning (ML) models were studied and introduced into the OpenFOAM open source CFD solver.
The training data consists of interface-tracking simulation results of the early bubble growth stages. Using the level-set and phase-change capabilities of PHASTA the transient evolution of evaporating microlayer profiles was computed for three different superheats for water at atmospheric pressure.
Data mining was then applied to pre-process and feed these results to a neural network in order for it to learn how to predict the microlayer volume depending on different wall superheats and bubble departure sizes.
The computed microlayer-to-bubble volume ratio allowed the trained NN model to be embedded into the RPI wall boiling model of OpenFOAM, which was extended in order to account for an additional microlayer evaporation term. Whilst the overall evaporation component remains unchanged in magnitude, the proposed model does distinguish between the evaporation contributions from the upper curved bubble surface and from the microlayer region.
The NN extended RPI wall boiling model is applied to two demonstration cases: the DEBORA wall boiling case [1], for which no microlayer contribution is expected, and the experimental case of Lee et al. [2] for water under atmospheric pressure, for which the microlayer evaporation is expected to be significant. The NN extended RPI wall boiling model is shown to predict reasonable contributions of the different evaporation mechanisms.
The application of ML techniques, where experimental and computational limits hinder sufficient data collection, seems a promising alternative to the conventional development of Euler-Euler models. In the future the NN model presented here can be fed with additional DNS data as well as experimental data for more refined results under various boiling conditions and for different working fluids. The particular implementation of the ML models in the RPI wall boiling model needs to be further researched and discussed with the broad scientific community.

[1] J. Garnier, E. Manon, G. Cubizolles, Local measurments on flow boiling of refrigerant r12 in a vertical tube, Multiphase Science and Technology 13 (2001) 1–111.
[2] T. Lee, G. Park, D. Lee, Local flow characteristics of subcooled flow boil-ing flow of water in a vertical concentric annulus, International Journal of Multiphase Flow 28 (2002) 1351–1368.