TY - GEN

T1 - Optimization of Computations in Simulating Chemical Kinetics Using Surrogate Neural Models

AU - Kazakov, Gleb

AU - Ivanov, Kirill

AU - Penenko, Alexey

AU - Marchenko, Mikhail

N1 - Conference code: 20

PY - 2024

Y1 - 2024

N2 - We investigate how deep learning-based methods can be applied to solve stiff systems of ordinary differential equations modeling complex chemical kinetic systems. We use artificial neural networks (ANN) as a time-step integrator: ANN reproduces the evolution of a given chemical state for a short period of time. Thus, it can potentially replace computationally expensive stiff solvers, giving a substantial speed-up. This improvement is crucial in simulating a large number of simultaneous reactions that can happen, for example, in almost all combustion problems in aircraft and rocket jet engines, industrial installations for the pyrogenic synthesis of nanomaterials, processes in chemical and technological industrial installations, for example, in chemical reactors. In order to achieve an acceptable improvement in inference time while maintaining prediction accuracy, different techniques are applied to help ANNs assimilate chemical system structure: clustering of phase space and auto-regression. The last one is enhanced with curriculum learning methods. The described approach is tested on a basic Robertson problem. The approach may be promising to further speed up of numerical modeling while maintaining acceptable accuracy of ODE solution.

AB - We investigate how deep learning-based methods can be applied to solve stiff systems of ordinary differential equations modeling complex chemical kinetic systems. We use artificial neural networks (ANN) as a time-step integrator: ANN reproduces the evolution of a given chemical state for a short period of time. Thus, it can potentially replace computationally expensive stiff solvers, giving a substantial speed-up. This improvement is crucial in simulating a large number of simultaneous reactions that can happen, for example, in almost all combustion problems in aircraft and rocket jet engines, industrial installations for the pyrogenic synthesis of nanomaterials, processes in chemical and technological industrial installations, for example, in chemical reactors. In order to achieve an acceptable improvement in inference time while maintaining prediction accuracy, different techniques are applied to help ANNs assimilate chemical system structure: clustering of phase space and auto-regression. The last one is enhanced with curriculum learning methods. The described approach is tested on a basic Robertson problem. The approach may be promising to further speed up of numerical modeling while maintaining acceptable accuracy of ODE solution.

KW - Stiff chemical kinetics

KW - auto-regression

KW - clustering

KW - curriculum learning

KW - deep learning

KW - neural networks

KW - power transform

UR - https://www.scopus.com/record/display.uri?eid=2-s2.0-85208922644&origin=inward&txGid=f0d820b8a09a83a98afdc32bc759baab

UR - https://www.mendeley.com/catalogue/34c75b6e-a4b9-3bf2-b381-9a3bd267fa9c/

U2 - 10.1109/OPCS63516.2024.10720440

DO - 10.1109/OPCS63516.2024.10720440

M3 - Conference contribution

SN - 9798331517625

T3 - Proceedings - 2024 20th International Asian School-Seminar on Optimization Problems of Complex Systems, OPCS 2024

SP - 60

EP - 64

BT - Proceedings - 2024 20th International Asian School-Seminar on Optimization Problems of Complex Systems, OPCS 2024

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 20th International Asian School-Seminar on Optimization Problems of Complex Systems

Y2 - 19 July 2024 through 30 July 2024

ER -