Lakehead University Knowledge Commons

Knowledge Commons is an open access repository for scholarship and research produced at Lakehead University. It is a free and secure repository for LU faculty, students, staff, and researchers to preserve and present their scholarship.

Communities in KnowledgeCommons

Select a community to browse its collections.

Now showing 1 - 3 of 3

Recent Submissions

Hydrologic regime drives greenhouse gas emissions and cross-assemblage convergence of soil protists and phototrophic microbiota along methane gradients in constructed mineral-soil wetlands
(2026) Haak, Alexandra; Pendea, I. Florin; Basiliko, Nathan; Wang, Jian; Levasseur, Patrick
Constructed wetlands are increasingly implemented as nature-based solutions for flood management, habitat creation, and climate mitigation. However, greenhouse gas (GHG) dynamics in these systems remain uncertain, particularly in mineral-soil wetlands where hydrologic conditions strongly influence carbon cycling processes. Understanding how hydrologic regime shapes greenhouse gas emissions and microbial community structure is therefore important for evaluating the climatic implications of wetland construction and management. Greenhouse gas fluxes and soil microbiota were examined across three emergent-vegetation wetlands in southern Ontario, Canada, including two constructed wetlands established on mineral soils with contrasting hydrologic regimes—one permanently flooded and one seasonally flooded—and a natural comparison marsh developed on organic soil. Carbon dioxide (CO₂), methane (CH₄), and nitrous oxide (N₂O) fluxes were measured using static chambers over three years (2023–2025). Soil surface microbiota assemblages were characterized from samples containing heterotrophic protists - primarily testate amoebae - and phototrophic microorganisms including diatoms, green algae, and cyanobacteria. Carbon dioxide fluxes exhibited strong seasonal patterns across all sites, increasing with temperature and peaking during summer months. In contrast, methane emissions were strongly structured by hydrologic regime. The permanently flooded constructed wetland exhibited sustained methane emissions throughout the year, including winter, whereas the seasonally flooded wetland maintained near-zero methane flux even during peak summer conditions. The natural comparison marsh showed intermediate methane emissions despite its organic substrate, reflecting shallower and more variable inundation. When expressed as CO₂-equivalent fluxes, methane accounted for most radiative forcing at the permanently flooded site but contributed minimally at the seasonally flooded site, indicating that hydroperiod exerts stronger control on climatic impact than wetland origin or substrate type. Microbial community analyses revealed that heterotrophic protist and phototrophic microbiota assemblages were structured along shared environmental gradients. Co-inertia analysis indicated significant cross-assemblage coupling, suggesting that both groups respond to integrated hydrologic and biogeochemical conditions within the wetlands. The dominant community gradient was positively associated with methane flux, indicating that microbial community organization reflects ecosystem states linked to methane production and transport, whereas no comparable relationship was observed with carbon dioxide flux. Together, the results demonstrate that hydrologic regime is a primary control on methane emissions and overall warming potential in mineral-soil constructed wetlands, while microbial community structure reflects underlying biogeochemical gradients associated with methane dynamics. These findings highlight the importance of hydroperiod in wetland design and management and contribute to understanding the climatic implications of constructed wetlands as nature-based solutions. Keywords: hydrologic regime, greenhouse gas fluxes, methane (CH₄), constructed wetlands, mineral-soil wetlands, carbon dioxide (CO₂), nitrous oxide (N₂O), testate amoebae, diatoms, co-inertia analysis, soil surface microbiota, variation partitioning
Multimodal deep learning for multi-horizon corporate revenue forecasting
(2026) Wu, Qiping; Yassine, Abdulsalam
Corporate revenue forecasting matters for valuation, portfolio management, and capital allocation. However, it is difficult because financial statements mainly reflect the past, while investors and firms often need forecasts from the next quarter to a rolling one-year horizon. This challenge becomes even greater over longer horizons, especially in fast-changing industries. This thesis addresses the problem by building a forecasting framework that starts with a broad quantitative baseline and then extends to a multimodal approach. First, this thesis develops a Temporal Fusion Transformer (TFT) baseline for next-quarter revenue forecasting across 155 continuously listed S&P 500 firms. Under a strict chronological evaluation protocol, the TFT model achieves a test Mean Absolute Percentage Error (MAPE) of 9.31%, a Root Mean Squared Error (RMSE) of 1,973 million USD, and a Mean Absolute Error (MAE) of 1,790 million USD. Controlled ablation analysis further shows that accurate short-horizon forecasting depends not only on autoregressive revenue history, but also on structured firm context, including sector identity, year-over-year growth, and firm scale variables such as total assets and equity. Second, the framework is extended from one-quarter-ahead to four-quarter-ahead forecasting. The results show that forecast accuracy deteriorates as the horizon expands, with MAPE rising from 9.31% at one quarter ahead (𝑡 + 1) to 12.07% at four quarters ahead (𝑡 + 4). A