Deploying convolutional neural networks (CNNs) on resource-constrained field-programmable gate arrays (FPGAs) remains challenging because hardware cost is often driven not only by parameter count or numerical precision, but also by intermediate activation size and architectural structure. This paper presents a hardware–aware CNN optimization study for image classification on the ZedBoard platform using a Keras—hls4ml—Vivado workflow. A baseline CNN is compared with post–training quantization, quantization–aware training, structured pruning, and activation–aware architectural redesign under the same evaluation framework. The results show that quantization alone provides limited hardware benefit, while structured pruning reduces resource utilization but remains insufficient to satisfy all ZedBoard constraints. In contrast, the final activation- -aware architecture is the first feasible design among the investigated branches. At the HLS synthesis level, it achieves 93.03% test accuracy with 183 BRAM_18K, 201 DSP48E, 40859 FF, 44079 LUT, and 20611 latency cycles. The final accelerator is then implemented in Vivado and physically validated on the ZedBoard. The post–route design uses 34505 LUTs, 35979 FFs, 158 BRAM_18K, and 177 DSP48E1, meets timing at 50 MHz with WNS = +1.547 ns, and achieves 92.97% hardware accuracy on the complete 10000–image MNIST test set with zero execution timeouts and zero BRAM access failures. These results demonstrate that activation-aware topology redesign is an effective strategy for achieving deployable CNN inference on constrained FPGA platforms.
In recent years, the building sector has become a central focus in global decarbonization efforts, highlighting the limitations of traditional energy-only assessments and the need for time-resolved approaches to quantify operational emissions. In the European Union, the EPBD IV revision further reinforces this need by introducing Zero-emission Buildings (ZeBs) and operational-GHG targets. To avoid ambiguity between regulatory compliance and carbon–neutral accounting, this work adopts a dual reporting framework that distinguishes (i) an EPBD-oriented operational emission intensity indicator and (ii) a separate net-emissions balance including explicitly defined compensation terms.An integrated methodology is proposed to assess ZeB performance by formalizing system boundaries, temporal resolution, and emissions accounting rules. The approach is demonstrated on a fully electric university building (946 m2 conditioned area) equipped with on-site PV generation (40.32 kW) and monitored from 2021 to 2024.Results show that changing the temporal resolution of input data and conversion factors can alter annual net-balance estimates by up to 24 %, with a concurrent reduction of the operational emission intensity from 93.2 to 61.4 kgCO2,eq /m2·y under coarse aggregation. For improvement strategies, demand reduction and increased PV producibility lead to net-balance reductions of 27.9 % and 24.8 %, while operational emission intensity decreases by about 17 % and 16 %, respectively. These findings confirm the sensitivity of ZeB assessments to methodological assumptions and support the proposed framework as a decision-support tool for robust, comparable and policy-relevant evaluations of building decarbonization pathways.
Wire drawing induces severe strain hardening in copper through dislocation multiplication, requiring subsequent thermal annealing to restore ductility. Although continuous direct current annealing is conventionally applied, high energy consumption and open-loop control instabilities limit processing efficiency. Electropulse-assisted treatment has emerged as an ultra-fast alternative, yet real-time dynamic current regulation and strict energy-equivalent metallurgical comparisons remain underexplored. This study presents an integrated framework combining closed-loop electropulse generator control, coupled multiphysics thermal-kinetics modeling, and an energy-equivalent physical benchmark (∼14,900 J) comparing direct current annealing and electropulse-assisted treatment across pulse durations of 35, 70, and 140μs on cold-drawn pure copper (99.90%) wires. A custom STM32 microcontroller platform regulates root-mean-square current via peak-current limiting and real-time frequency modulation, maintaining theoretical current accuracy within <0.16% and oscilloscope verification within <6.27%. The accompanying transient heat balance model accurately predicts surface temperature evolutions (R2=0.84–0.99), while a strain-compensated Johnson–Mehl–Avrami–Kolmogorov Arrhenius model captures recrystallization kinetics. Tensile testing reveals that under matched energy input, both direct current annealing and electropulse-assisted treatment successfully recover macroscopic ductility, yielding elongation of 34–37% and ultimate tensile strength (UTS) of 233–241MPa compared to cold-drawn baselines (1% elongation, ∼340MPa UTS). However, a key microstructural divergence emerges: direct current treatment reduces Vickers hardness by 66% and yields 92% recrystallization, whereas electropulse-assisted treatment induces a slight hardness change (＋0.75% to ＋7.8%) and a calculated recrystallization fraction of 38–47%, despite both triggering comparable grain growth (>10μm). Electrical resistivity exhibited no statistically significant variation across treated groups (p=0.267). These results prove that closed-loop electropulse-assisted treatment achieves macro-mechanical restoration comparable to direct current annealing under matched energy inputs while maintaining a distinct sub-microstructural state.
The rheological behavior of fresh cement paste is strongly influenced by both the phase composition of the portland cement and the molecular structure of polycarboxylate ether-based superplasticizers. Predicting flow time from these compositional variables remains challenging due to the nonlinear nature of the underlying interactions. This study employs tree-based ensemble machine learning models to predict the flow time of cement paste using a dataset comprising 616 observations and 14 input features that encompass cement phase components (C3S, C2S, C3A, C4AF) and key characteristics of the superplasticizer. Three algorithms, Gradient Boosting, Random Forest, and Extra Trees, were trained, optimized via grid search with k-fold cross-validation, and evaluated on an independent test set. Among these, the Extra Trees model delivered the highest predictive accuracy, achieving an R2 of 0.976 and a mean absolute error of 1.24 s. SHAP (SHapley Additive exPlanations) analysis was subsequently applied to interpret the model outputs. The results identify water-reducing admixture dosage as the single most influential predictor of flow time, followed at a considerable distance by superplasticizer molecular weight and cement specific surface area. Other material parameters, including C4AF content and pH, exerted only marginal effects within the investigated range. The findings demonstrate that ensemble learning methods, when coupled with interpretability techniques, can effectively capture the complex relationships governing cement paste rheology and provide a reliable, data-driven framework to support mixture proportioning and admixture compatibility assessment.
The mechanical performance of three-dimensional concrete printing (3DCP) mixtures is governed by a complex interaction between mixture composition, fiber characteristics, aggregate properties, and layer-wise deposition effects. Although compressive strength (CS) has been widely investigated, flexural strength (FS) is equally critical because printed elements are vulnerable to crack propagation, interlayer weakness, and direction-dependent behaviour. This study develops interpretable machine learning models for the prediction of FS and CS of 3D-printable concrete using literature-derived experimental datasets. Extra Trees (ET), XGBoost, and Random Forest (RF) models were trained using selected mixture, binder, admixture, fiber, and aggregate descriptors. Recursive Feature Elimination and model hyperparameter optimization were integrated within a random search procedure with 10-fold cross-validation and applied separately for each model and target variable. Model performance was assessed using R2, MAE, RMSE, MAPE, actual-predicted plots, Taylor diagrams, and residual analysis. SHAP analysis was then used to interpret the relative and directional effects of the selected variables. All three models provided accurate predictions, with XGBoost showing the best cross-validation performance for both FS and CS. XGBoost achieved independent-test R2 values of 0.9856 and 0.9781 for FS and CS, respectively, with the corresponding MAE values of 0.5211 MPa and 2.8523 MPa. SHAP results showed that FS predictions were influenced most strongly by water-reducing admixture content, water-to-binder ratio, fiber ratio, aggregate mass, and fiber density. In contrast, CS predictions were influenced primarily by cement mass, water-to-binder ratio, fiber ratio, and fiber density.
