Research Article

L-SCOPE: An LLM-Assisted Interactive Platform
for Efficient Building Design Optimization
Qianyun Zhou1, Fan Xue2*
1 PhD student, Department of Real Estate and Construction, The University of Hong Kong
2 Associate Professor, Department of Real Estate and Construction, The University of Hong
Kong
*Correspondence: xuef@hku.hk

This is the peer-reviewed post-print version of the conference paper:
Zhou, Q., & Xue, F. (2025). L-SCOPE: An LLM-Assisted Interactive Platform for EXicient
Building Design Optimization. In Proceedings of Smart and Sustainable Built
Environment Conference Series, 96-105. Publisher: Delft, Netherlands: International
Council for Research and Innovation in Building and Construction (CIB).
https://isasbec.abc2.net/index.php/sasbe/article/view/2554
This ﬁle is shared for personal and academic use only, under the license CC BY-NC 4.0
(Non-Commercial and with an Attributed citation when you use). The ﬁnal published version
of this paper can be found at: https://isasbec.abc2.net/index.php/sasbe/article/view/2554.
Abstract
Copyright: Copyright: © 2025 by the
authors.
SASBE is an open-access
proceedings distributed under the
terms of the Creative Commons
Attribution 4.0 International License
(CC BY 4.0).
View this license’s legal deed at
https://creativecommons.org/license
s/by/4.0/

The building and construction sector accounts for a signiﬁcant share of global energy use
and greenhouse gas emissions, highlighting the importance of focusing on sustainable
design for climate change mitigation. Early-stage design decisions are crucial in shaping a
building’s life cycle energy performance. However, sustainable building design exploration
often relies on simulation-based optimization (SBO), of which complex objectives and
numerous design parameters result in high computational complexity and limited user
understanding of how individual parameters inﬂuence the outcome. This study introduces
L-SCOPE, a large language model (LLM)-assisted interactive platform designed to enhance
the interpretability and eXiciency of complex SBO. The platform provides an end-to-end
workﬂow that includes environment setup, information-theoretic analysis, and automated
interval reﬁnement of design parameters. Stakeholders directly participate in the analytical
process through a prompt-driven, multi-turn dialogue interface, where L-SCOPE facilitates
seamless human-AI interaction by implementing natural language prompts, generative code
reviews, and interpretable visual feedback. A case study demonstrates that the platform
achieved robust and reliable results, with a tenfold reduction in convergence analysis time
and a 0.49% improvement in energy performance compared to conventional SBO methods
under the same computational budget. The contributions of this study are twofold: (i) a
novel approach that combines LLM capabilities with information-theoretic analysis to
advance interpretable and eXicient building design optimization; and (ii) an interactive
language-based interface that facilitates active participation and decision-making by nonexpert stakeholders.

Keywords: Large Language Model (LLM); Human-AI Interaction; Sustainable Design
Optimization; Parameter Converging Tool; Design Decision Support

Highlights

• A novel LLM-assisted platform integrating information-theoretic analysis with SBO.
• Enhances interpretability and stakeholder participation via interactive dialogue interface.
• Achieves 10× faster convergence analysis time with 0.49% energy performance
improvement under same budget.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

