ResearchSpace

Communities in ResearchSpace

Welcome to ResearchSpace, the institutional repository of the CSIR. ResearchSpace is an open access electronic archive collecting, preserving and distributing scholarly digital materials created by the CSIR.

Recent Submissions

Item

High-performance software-defined radio spectrometer: A GPU implementation for multi-GHz radio astronomical signal processing

(2025-12) Gomomo, Lwando; Gaffar, YA; Winberg, S

This paper presents a GPU-accelerated polyphase filterbank (PFB) spectrometer enhanced with spectral kurtosis (SK) capability for real-time operation in radio astronomical signal processing. Modern wideband radio telescopes employ analog-to-digital converters (ADCs) that sample signals at multi GHz rates, generating tens of gigabits per second (Gbps) of data per antenna that demand high-speed processing while operating in radio frequency interference (RFI)-contaminated environments. Existing solutions often separate FPGA- or GPU based spectrometer backends and RFI detection systems, but this work unifies both within a single GPU pipeline. Using NVIDIA’s PyCUDA framework on a Tesla T4 GPU, the system employs a four-tap Hamming-windowed PFB for channelisation, an averager and a higher-order spectral moment estimator based on the SK algorithm for automated RFI detection in the mea sured spectrum. The implementation achieved 8.88 GSamples/s throughput with high spectral fidelity. Tests using radio telescope data confirmed accurate detection of stationary and transient interference while maintaining false-alarm rates only 1.3% above theoretical expectation. This research demonstrates that modern GPUs can enable efficient and flexible high-throughput spectrom eter architectures for advanced digital signal processing (DSP) in radio astronomy.

Item

Not all CO2 is equal: Source-specific constraints and viability trade-offs in methanol synthesis from industrial emissions

(2026-05) Macheli, L; Mukeru, BM; Duma, Zama G; Patel, B; Jewell, LL

Methanol synthesis from captured CO2 is widely regarded as a promising pathway for carbon utilization, yet its feasibility depends heavily on the characteristics and constraints of the CO2 source. This review evaluates four industrial point sources—biogas, steel plants, cement kilns, and waste-to-energy facilities—highlighting key differences in CO2 purity, contaminant load, hydrogen integration, and catalyst stability. We propose a five-axis viability framework, developed through a synthesis of current literature, to structure source-specific comparison and guide system-level evaluation. The framework includes CO2 usability, hydrogen vulnerability, contaminant burden, integration potential, and policy exposure. By applying this structured lens, the review identifies key performance-limiting trade-offs, techno-economic constraints, and integration barriers across point sources. Results show that biogas and steel off-gases offer favourable trade-offs (scores of 15–18/25), while cement and waste-to-energy streams face major integration and degradation challenges (≤9/25). Reforming pathways, gas conditioning requirements and modular deployment considerations are also discussed. This review concludes that effective CO2-to-methanol deployment requires source-specific process design, improved ontaminant-tolerant catalysts, and better alignment of infrastructure and policy to the heterogeneous nature of industrial CO2 sources.

Item

AI in the antenna design life cycle: An AI-assisted systematic review

(2025-12) Mafa, Ike M; Winberg, S

The increasing integration of Artificial Intelligence (AI) in antenna design presents opportunities for optimizing the design process. This paper presents a comprehensive AI-assisted systematic review of the use of AI in the antenna design life cycle. A Systematic Mapping Study (SMS) is first employed to map the landscape of AI applications in antenna design, identifying key research areas, trends, and gaps. The SMS categorizes studies into three primary domains: AI for Optimization, AI for Synthesis, and AI for Performance Prediction. In the second phase, a Systematic Literature Review (SLR) is conducted to critically evaluate and synthesize empirical findings from the selected studies. The SLR examines the various AI techniques used in each step of the lifecycle. The results from both methods provide a comprehensive overview of AI’s current applications in the field, highlighting key challenges and opportunities for future research. This hybrid approach, combining AI-assisted SMS and SLR, offers both a broad mapping of AI applications and a deep, critical analysis of their real world effectiveness, making it a valuable contribution to the field of AI-enhanced antenna design and AI-assisted systematic reviews.

Item

Beyond posters: A user-centric digital twin framework for cybersecurity awareness

(2026-03) Makharamedzha, Fhatuwani; Baloyi, Errol; Mmbodi, Rendani; Hlongwane, Ndabezinhle E

Traditional cybersecurity awareness (CSA) methods, such as posters, flyers, and static training modules often fail to engage users or drive lasting behavioural change. To address these limitations, this paper proposes a novel, user-centric approach to CSA using Digital Twin (DT) technology integrated with machine learning (ML). The proposed framework introduces the concept of a User-Centric Digital Twin (UCDT)-CSA, a dynamic digital replica of each user modelled on their cybersecurity knowledge, behaviours, and risk profile. While UCDTs have been applied in domains such as construction, aquaculture, and healthcare, this work pioneers their use in the cybersecurity context. The system begins with a pre-assessment to capture individual user responses, which are used to configure a personalized training path. Through ongoing interaction with adaptive simulations and scenario-based learning, the UCDT-CSA evolves in real time, enabling training that continuously adjusts to user performance and behaviour. ML models analyse these interactions to refine each twin’s profile, delivering increasingly targeted content and interventions aimed at improving secure behaviours. This approach transforms CSA from a static, compliance-focused exercise into an engaging, data-driven, and behaviourally adaptive learning experience. The paper outlines the architecture of the UCDT-CSA framework, discusses key implementation considerations, and sets the stage for future empirical validation and deployment in government, Small and Medium-Sized Enterprises (SMEs) and academic environment.

Item

Pipeline for efficient quantization of large language models for resource-constrained deployment

(2025-12) Tchiwewe@csir.co.za; Onumanyi, Adeiza J

Recent years have seen breakthroughs in large language models (LLMs), such as the GPT and LLaMA family of models that have transformed natural language processing. Despite this, their considerable computational and memory requirements inhibit their deployment in edge and mobile environments. In this paper we introduce a modular quantization pipeline that reduces the memory footprint of LLMs while preserving core performance. We evaluate the basic and advanced quantization techniques, including Absolute Maximum Quantization, Zero-Point Quantization, GPTQ, and NF4, and make use of popular toolkits that include bitsandbytes and AutoGPTQ. Experimental results on representative tasks show that 4- and 8- bit quantized models can be run on commodity GPUs and CPUs with acceptable quality loss. Our experiments demonstrated up to 8x compression, making our pipeline suitable for LLM deployment in both edge and industrial scenarios.