arXiv subjects

Mohammad Umair

Publications and source records attributed to Mohammad Umair.

4 recordsLinked to original sources

arXiv2026

A Hybrid Generative Reduced-Order Model for the Minimal Flow Unit

A data-driven reduced-order modelling framework is proposed for wall-bounded turbulent flows to forecast the intermittent near-wall dynamics over extended time horizons from sparse sensor measurements. The approach combines a $\beta$-VAE-GAN, which compresses high-dimensional flow fields into a low-dimensional latent space, with a sensor-conditioned Transformer that forecasts the evolution of the latent variables. The temporal module employs Easy Attention, a static time-mixing operator that replaces the learnable query-key mechanism of standard self-attention at reduced computational cost, combined with an adapted AdaLN-Zero modulation mechanism for sensor-based conditioning. Evaluated on the Minimal Flow Unit ($Re_\tau = 200$) at $y^+ = 14$, the compression stage recovers $87\%$ of the turbulent kinetic energy within a four-dimensional latent space, exceeding the standard $\beta$-VAE baseline by more than $10\%$. The latent dimensions autonomously encode the characteristic timescales of the flow, with specific coordinates capturing the low-frequency signature of the near-wall regeneration cycle ($T^+ \approx 1724$), establishing the physical interpretability of the learnt representation. The sensor-conditioned Transformer maintains accurate forecasts over $17{,}288\,t^+$ from an initialisation window of only $128\,t^+$, whilst end-to-end inference reconstructs $82\%$ of the turbulent kinetic energy. The principal limitation is the attenuation of rare, extreme-amplitude events, a consequence of the encoder prioritising the most statistically recurrent flow states within the low-dimensional bottleneck. Nevertheless, the framework accurately reproduces the alternating active and quiescent phases of the regeneration cycle, demonstrating its suitability as a surrogate model for the intermittent dynamics of wall-bounded turbulence.

physics.flu-dyn↗

arXiv2026

Multi-Partner Project: Multi-GPU Performance Portability Analysis for CFD Simulations at Scale

As heterogeneous supercomputing architectures leveraging GPUs become increasingly central to high-performance computing (HPC), it is crucial for computational fluid dynamics (CFD) simulations, a de-facto HPC workload, to efficiently utilize such hardware. One of the key challenges of HPC codes is performance portability, i.e. the ability to maintain near-optimal performance across different accelerators. In the context of the \textbf{REFMAP} project, which targets scalable, GPU-enabled multi-fidelity CFD for urban airflow prediction, this paper analyzes the performance portability of SOD2D, a state-of-the-art Spectral Elements simulation framework across AMD and NVIDIA GPU architectures. We first discuss the physical and numerical models underlying SOD2D, highlighting its computational hotspots. Then, we examine its performance and scalability in a multi-level manner, i.e. defining and characterizing an extensive full-stack design space spanning across application, software and hardware infrastructure related parameters. Single-GPU performance characterization across server-grade NVIDIA and AMD GPU architectures and vendor-specific compiler stacks, show the potential as well as the diverse effect of memory access optimizations, i.e. 0.69$\times$ - 3.91$\times$ deviations in acceleration speedup. Performance variability of SOD2D at scale is further examined on the LUMI multi-GPU cluster, where profiling reveals similar throughput variations, highlighting the limits of performance projections and the need for multi-level, informed tuning.

cs.DC↗

arXiv2025

Effect of filter kernel on scale-energetics of near-wall turbulent structures

Inter-scale energy fluxes, $Π^λ$, are widely used as a diagnostic tool to analyse energy transfer across length scales, $λ$, in turbulence data. Here, we investigate how the choice of filter kernel (sharp spectral, Gaussian, box) affects the computed energy fluxes at constant filter width. We apply spatial filtering to a turbulent pipe flow simulation dataset and assess the effect on the local structure of $Π$. While the mean energy flux profile at each wall-normal distance is qualitatively robust across kernels, we observe significant differences in the intensity and spatial distribution of localised $Π$ events. Correlations between typical flow structures in the buffer layer (streaks, vortices, and Q-events) and regions of forward/backward transfer in the instantaneous $Π$ field differ markedly between kernel types. Cross-correlations appear strongly upstream--downstream symmetric when using the sharp spectral kernel, but asymmetric for the Gaussian and box kernels. For the Gaussian and box kernels $Π$ events tend to localise along the inclined meander of streaks, while they are centred on top of the streaks for the sharp spectral kernel. Moreover, using the sharp spectral kernel, we observe a coincidence of backward scatter and fluid transport away from the wall ($Q_1$), which does not appear with the Gaussian and box kernels. All kernels, however, predict backward scatter directly downstream of $Q_1$ events. The results suggest that interpretations of inter-scale energy flux based on sharp spectral scale separation should be treated with caution, since such kernels act non-local in physical space, whereas $Π$ events are inherently localised. Our python post-processing tool eFlux for scale separation and energy flux analysis in pipe flows is freely available and readily adaptable to other flow configurations and filter widths.

physics.flu-dyn↗

arXiv2021

Transferring Semantic Knowledge Into Language Encoders

We introduce semantic form mid-tuning, an approach for transferring semantic knowledge from semantic meaning representations into transformer-based language encoders. In mid-tuning, we learn to align the text of general sentences -- not tied to any particular inference task -- and structured semantic representations of those sentences. Our approach does not require gold annotated semantic representations. Instead, it makes use of automatically generated semantic representations, such as from off-the-shelf PropBank and FrameNet semantic parsers. We show that this alignment can be learned implicitly via classification or directly via triplet loss. Our method yields language encoders that demonstrate improved predictive performance across inference, reading comprehension, textual similarity, and other semantic tasks drawn from the GLUE, SuperGLUE, and SentEval benchmarks. We evaluate our approach on three popular baseline models, where our experimental results and analysis concludes that current pre-trained language models can further benefit from structured semantic frames with the proposed mid-tuning method, as they inject additional task-agnostic knowledge to the encoder, improving the generated embeddings as well as the linguistic properties of the given model, as evident from improvements on a popular sentence embedding toolkit and a variety of probing tasks.

cs.CL↗