1 Introduction
The building and construction sector signiﬁcantly contributes to global climate change, accounting for
approximately 34% of global energy consumption and 37% of greenhouse gas emissions (UNEP 2024).
The increasing adoption of sustainable design practices in the building sector has demonstrated that
early-stage performance-driven design plays a critical role in shaping long-term energy eHiciency and
life-cycle environmental impact, thereby helping to mitigate environmental burdens (Wang et al. 2024;
Zhan et al. 2024).
Simulation-based optimization (SBO) has become a widely used approach for performancedriven building design. By integrating parametric modeling and performance simulations, SBO enables
iterative exploration of design alternatives and evaluation of parameter impacts on performance
objectives (Luca et al. 2024). However, SBO in early design often involves numerous parameters and
objective functions with black-box characteristics, leading to a complex design search space that
increases computational demand and limits the interpretability of design decisions (Zhou & Xue 2025).
Conventional optimizers struggle to approximate near-global solutions within limited iterations,
underscoring the need for more eHicient strategies to guide and shrink the design search space under
constrained computational budgets.
Mainstream strategies employed surrogate models to approximate the design search space and
sensitivity analysis to identify important design parameters (Danhaive & Mueller 2021; Hinkle et al. 2024;
Tian et al. 2024). However, both techniques rely heavily on extensive simulation datasets and often
require re-execution for diHerent optimization tasks, resulting in signiﬁcant time and labour costs. To
address these limitations, Zhou and Xue (2025) introduced the Automatic Information Gain-guided
Convergence (AIGGC) method, which instead employs small-scale historical optimization data to
perform information gain-guided analysis of sub-intervals within each design parameter. The AIGGC
method reduces the complexity of the design search space and accelerates convergence while
enhancing interpretability. Nevertheless, the absence of an integrated, visual, and interactive platform
has limited its broader application in practice.
Large language models (LLMs) have shown considerable potential in natural language
understanding and code generation (Fang et al. 2024). By lowering the barrier between domain expertise
and computational modelling, LLMs enable non-expert users to specify complex analytical tasks
without extensive programming skills. In architecture, engineering, and construction (AEC), recent
studies have applied LLMs to support more interactive and accessible human-AI collaboration. These
applications are typically grouped into three categories: information retrieval (Guo et al. 2025; Li & Wang
2025), where LLMs translate natural-language queries into building or BIM semantic data; task
automation (Gao et al. 2025; Zhang et al. 2025), where LLM-based agents convert user requirements
into executable modelling workﬂows; and interpretability enhancement (Zhang & Chen 2024), where
LLMs are combined with interpretable machine learning to clarify rule-based rationale. However, the
applications of LLMs in complex SBO tasks remain limited.
This study presents L-SCOPE, an LLM-assisted interactive platform for design Space
Convergence and Optimized Parameter Exploration. Operating within the Grasshopper ecosystem, LSCOPE seamlessly integrates with existing SBO tools. Through a prompt-driven, multi-turn dialogue
interface, users can issue natural language queries, customize analysis indices, and review editable
code before execution. This human-in-the-loop process mitigates risks of hallucination, improves
interpretability through visualization, and preserves user control. The platform integrates optimization
Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

outputs, design parameter analysis, and automated domain convergence into an interoperable
workﬂow, extending conventional SBO with interpretable, stakeholder-oriented insights and AI-assisted
decision support.
The contributions of this study are twofold: (i) an integrated approach that combines LLM
capabilities with information-theoretic analysis to enhance the interpretability and eHiciency of building
design optimization; and (ii) an interactive language-based interface that facilitates active participation
and decision-making by non-expert stakeholders.

2 Related works
In early-stage sustainable building design, designers generally adopt Rhino-Grasshopper workﬂows
with SBO tools to explore and evaluate design alternatives (Luca et al. 2024). Common plugins such as
Ladybug for daylighting, Honeybee for energy simulation, and various optimization solvers enable
automated design iterations. Popular optimizers, including Genetic Algorithm (GA), CMA-ES, and
RBFOpt-based methods, run predeﬁned iterations and output the best-performing solution (Brown et
al. 2020). Since the objective functions are black boxes computed by external simulation engines using
climate and physical models, direct mathematical formulations are infeasible. Combined with
numerous design parameters in early stages, the black-box nature of SBO results in complex search
spaces that limit interpretability and computational eHiciency (Zhou & Xue 2025).
Recent research has explored deep generative models and surrogate-based methods to
approximate complex search spaces and support exploration of speciﬁc design regions (Brown et al.
2020; Danhaive & Mueller 2021). Others have applied metamodels for real-time interactive optimization
and sensitivity analysis to identify inﬂuential design parameters (Hinkle et al. 2024). However, most
sensitivity-based methods need large-scale sampling, which is impractical for time-intensive
simulations. Similarly, machine learning-based feature selection requires retraining when design
parameters or ranges change, increasing computational costs. Moreover, the problem-speciﬁc nature
of surrogate models limits the generalizability across diHerent design scenarios, while their predictive
accuracy tends to degrade in high-dimensional or sparsely sampled design spaces (Zhou & Xue 2025).
The AIGGC method addresses these limitations by reallocating a small portion of the
optimization budget to perform information gain analysis on design parameters’ sub-intervals (Zhou &
Xue 2025). The method reduces search space complexity, accelerates optimization convergence, and
improves SBO interpretability, with demonstrated robustness across diHerent scales and optimizers.
However, the absence of an integrated, visual, and interactive platform, particularly within the RhinoGrasshopper ecosystem, limits accessibility for practitioners, making it diHicult for users to understand
and apply the method eHectively.