comparison with an LSTM baseline under the same chronological setting further suggests that this deterioration is not specific to a single model, but reflects a broader limitation of purely financial forecasting approaches. The effect is especially pronounced in technology-oriented firms, highlighting the limits of relying only on lagged financial data in non-linear growth environments. Third, the work proposes a multimodal TFT framework that integrates earnings-call-derived textual signals into the forecasting pipeline. Focusing on the Mega-Cap 5 companies, the framework uses both Financial Bidirectional Encoder Representations from Transformers (FinBERT) and a locally deployed Llama-3 8B model to extract finance-domain sentiment and richer generative narrative features from quarterly earnings call transcripts. These results show that transcript-based narrative features improve long-horizon forecasting. Among the models, the Llama-3 representation delivers the biggest improvement. For example, the pure TFT has a MAPE of 53.85%, while the FinBERT+TFT and Llama-3+TFT hybrids reduce it to 48.70% and 43.01%, respectively. Overall, this thesis presents a practically deployable multimodal forecasting framework that bridges the gap between backward-looking financial fundamentals and forward-looking managerial narratives in corporate revenue forecasting.
Crop disease analysis through hyperspectral images using deep learning models
(2026) Khan, Mohammed Salman; Ahmed, Saad B.; Alkhateeb, Abedalrhman; Atoofian, Ehsan
Modern precision agriculture increasingly relies on high-resolution Unmanned Aerial Vehicle (UAV) hyperspectral imagery to map diverse vegetation species and monitor complex crop health. However, processing these massively high-dimensional data cubes has historically required classical deep learning models with unsustainable computational bloat, such as heavy vision transformers or extremely deep convolutional networks. Furthermore, standard optimization pipelines routinely collapse when confronted with the complicated structural complexities of real-world agricultural datasets, which naturally feature severe class imbalances and highly overlapping spatial boundaries. This thesis directly addresses these critical computational and mathematical vulnerabilities by engineering ultra-lightweight, parameter-efficient hybrid quantum-classical architectures. By entirely replacing massive classical dense layers with a parameterized 4-qubit variational quantum circuit, this research demonstrates that quantum mechanics can natively and efficiently synthesize the highly complex, non-linear global dependencies required for accurate field classification. To overcome the distinct spatial and spectral challenges of agricultural data, this work introduces two novel evolutionary frameworks. The first, the Quantum Patch-Graph Transformer (QPGF), mathematically preserves orthogonal crop row geometry by structuring spatial patches into row-normalized 4- nearest neighbor graphs, seamlessly fusing local graph attention with quantum global feature extraction. The second methodology is the Quantum Enhanced CNN-BiSpectralMamba-Quantum architecture, which actively bypasses standard memory bottlenecks by utilizing bidirectional Mamba state-space models to aggressively process continuous spectral sequences at linear complexity. Both architectures are stabilized by a custom Hybrid Cross-Entropy and Log-Cosh Dice loss function. This highly specialized optimization pipeline strictly forces the networks to penalize dominant staple crops and accurately map the topological boundaries of rare, minority vegetation. Rigorous empirical validation on the highly imbalanced, 200-band, 30-class UAV-HSI-Crop dataset proves the absolute efficacy of these hybrid designs. The classical-quantum fusion drastically reduced the total trainable parameter count compared to state-of-the-art classical benchmarks. Despite this incredibly lightweight computational footprint, the QPGF established a robust baseline of 81.92% overall accuracy, while the advanced Quantum Enhanced CNN-BiSpectralMamba achieved a highly competitive peak of 84.83% overall accuracy and 82.07 kappa score. Ultimately, this thesis proves that fusing targeted classical spatial-sequence extractors with quantum state entanglement provides a mathematically elegant and resource efficient diagnostic engine for the future of precision agriculture.
A data-driven machine learning framework to model turbulence modulation and preferential concentration by particles
(2026) Waseem, Maryum; Tarokh, Ali; Wang, Wilson; Elyasi, Saimak
Particle-laden turbulent flows appear across a wide range of engineering and environmental systems, from spray combustion and aerosol dispersion to sediment transport and industrial multiphase reactors. Yet accurately predicting particle behaviour within these flows remains a persistent challenge, particularly in resolving how particle inertia couples to turbulent flow structures and how that coupling can be retained at computational scales where full resolution is prohibitive. This dissertation addresses both challenges through two connected studies. The first study establishes a quantitative framework linking particle inertial properties to the topological features of homogeneous isotropic turbulence that govern preferential concentration. Using high-fidelity DNS with two-way Eulerian–Lagrangian coupling on a 1283 grid at Reλ = 120, we systematically investigate the influence of the Stokes number St and the particle-to-fluid density ratio ρp/ρf on particle clustering. The centrifuge mechanism is quantitatively confirmed: clustering intensity peaks at St = 1, where particles