Hongjo Kim
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Hongjo Kim has a total of 77 co-authors. Here are the top co-authors by number of shared publications:
Hyoungkwan Kim
Professor, School of Civil and Environmental Engineering, Yonsei Univ., 50 Yonsei-ro, Seodaemun-gu, Seoul 03722, South Korea
17 shared publications- Hyoungkwan Kim
Professor, School of Civil and Environmental Engineering, Yonsei Univ., Seoul 03722, Korea
8 shared publications Seongdeok Bang
Ph.D. Candidate, School of Civil and Environmental Engineering, Yonsei Univ., 50 Yonsei-ro, Seodaemun-gu, Seoul 03722, Korea
8 shared publicationsTaegeon Kim
Yonsei Univ
7 shared publicationsHoyoung Jeong
4 shared publicationsKichang Choi
4 shared publications
Latest Publications
- 1 November 2026
- Published by American Society of Civil Engineers (ASCE) in Journal of Management in Engineering
- Vol. 42 (6) , 04026049
Scenario-Based Framework for Economic Evaluation of Safety Technology Adoption in Construction
Despite growing interest in emerging safety technologies in the construction industry, their adoption in practice remains limited due to the lack of structured quantitative models that evaluate the trade-off between safety performance improvements and investment costs. Consequently, it remains unclear under which conditions safety technology adoption can achieve economic justification. To address this limitation, this study proposes the economic evaluation of safety technology adoption (ECOSTAT) framework, which integrates net value calculations, Monte Carlo simulation, and scenario analysis. To account for accident-prevention benefits in this evaluation, this study considers five cost components arising from construction accidents. These components include settlement costs, legal costs, accident investigation costs, interrupted construction costs, and fines. Among them, settlement costs, interrupted construction costs, and fines exhibit substantial variability depending on accident occurrence and characteristics and are therefore modeled using probability distributions. The applicability of the ECOSTAT framework is demonstrated through case studies of two construction project types: a building project and a railway infrastructure project. The results indicate that as total construction costs increase, the minimum accident-prevention efficiency required for economic feasibility decreases, reflecting economies of scale. For example, in the railway project, increasing total construction costs from $200 million to $350 million reduced the required accident-prevention efficiency from 5.36% to 3.95% at an investment ratio of 0.01%. Furthermore, the ECOSTAT framework is implemented as a web-based decision-support system that enables real-time evaluation of economic feasibility by allowing users to adjust project parameters.
- 1 November 2026
- Published by American Society of Civil Engineers (ASCE) in Journal of Management in Engineering
- Vol. 42 (6) , 04026049
Scenario-Based Framework for Economic Evaluation of Safety Technology Adoption in Construction
Despite growing interest in emerging safety technologies in the construction industry, their adoption in practice remains limited due to the lack of structured quantitative models that evaluate the trade-off between safety performance improvements and investment costs. Consequently, it remains unclear under which conditions safety technology adoption can achieve economic justification. To address this limitation, this study proposes the economic evaluation of safety technology adoption (ECOSTAT) framework, which integrates net value calculations, Monte Carlo simulation, and scenario analysis. To account for accident-prevention benefits in this evaluation, this study considers five cost components arising from construction accidents. These components include settlement costs, legal costs, accident investigation costs, interrupted construction costs, and fines. Among them, settlement costs, interrupted construction costs, and fines exhibit substantial variability depending on accident occurrence and characteristics and are therefore modeled using probability distributions. The applicability of the ECOSTAT framework is demonstrated through case studies of two construction project types: a building project and a railway infrastructure project. The results indicate that as total construction costs increase, the minimum accident-prevention efficiency required for economic feasibility decreases, reflecting economies of scale. For example, in the railway project, increasing total construction costs from $200 million to $350 million reduced the required accident-prevention efficiency from 5.36% to 3.95% at an investment ratio of 0.01%. Furthermore, the ECOSTAT framework is implemented as a web-based decision-support system that enables real-time evaluation of economic feasibility by allowing users to adjust project parameters.