3 L-SCOPE: LLM-assisted design Space Convergence for Optimized
Parameter Exploration
L-SCOPE is an interactive platform designed to enhance the eHiciency of complex SBO in early-stage
design by integrating an LLM-assisted information-theoretic analysis workﬂow.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

3.1 L-SCOPE workﬂow and components
As illustrated in Figure 1, the platform operates within the Rhino-Grasshopper environment. The
workﬂow begins with an initial SBO run, where stakeholders deﬁne design parameters, performance
objectives, and an initial iteration budget (N1) using relevant Grasshopper plugins for modelling,
simulation, and optimization. The resulting dataset (e.g., .csv) is transferred to L-SCOPE, which
interprets stakeholder prompts and generates executable Python code for information-theoretic
analysis. Based on information gain rankings, L-SCOPE guides stakeholders in reﬁning design
parameters’ domain intervals. The reﬁned domains are then applied in a second SBO run under the
remaining budget (N2), thereby converging the search toward optimal building design. Each
LLM-generated script was cross-validated with a manually implemented reference during development
to ensure correctness and reproducibility. In practical use, the platform returns the generated code to
the user interface for human inspection before execution, reducing hallucination risks and preserving
user control.

Figure 1. The proposed workﬂow for L-SCOPE.

As shown in Figure 2, L-SCOPE integrates four Grasshopper components into a three-step
pipeline covering environment setup, interactive information-theoretic analysis, and automated
reﬁnement for design space convergence. Step 1 establishes the local execution environment and
backend server, including components for creating a virtual environment with Python dependencies and
launching the LLM server. Step 2 is the core interactive module, where a control panel enables users to
issue natural language prompts, review LLM-generated Python code and execution feedback, and
visualize key results such as elite probability distributions of sub-intervals and information gain rankings.
Based on these outputs, Step 3 automatically reﬁnes the sub-intervals of selected design parameters,
adapts input ports to match the detected parameters, and outputs updated domain intervals with userdeﬁned precision for subsequent optimization.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

Figure 2. The four components of L-SCOPE.

3.2 L-SCOPE interactive panel
Figure 3 displays the layout of the interactive panel within L-SCOPE. The platform supports multi-modal
outputs, including Python code, console messages, chat history, structured data, and visualizations.
After loading the dataset, the panel displays a parallel coordinates plot to help users inspect the
distribution of design parameters. On the left, analysis factors can be speciﬁed through sliders or
natural language prompts. Intermediate results are shown through console outputs and visual feedback,
ensuring transparency and traceability. The ﬁnal output is a structured summary of key information,
such as the sub-intervals targeted for reﬁnement and design parameters classiﬁed as constant.

Figure 3. Details of L-SCOPE interactive panel.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

4 Experimental results
4.1 Case study and experimental settings
To evaluate the eHectiveness of the proposed platform, a case study was conducted based on the
published work of Zhou and Xue (2023). The study focused on a Modular Integrated Construction (MiC)
project for a university student hostel located in Hong Kong. In this case, the optimization objective was
to minimize the Energy Use Intensity (EUI), deﬁned as the total annual energy consumption divided by
the gross ﬂoor area (kWh/m2·yr). The EUI value was calculated via the Honeybee-Radiance (Ver.1.8)
and Honeybee-OpenStudio (Ver.1.8) components, with a sensor grid size of 0.5 meters. As shown in
Figure 4, the project features a symmetrical layout in which modules Bw, Cw, and Ct remain ﬁxed
throughout the study, and modules MA, MB, and MC were reorganized during the optimization process. In
addition to module layout, window size and material properties were also considered due to their
signiﬁcant impact on energy performance (Qin & Pan 2020). The design parameters and their
corresponding domain intervals are summarized in Table I.

