Load-Balanced Neural-Network-Aided Adaptive Tabulation of Phase Equilibria for Transcritical Multiphase Flows
Abstract
Simulating high-pressure transcritical flows with phase change is critical for advanced propulsion, yet traditional first-principles approaches, vapor-liquid equilibrium (VLE) calculations, consume most of the runtime and frequently fail to converge. Standalone machine learning models introduce heavy evaluation overhead across large grids, while dynamic tabulation still relies on fragile iterative solvers. To resolve these limitations, this study introduces a hybrid In-Situ Adaptive Tabulation and Artificial Neural Network (ISAT-ANN) framework. A pre-trained ANN replaces the iterative solver during table expansion to guarantee crash-free operation, while ISAT enables rapid data retrieval via local linear interpolation. To eliminate parallel load imbalances, an O(N) greedy load-balancing algorithm redistributes computational tasks across cores. Coupled with Adaptive Mesh Refinement (AMR), the framework was validated on a Mach 5 shock-droplet interaction benchmark. It accurately captured shock-interface dynamics without instability, achieving a 2.30 times speedup in thermodynamic calculations on a single core and 2.62 times in parallel over standalone ANNs, with AMR providing an additional 2.92 times overall efficiency boost. This plug-and-play framework offers a stable, scalable path for high-fidelity multiphase flow simulations.
Related
- Phase equilibria in MnSb2Te4 – GeSb2Te4 system and magnetic properties of Mn1-xGexSb2Te4 solid solutions
- Seifert type magnetohydrostatic equilibria in rotationally symmetric toroidal domains
- Quantum-enhanced multiphase estimation with spin-1 Bose-Einstein condensates
- Analytic non-axisymmetric 3D MHD equilibria without stellarator symmetry
- An electromechanically coupled multiphase-field model with generalized kinetic relations for ferroelectrics
BibTeX
@misc{srinivasan2026loadbalanced,
title = {{Load-Balanced Neural-Network-Aided Adaptive Tabulation of Phase Equilibria for Transcritical Multiphase Flows}},
author = {Navneeth Srinivasan and Po-Yu Lai and Suo Yang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.26984},
eprint = {2610.26984},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26984},
}
RIS
TY - UNPB
TI - Load-Balanced Neural-Network-Aided Adaptive Tabulation of Phase Equilibria for Transcritical Multiphase Flows
AU - Navneeth Srinivasan
AU - Po-Yu Lai
AU - Suo Yang
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2610.26984
UR - https://arxiv.org/abs/2610.26984
AB - Simulating high-pressure transcritical flows with phase change is critical for advanced propulsion, yet traditional first-principles approaches, vapor-liquid equilibrium (VLE) calculations, consume most of the runtime and frequently fail to converge. Standalone machine learning models introduce heavy evaluation overhead across large grids, while dynamic tabulation still relies on fragile iterative solvers. To resolve these limitations, this study introduces a hybrid In-Situ Adaptive Tabulation and Artificial Neural Network (ISAT-ANN) framework. A pre-trained ANN replaces the iterative solver during table expansion to guarantee crash-free operation, while ISAT enables rapid data retrieval via local linear interpolation. To eliminate parallel load imbalances, an O(N) greedy load-balancing algorithm redistributes computational tasks across cores. Coupled with Adaptive Mesh Refinement (AMR), the framework was validated on a Mach 5 shock-droplet interaction benchmark. It accurately captured shock-interface dynamics without instability, achieving a 2.30 times speedup in thermodynamic calculations on a single core and 2.62 times in parallel over standalone ANNs, with AMR providing an additional 2.92 times overall efficiency boost. This plug-and-play framework offers a stable, scalable path for high-fidelity multiphase flow simulations.
ER -