- Open Access
- 2 October 2026
- Published by Elsevier in SSRN Electronic Journal
Crack Matting beyond Binary Segmentation: Fractional Occupancy Estimation for Crack Geometry Representation
Accurate crack geometry measurement supports condition assessment and maintenance planning for infrastructure and buildings. Automated image-based inspection commonly measures cracks using segmentation masks. However, binary masks lose geometric information when cracks occupy only part of a pixel. This study proposes Crack Matting to preserve this information by estimating the fraction of each pixel occupied by a crack. To estimate these fractions, the CrackMatte Pipeline adapts pretrained image matting models for continuous foreground estimation using segmentation-derived guidance. Occupancy targets constructed from existing annotations and measured widths support this adaptation and form the CrackAlpha dataset. Evaluation shows that fractional occupancy better preserves width accuracy and crack traces than binary encoding of the same geometry. The pipeline achieves a mean absolute width error of 0.608 pixels versus 0.858 for the best evaluated segmentation baseline on CrackAlpha-S and outperforms evaluated sub-pixel width methods after target-domain adaptation. These methods establish a baseline for crack-specific matting.
- 2 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCrack Matting beyond Binary Segmentation: Fractional Occupancy Estimation for Crack Geometry Representation
Accurate crack geometry measurement supports condition assessment and maintenance planning for infrastructure and buildings. Automated image-based inspection commonly measures cracks using segmentation masks. However, binary masks lose geometric information when cracks occupy only part of a pixel. This study proposes Crack Matting to preserve this information by estimating the fraction of each pixel occupied by a crack. To estimate these fractions, the CrackMatte Pipeline adapts pretrained image matting models for continuous foreground estimation using segmentation-derived guidance. Occupancy targets constructed from existing annotations and measured widths support this adaptation and form the CrackAlpha dataset. Evaluation shows that fractional occupancy better preserves width accuracy and crack traces than binary encoding of the same geometry. The pipeline achieves a mean absolute width error of 0.608 pixels versus 0.858 for the best evaluated segmentation baseline on CrackAlpha-S and outperforms evaluated sub-pixel width methods after target-domain adaptation. These methods establish a baseline for crack-specific matting.
- Open Access
- 27 September 2026
- Published by Elsevier in SSRN Electronic Journal
Surrogate-guided diffusion for performance-aware generation of residential floorplans and window-to-wall ratios
In many generate-then-evaluate workflows, performance information arrives only after a layout exists: generators produce the artifact while feedback remains outside, and climate-based simulation is too slow to consult at the pace alternatives are produced. Daylight makes this problem concrete because room exposure to the exterior boundary and available window area are largely settled by early layout decisions. This paper presents surrogate-guided diffusion, a vector floorplan generator that encodes structured floorplan requirements as conditions and produces exterior-wall window-to-wall ratios together with room geometry. During sampling, a learned daylight surrogate steers both geometry and window ratios, while geometry-stabilizing terms limit distortion. Because guidance is applied at inference, the same generator serves three distinct daylight objectives without retraining. Annual Radiance simulation under a Seoul climate independently verifies all daylight outcomes. Among each configuration’s own overlap-free outputs generated without a prescribed boundary, guided sampling averaged 84.4% versus 71.5% for unguided sampling in spatial Daylight Autonomy, 11.7% versus 14.9% in the over-illumination band of Useful Daylight Illuminance, and 35.5% versus 17.8% for the worst room. Under both free and boundary-conditioned generation, every objective moved its target in the intended direction relative to matched unguided and best-of-five selection baselines. A separate surrogate audit distinguishes level accuracy from directional behavior: under finite perturbations of completed generated plans’ window ratios, surrogate and simulation changes agreed in direction more often than not. The results show how simulation-derived performance information can act on structured residential layouts while they are being formed.