Figure 4. MiC case layout and schematic of module types.
Table I. Description of design parameters for MiC parametric design.
Design parameters

Parameter type

Unit

Parameter domain intervals

Precision

MiC Module Type (M1, M2, M3, M4, M5, M6, M7,
M8, M9, M10, M11, M12, M13, M14)
Window-to-Wall Ratio (WWRN, WWRS)

Categorial

-

[MA, MB, MC]

-

Continuous

-

0.10 ≤ WWRN, WWRS ≤ 0.60

0.01

Window Height (WHN, WHS)

Discrete

m

1.0 ≤ WHN, WHS ≤ 2.5

0.1

Window U-value (UW)

Continuous

W/(m2·K)

0.50 ≤ UW ≤ 5.78

0.01

Window Solar Heat Gain CoeRicient (g)

Continuous

-

0.100 ≤ UW ≤ 0.775

0.001

Window Visible Transmittance (τv)

Continuous

-

0.300 ≤ UW ≤ 0.881

0.001

The experiments were conducted on a desktop computer with an Intel (R) Core i7-10700 CPU @
2.90 GHz processor and 32 GB memory. In this study, DeepSeek serves as the backbone LLM for code
generation and analysis. The experimental conﬁguration in this study follows the settings recommended
in our previous work (Zhou & Xue 2025). The total iteration budget was set to 300, divided into an initial
stage (N1=100) and a second stage (N2=200), optimized using a genetic algorithm (GA) through the
Galapagos component in Grasshopper. The experiments were repeated 10 times (as each run set
required 36 h) to mitigate the stochastic eHects of GA and evaluate the robustness of L-SCOPE. An elite
ratio re=20% was used to classify initial-stage results into ‘elite’ and ‘non-elite’ groups. Each design
parameter’s domain interval was divided into 10 sub-interval bins for calculating elite probability and
Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

entropy. Based on the calculated information gain, design parameters were then ranked and categorized
into three groups using two thresholds: rconv and rconst, both set to one-third. Parameters with high
information gain were reﬁned by excluding sub-intervals with zero elite probability; those with low
information gain were ﬁxed to constant values; and the remaining parameters were left unchanged. For
detailed deﬁnitions and algorithmic procedures, please refer to (Zhou & Xue 2025).

4.2 Experimental results and analysis
4.2.1 Optimized results with L-SCOPE
The experimental results, summarized in Table II, compare the performance of the Genetic Algorithm
(GA) with and without the L-SCOPE platform over ten independent runs. On average, L-SCOPE achieved
a lower EUI of 113.466 kWh/m2·yr, compared with 114.025 kWh/m2·yr obtained from conventional SBO,
representing a 0.49% performance gain under the same computational budget. For the best-optimized
solutions, L-SCOPE further reduced the EUI from the baseline value of 116.747 to 112.684 kWh/m2·yr,
corresponding to a 3.48% improvement. A two-tailed independent t-test yielded a statistically
signiﬁcant result (p = 0.019 < 0.05), supporting the robustness and reliability of the observed
improvements.
Table II. The magnitude and signiﬁcance of L-SCOPE’s performance in ten runs.
Objective function f
(Baseline value v0)

SBO
algorithm

Average optimization results

With
% Gain.*
Sig.#
Conventional SBO
L-SCOPE
EUI (kWh/m2·yr)
114.025
113.466
0.49%
0.019
GA
Δ = f − v0
− 2.722
− 3.281
% Imp.†
2.33%
2.81%
†: Improvement relative to baseline EUI, computed as| Δ |/v0 × 100%, where v0 =116.747 kWh/m2·yr.
*: Relative performance gain of L-SCOPE over Conventional SBO in average EUI results.
#: Two-tailed p-value of independent t-test, bold when p-value < 0.05.

Best solution
Conventional
SBO
113.388
− 3.359
2.88%

With
L-SCOPE
112.684
− 4.063
3.48%

Table III shows a visual comparison between the baseline and L-SCOPE-optimized layouts. The
optimized design obtained a 4.063 kWh/m2·yr reduction in EUI, with a clearer preference for using
Module B and C over A. The optimized MiC layout also exhibits a more compact and ﬂattened horizontal
proﬁle.
Table III. Comparison between the best of optimized solutions with L-SCOPE and the baseline solution.

