Seminar on Machine learning augmented massively parallel flow solvers | Fri 26 Sep @ 2:15 pm | ME Auditorium
Venue:
ME Auditorium
September 26, 2025
Seminar on "Machine learning augmented massively parallel flow solvers: application to combustion", by Prof. K. Aditya from the Department of Computational and Data Sciences, IISc, on Friday, 26th Sept 2025 at 2:15 pm. The venue is Mech Eng Department Auditorium.
The details of the talk and bio of the speaker are given below.
Machine learning augmented massively parallel flow solvers: application to combustion
Konduri Aditya
Assistant Professor
Department of Computational and Data Sciences
Indian Institute of Science
High-fidelity direct numerical simulations (DNS) of turbulent combustion are often performed to gain fundamental insights into flow–chemistry interactions under conditions relevant to practical engines. These simulations solve highly nonlinear partial differential equations, which require massive computations on large supercomputers. A key challenge is to perform the simulations efficiently and in a scalable manner. In this talk, we introduce two methods that can significantly improve the scalability of DNS solvers: First, an asynchronous computing method that significantly minimizes the data movement costs at extreme scales. Second, a low-dimensional manifold is used to reduce the chemistry computation costs.
Current state-of-the-art direct numerical simulations are routinely performed on hundreds of thousands of processing elements (PEs). At an extreme scale, it was observed that data movement and its synchronization pose a bottleneck to the scalability of solvers. We introduce an asynchronous computing method that relaxes communication synchronization at the mathematical level and has shown significant promise in improving the scalability of PDE solvers. In this method, communication synchronization between PEs owing to halo exchanges is relaxed, and computations proceed regardless of the communication status. It was shown that the numerical accuracy of standard schemes, such as finite differences implemented with relaxed communication synchronization, is significantly affected. Subsequently, new asynchrony-tolerant schemes have been developed to compute accurate solutions and demonstrate good scalability. This section presents an overview of the status of asynchronous computing methods for PDE solvers and their applicability to exascale simulations.
Identifying low-dimensional manifolds (LDMs) to represent the thermochemical state in reacting flows is crucial for significantly reducing the computational costs. Widely used principal component analysis (PCA) achieves this by obtaining an eigenvector basis for the LDM through eigenvalue decomposition of the data covariance matrix. However, this may not effectively capture the stiff chemical dynamics when the reaction zones are localized in space and time. Alternatively, we propose a co-kurtosis PCA (CoK-PCA), wherein the principal components are obtained from the singular value decomposition (SVD) of the matricized co-kurtosis tensor. The efficacy of the CoK-PCA-based reduced manifold was assessed by simulating spontaneous ignition in a homogeneous reactor. The time-evolved profiles of the PCs and reconstructed thermochemical scalars demonstrate the robustness of the CoK-PCA-based low-dimensional manifold in accurately capturing the ignition process. The results of this study show the potential of CoK-PCA-based manifolds to be implemented in massively parallel reacting flow solvers.
Biography: Konduri Aditya is an Assistant Professor in the Department of Computational and Data Sciences at the Indian Institute of Science, Bengaluru, where he leads the FLAME Laboratory. His research spans computational fluid dynamics, turbulent combustion and propulsion, high-performance computing, and scientific machine learning.