- 27 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessSurrogate-guided diffusion for performance-aware generation of residential floorplans and window-to-wall ratios
In many generate-then-evaluate workflows, performance information arrives only after a layout exists: generators produce the artifact while feedback remains outside, and climate-based simulation is too slow to consult at the pace alternatives are produced. Daylight makes this problem concrete because room exposure to the exterior boundary and available window area are largely settled by early layout decisions. This paper presents surrogate-guided diffusion, a vector floorplan generator that encodes structured floorplan requirements as conditions and produces exterior-wall window-to-wall ratios together with room geometry. During sampling, a learned daylight surrogate steers both geometry and window ratios, while geometry-stabilizing terms limit distortion. Because guidance is applied at inference, the same generator serves three distinct daylight objectives without retraining. Annual Radiance simulation under a Seoul climate independently verifies all daylight outcomes. Among each configuration’s own overlap-free outputs generated without a prescribed boundary, guided sampling averaged 84.4% versus 71.5% for unguided sampling in spatial Daylight Autonomy, 11.7% versus 14.9% in the over-illumination band of Useful Daylight Illuminance, and 35.5% versus 17.8% for the worst room. Under both free and boundary-conditioned generation, every objective moved its target in the intended direction relative to matched unguided and best-of-five selection baselines. A separate surrogate audit distinguishes level accuracy from directional behavior: under finite perturbations of completed generated plans’ window ratios, surrogate and simulation changes agreed in direction more often than not. The results show how simulation-derived performance information can act on structured residential layouts while they are being formed.
- 17 April 2026
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 74
RAMASC: A retrieval-augmented multi-agent framework for automated structural calculation
- 17 April 2026
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 74
RAMASC: A retrieval-augmented multi-agent framework for automated structural calculation
- Open Access
A Comparison of Human Capabilities and Large Language Models for Knowledge Representation with Ontologies of Non-Destructive Testing in Bridge Engineering
Bridge structures are considered complex and significant. Accordingly, the knowledge of the engineering domain of bridge construction and related specialist areas is multidimensional and highly specific. Sometimes this knowledge is explicitly documented in standards, technical regulations, or information sheets. At other times, it resides implicitly in the expertise of the specialists involved. Ontologies are used to structure and formalize such domain knowledge, but creating them is resource-intensive and requires specialized expertise. Large language models (LLMs) offer one way to automate ontology creation through their natural language processing capabilities. This article examines LLMs’ ability to generate ontologies in the specialized field of structural non-destructive testing (NDT) in bridge construction. Four different LLM-based approaches are employed. The results are compared with a previously created human-generated ontology and subsequently evaluated by external experts. Experts rate the human-developed SODIA ontology highest, with an average score of 3.44 out of 5 points. Only the ChatGPT 4.0-created ontology performed similarly well, with a score of 3.3 out of 5.00. All other LLM-based ontologies with ratings below 3.0 are of minor quality. These results underscore the potential and constraints of using LLMs to structure and formalize engineering domain knowledge into ontologies.
Open AccessA Comparison of Human Capabilities and Large Language Models for Knowledge Representation with Ontologies of Non-Destructive Testing in Bridge Engineering
Bridge structures are considered complex and significant. Accordingly, the knowledge of the engineering domain of bridge construction and related specialist areas is multidimensional and highly specific. Sometimes this knowledge is explicitly documented in standards, technical regulations, or information sheets. At other times, it resides implicitly in the expertise of the specialists involved. Ontologies are used to structure and formalize such domain knowledge, but creating them is resource-intensive and requires specialized expertise. Large language models (LLMs) offer one way to automate ontology creation through their natural language processing capabilities. This article examines LLMs’ ability to generate ontologies in the specialized field of structural non-destructive testing (NDT) in bridge construction. Four different LLM-based approaches are employed. The results are compared with a previously created human-generated ontology and subsequently evaluated by external experts. Experts rate the human-developed SODIA ontology highest, with an average score of 3.44 out of 5 points. Only the ChatGPT 4.0-created ontology performed similarly well, with a score of 3.3 out of 5.00. All other LLM-based ontologies with ratings below 3.0 are of minor quality. These results underscore the potential and constraints of using LLMs to structure and formalize engineering domain knowledge into ontologies.