EUI (kWh/m2·yr)

Baseline model layout

Best optimized solution based on L-SCOPE

116.747

112.684

4.2.2 Robustness evaluation and time e@iciency of L-SCOPE
To evaluate the robustness of the LLM-generated code, the platform was tested using a ﬁxed sequence
of seven prompts across ten independent runs. In all cases, the LLM reliably produced syntactically
correct and executable Python code. Only minor formatting artifacts, such as Markdown-style
annotations (e.g., triple backticks), occasionally appeared but were easily removed without aHecting
execution or reproducibility. No runtime errors or logic inconsistencies were observed, demonstrating
the platform’s robustness in maintaining context and variable consistency across multi-step prompt
interactions. Moreover, validation against a manually implemented reference conﬁrmed that the LLMgenerated code replicated the core information-theoretic analytical logic, achieving consistent
Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

information gain rankings, accurate identiﬁcation of converging design parameters, and reﬁnement of
their corresponding sub-interval domains.
As illustrated in Figure 5, L-SCOPE consistently translated sequential prompts into valid code
and coherent visualizations. All ten runs exhibited uniform colour schemes and layout structures, with
only minor variations in bar chart rendering. These results conﬁrm the platform’s reliability in producing
reproducible, interpretable outputs through an end-to-end analysis workﬂow.

Figure 5. End-to-end outputs from sequential prompt execution in L-SCOPE.

The platform also improves analysis eHiciency. As shown in Table IV, manual informationtheoretic analysis typically takes 40 to 60 minutes, covering tasks such as code development,
debugging, and identifying converging design parameter intervals. In contrast, L-SCOPE completes the
whole process in 4 to 6 minutes, reducing analysis time by a factor of ten and accelerating early-stage
design exploration.
Table IV. Estimated time range of manual and LLM-assisted analysis.
Estimated time (min)

Manual

L-SCOPE

40-60

4-6

5 Discussion
This study presents the ﬁrst implementation of the L-SCOPE platform for early-stage SBO in
architectural design. The platform integrates a large language model (Deepseek) with an information
gain-guided convergence framework to address the limitations of low interpretability and search
ineHiciency in existing tools. L-SCOPE enables non-expert users to participate directly in the analysis
through a prompt-driven, human-in-the-loop workﬂow that interactively guides the reﬁnement of design
parameters’ domain intervals. Compared to conventional SBO methods, L-SCOPE achieved a tenfold
reduction in convergence analysis time and a 0.49% enhancement in energy performance.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2

Experimental results conﬁrmed the reliability of the LLM-generated code and the robustness of
analytical outputs across repeated runs.
The limitations of this study are twofold: ﬁrst, it lacks comparative evaluation against alternative
design space reduction techniques; second, it has not been extended to address multi-objective
optimization tasks, such as daylighting, thermal comfort, or construction cost, which remain a direction
for future development.

6 Conclusions
This study proposes L-SCOPE, an LLM-assisted interactive platform designed to support interpretable
and eHicient early-stage building design optimization. The platform consists of four Grasshopper
components, enabling an end-to-end workﬂow that covers environment setup, server initialization,
information-theoretic analysis, and automated interval reﬁnement. By combining large language
models with information gain-guided convergence, L-SCOPE facilitates seamless human-AI interaction
and empowers non-expert users to engage in design space convergence through natural language
prompts. Experimental results demonstrate the platform’s robustness, reproducibility, and signiﬁcant
improvements in analysis eHiciency. L-SCOPE contributes to both the enhanced interpretability and
computational eHiciency of the optimization process, and the promotion of broader stakeholder
participation through an interactive language-based interface.
Future developments will focus on evaluating L-SCOPE’s broader applicability to diverse
performance objectives, benchmarking alternative language models, and validating user accessibility
and decision support eHectiveness through practitioner studies.
Acknowledgements

The study was supported by Hong Kong Research Grants Council (RGC) (Nos. C7080-22GF, T22-504/21-R).

Data Availability Statement

A packaged version of the L-SCOPE plugin and example Grasshopper ﬁles are publicly available at:
https://github.com/Joannazhou-qianyun/L-SCOPE. The repository also contains documentation and instructions for
installation and use. All data supporting the ﬁndings of this study can be reproduced using the provided example ﬁles.

Conflicts of Interest

The authors declare that they have no known competing ﬁnancial interests or personal relationships that could have
appeared to inﬂuence the work reported in this paper.

Authorship Contribution Statement

Qianyun Zhou: Writing – original draft, Visualization, Validation, Software, Resources, Methodology, Data curation,
Conceptualization. Fan Xue: Writing – review & editing, Supervision, Project administration, Methodology, Funding acquisition,
Conceptualization.