are most effectively expelled from vortical regions and accumulate in straindominated zones identified by the Q-criterion. A non-monotonic dependence on density ratio is observed, with the maximum correlation between Q and particle number density occurring at ρp/ρf = 500, reflecting an optimal balance between inertial decoupling and fluid responsiveness. A split-correlation analysis further reveals that vortex exclusion is a robust and nearly universal mechanism, whereas strain-field accumulation exhibits stronger sensitivity to particle inertia and density. The relatively weak magnitude of the overall linear correlations demonstrates that the Q-criterion alone is insufficient to fully characterize preferential concentration, indicating that turbulence intermittency, multiscale particle–structure interactions, and trajectory memory effects also play important roles. The second study builds on these physical insights by developing a machine-learningenhanced framework for particle-laden turbulence simulations. The objective is to recover the small-scale velocity structures that are removed by filtering operations in coarseresolution simulations but are essential for accurately predicting inertial particle dynamics. To achieve this, a multilayer perceptron (MLP) model is trained using DNS data from the first study to reconstruct subgrid-scale velocity corrections from filtered flow fields. The reconstructed velocity field restores a substantial portion of the gradient intensity and high-wavenumber energy content that is absent in the filtered representation. The reconstructed field is then used to drive an Eulerian–Lagrangian particle-tracking solver with full two-way coupling, allowing particle drag forces to be consistently returned to the carrier phase as momentum source terms. Because particles evolve within a velocity field that contains reconstructed small-scale structures, the simulation recovers key features of particle-turbulence interaction that are typically suppressed in filtered simulations. In particular, the ML-enhanced framework restores stronger preferential concentration, more realistic particle slip velocities, and spatially intermittent particle source terms associated with strain-dominated regions of the flow. Together, the two studies form a coherent progression. The first establishes the physical mechanisms governing particle clustering and identifies the flow-topology conditions under which preferential concentration occurs. The second demonstrates that these critical small-scale flow structures can be reconstructed using a data-driven model, enabling improved particle tracking and two-way coupling predictions in simulations where full DNS resolution is not computationally feasible. The combined framework therefore provides a computationally efficient pathway toward higher-fidelity prediction of particle-laden turbulent flows, with direct relevance to industrial and environmental multiphase systems.
Energy-efficient deep neural networks using Domainwall Memory cache for general-purpose graphics processing units
(2026) Namvarmotlagh, Alireza; Atoofian, Ehsan; Deng, Yong; Wu, Kaiyu; Zhou, Yushi
Deep neural networks (DNNs) have become the dominant computational paradigm across computer vision, natural language processing, and generative modelling, yet achieving state-of-the-art accuracy increasingly requires models with billions of parameters and commensurately large memory footprints. These models place extreme bandwidth and capacity demands on the on-chip memory hierarchy of modern general-purpose graphics processing units (GPGPUs), making the shared L2 cache a major contributor to energy consumption. At the same time, aggressive SRAM scaling leads to rapidly increasing leakage power, presenting a fundamental challenge for future high-performance computing architectures. Domain Wall Memory (DWM) is a promising alternative for large on-chip caches due to its ultra-high density and near-zero leakage, but its shift-based access mechanism introduces variable and often high access latency that must be addressed before practical deployment. This thesis presents a hardware–software co-design framework that integrates a DWM-based L2 cache into tensor core (TC)-equipped GPGPUs while mitigating DWM’s shift penalty. On the hardware side, the conventional SRAM data array is replaced with DWM, and tape-head prediction policies are employed that proactively reposition track heads based on predicted access patterns. A hybrid predictor combining stride and two-level context-based prediction achieves the lowest shift overhead among all evaluated strategies. On the software side, structured pruning is applied to representative CNN and transformer models to reduce parameter count and regularize memory accesses, and TC-optimized kernels are implemented that efficiently exploit the pruned structures. Across a suite of seven convolutional and attention-based DNN models, pruned DWM-based L2 caches achieve an average energy saving of 73.2% compared to an unpruned SRAM-based L2 cache, while delivering an average performance improvement of 13.5% that effectively mitigates performance degradation across all evaluated models. Under iso-area conditions, the DWM-based L2 cache achieves 17× more capacity than SRAM, enabling it to outperform SRAM by 7% to 37.4% in execution time and reduce energy consumption by 53.3% to 71.6%. The resulting Energy–Delay Product (EDP) of SRAM is 2:3× to 4:58× higher than that of DWM. These results demonstrate that carefully cooptimizing emerging non-volatile memories at both software and hardware levels can deliver energy-efficient DNN acceleration without sacrificing performance.