- 28 March 2026
- Published by Elsevier in Automation in Construction
- Vol. 187
Tiny object detection using Distance-guided, Signed, and Densified Learning (DSDL) for construction site safety monitoring
- 28 March 2026
- Published by Elsevier in Automation in Construction
- Vol. 187
Tiny object detection using Distance-guided, Signed, and Densified Learning (DSDL) for construction site safety monitoring
- Open Access
- 27 March 2026
- Published by Elsevier in SSRN Electronic Journal
MANTA: Multivariate Attention Network for Short-Term Water-Level Forecasting
Accurate short-term water-level forecasting is a critical requirement for flood early-warning systems, particularly under extreme hydrological conditions characterized by rapid and non-stationary water-level fluctuations. In practical deployment scenarios, forecasting models must operate with limited recent observations and generalize to previously unseen monitoring locations, conditions that challenge many existing data-driven approaches. This study formulates short-term water-level forecasting as a real-time forecasting problem constrained by short observation windows, location-independent generalization, and robustness under extreme fluctuation conditions. A unified data representation is adopted to enable consistent model training across heterogeneous monitoring stations, and a forecasting framework is developed using an architecture that prioritizes stability under distribution shifts rather than average-case optimization. Extreme water-level fluctuation events are used to evaluate robustness under the most critical operational conditions. Experiments conducted on a nationwide monitoring network in South Korea (202 stations) demonstrate consistent performance improvements over baseline models, particularly during extreme events. While overall accuracy gains are moderate (MSE reduced from 0.0078 to 0.0051 compared to TSMixer), the proposed approach exhibits substantially reduced prediction errors at critical short lead times (t+1–t+6). Under extreme fluctuation conditions (≥0.6 m/hour), MANTA (Multivariate Attention Network for short-Term wAter-level Forecasting) reduces MSE from 1.273 to 0.932 with a pronounced improvement at t+1 where MSE decreases by approximately 42%, indicating improved reliability for flood early-warning decision-making. The framework has been deployed as a real-time forecasting system with continuous performance monitoring, supporting ongoing validation and data-centric refinement under real-world non-stationary conditions.
- 27 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessMANTA: Multivariate Attention Network for Short-Term Water-Level Forecasting
Accurate short-term water-level forecasting is a critical requirement for flood early-warning systems, particularly under extreme hydrological conditions characterized by rapid and non-stationary water-level fluctuations. In practical deployment scenarios, forecasting models must operate with limited recent observations and generalize to previously unseen monitoring locations, conditions that challenge many existing data-driven approaches. This study formulates short-term water-level forecasting as a real-time forecasting problem constrained by short observation windows, location-independent generalization, and robustness under extreme fluctuation conditions. A unified data representation is adopted to enable consistent model training across heterogeneous monitoring stations, and a forecasting framework is developed using an architecture that prioritizes stability under distribution shifts rather than average-case optimization. Extreme water-level fluctuation events are used to evaluate robustness under the most critical operational conditions. Experiments conducted on a nationwide monitoring network in South Korea (202 stations) demonstrate consistent performance improvements over baseline models, particularly during extreme events. While overall accuracy gains are moderate (MSE reduced from 0.0078 to 0.0051 compared to TSMixer), the proposed approach exhibits substantially reduced prediction errors at critical short lead times (t+1–t+6). Under extreme fluctuation conditions (≥0.6 m/hour), MANTA (Multivariate Attention Network for short-Term wAter-level Forecasting) reduces MSE from 1.273 to 0.932 with a pronounced improvement at t+1 where MSE decreases by approximately 42%, indicating improved reliability for flood early-warning decision-making. The framework has been deployed as a real-time forecasting system with continuous performance monitoring, supporting ongoing validation and data-centric refinement under real-world non-stationary conditions.
- Open Access
An Implementation of the Crack Topology Score with Extensions
The Crack Topology Score (CTS) is a recently proposed metric that focuses on evaluating the topological correctness of crack segmentation outputs. While pixel-wise metrics such as IoU or F1-score fail to capture structural validity, CTS offers a skeleton-based matching framework to measure the preservation of connectivity. This paper presents a faithful implementation of the CTS metric, along with optional preprocessing extensions designed to handle common prediction artifacts (e.g., small holes and edge noise) found in deep learning outputs. All extensions are disabled by default to ensure strict comparability with the original definition. The implementation supports PyTorch-based workflows and includes visualization tools for transparency. Code and archival resources will be made available at https://github.com/SH-Joo/crack-topology-score.
Open AccessAn Implementation of the Crack Topology Score with Extensions
The Crack Topology Score (CTS) is a recently proposed metric that focuses on evaluating the topological correctness of crack segmentation outputs. While pixel-wise metrics such as IoU or F1-score fail to capture structural validity, CTS offers a skeleton-based matching framework to measure the preservation of connectivity. This paper presents a faithful implementation of the CTS metric, along with optional preprocessing extensions designed to handle common prediction artifacts (e.g., small holes and edge noise) found in deep learning outputs. All extensions are disabled by default to ensure strict comparability with the original definition. The implementation supports PyTorch-based workflows and includes visualization tools for transparency. Code and archival resources will be made available at https://github.com/SH-Joo/crack-topology-score.
- Open Access
- 27 July 2025
- Published by International Association for Automation and Robotics in Construction (IAARC) in Proceedings of the 37th International Symposium on Automation and Robotics in Construction (isarc)
- Vol. 2025 Proce, 485-491
Paraphrasing-Based Data Augmentation for Responsible Personnel Classification in Crane Accidents
Crane accidents often result in severe injuries and fatalities, necessitating accurate analysis and responsibility prediction to prevent future incidents. However, the scarcity of crane accident-specific datasets and class imbalance pose challenges to developing robust predictive models. This study addresses these challenges by leveraging paraphrasebased data augmentation methodology to expand a dataset of 480 natural language accident descriptions into a balanced dataset of 900 samples. Using Google's Pegasus paraphrasing model, the augmented data improved model performance, as demonstrated by the F1 score increase from 0.29 to 0.5583. Three pre-trained transformer BERT-based models were fine-tuned on the augmented dataset to evaluate their effectiveness in predicting responsible personnel. The results highlight the effectiveness of the paraphrasing technique and transformer-based models in addressing class imbalance and improving classification accuracy. This research demonstrates the potential of natural language text data in enhancing safety analysis and proposes future directions, such as exploring alternative data augmentation methods, leveraging large language models, and experimenting with domain-specific finetuning.
- 27 July 2025
- Published by International Association for Automation and Robotics in Construction (IAARC) in Proceedings of the 37th International Symposium on Automation and Robotics in Construction (isarc)
- Vol. 2025 Proce, 485-491
Open AccessParaphrasing-Based Data Augmentation for Responsible Personnel Classification in Crane Accidents
Crane accidents often result in severe injuries and fatalities, necessitating accurate analysis and responsibility prediction to prevent future incidents. However, the scarcity of crane accident-specific datasets and class imbalance pose challenges to developing robust predictive models. This study addresses these challenges by leveraging paraphrasebased data augmentation methodology to expand a dataset of 480 natural language accident descriptions into a balanced dataset of 900 samples. Using Google's Pegasus paraphrasing model, the augmented data improved model performance, as demonstrated by the F1 score increase from 0.29 to 0.5583. Three pre-trained transformer BERT-based models were fine-tuned on the augmented dataset to evaluate their effectiveness in predicting responsible personnel. The results highlight the effectiveness of the paraphrasing technique and transformer-based models in addressing class imbalance and improving classification accuracy. This research demonstrates the potential of natural language text data in enhancing safety analysis and proposes future directions, such as exploring alternative data augmentation methods, leveraging large language models, and experimenting with domain-specific finetuning.