References
Brown, N. C., Jusiega, V. & Mueller, C. T. (2020). Implementing data-driven parametric building design with a
ﬂexible toolbox. Automation in Construction, 118, 103252. doi:10.1016/j.autcon.2020.103252
Danhaive, R. & Mueller, C. T. (2021). Design subspace learning: Structural design space exploration using
performance-conditioned generative modeling. Automation in Construction, 127, 103664.
doi:10.1016/j.autcon.2021.103664
Fang, C., Miao, N., Srivastav, S., Liu, J., Zhang, R., Fang, R., Tsang, R., Nazari, N., Wang, H. & Homayoun, H.
(2024). Large Language Models for Code Analysis: Do LLMs Really Do Their Job? 33rd USENIX Security
Symposium (USENIX Security 24) (pp. 829-846). Philadelphia: USENIX Association.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx

Qianyun Zhou1, Fan Xue2
Gao, M. Y., Li, C., Petzold, F., Tiong, R. L. & Yang, Y. (2025). Lifecycle framework for AI-driven parametric
generative design in industrialized construction. Automation in Construction, 174, 106146.
doi:10.1016/j.autcon.2025.106146
Guo, P., Xue, H., Ma, J. & Cheng, J. C. (2025). Advancing BIM information retrieval with an LLM-based querydomain-speciﬁc language and library code function alignment system. Automation in Construction,
178, 106374. doi:10.1016/j.autcon.2025.106374
Hinkle, L. E., Pavlak, G., Curtis, L. & Brown, N. C. (2024). Implementing dynamic subset sensitivity analysis for
early design datasets. Automation in Construction, 158, 105198. doi:10.1016/j.autcon.2023.105198
Li, M. & Wang, Z. (2025). BuildingGPT: Query Building Semantic Data Using Large Language Models and VectorGraph Retrieval-Augmented Generation. Building and Environment, 113855.
doi:10.1016/j.buildenv.2025.113855
Luca, F. D., Natanian, J. & Wortmann, T. (2024). Ten questions concerning environmental architectural design
exploration. Building and Environment, 261, 111697. doi:10.1016/j.buildenv.2024.111697
Qin, H. & Pan, W. (2020). Energy use of subtropical high-rise public residential buildings and impacts of energy
saving measures. Journal of Cleaner Production, 254, 120041. doi:10.1016/j.jclepro.2020.120041
Tian, L., Hao, T., He, X., Chan, I., Niu, J., Chan, P., Ng, W. & Huang, J. (2024). Examining the non-linear relationship
between urban form and air temperature at street level: A case of Hong Kong. Building and Environment,
264, 111884. doi:10.1016/j.buildenv.2024.111884
UNEP. (2024). Global status report for buildings and construction: Beyond foundations: Mainstreaming
sustainable solutions to cut emissions from the buildings sector. Nairobi: United Nations Environment
Programme. Retrieved 11 28, 2024, from https://www.unep.org/resources/report/global-status-reportbuildings-and-construction
Wang, X., Teigland, R. & Hollberg, A. (2024). Identifying inﬂuential architectural design variables for early-stage
building sustainability optimization. Building and Environment, 252, 111295.
doi:10.1016/j.buildenv.2024.111295
Zhan, J., He, W. & Huang, J. (2024). Comfort, carbon emissions, and cost of building envelope and photovoltaic
arrangement optimization through a two-stage model. Applied Energy, 356, 122423.
doi:10.1016/j.apenergy.2023.122423
Zhang, L. & Chen, Z. (2024). Large language model-based interpretable machine learning control in building
energy systems. Energy and Buildings, 313, 114278. doi:10.1016/j.enbuild.2024.114278
Zhang, L., Fu, X., Li, Y. & Chen, J. (2025). Large language model-based agent Schema and library for automated
building energy analysis and modeling. Automation in Construction, 176, 106244.
doi:10.1016/j.autcon.2025.106244
Zhou, Q. & Xue, F. (2023). Pushing the boundaries of modular-integrated construction: A symmetric skeleton
grammar-based multi-objective optimization of passive design for energy savings and daylight
autonomy. Energy and Buildings, 296, 113417. doi:10.1016/j.enbuild.2023.113417
Zhou, Q. & Xue, F. (2025). Automatic information gain-guided convergence for reﬁning building design
parameters: Enhancing ejectiveness and interpretability in simulation-based optimization. Building and
Environment, 275, 112788. doi:10.1016/j.buildenv.2025.112788
Disclaimer/Publisher’s Note

The statements, opinions, and data contained in all publications are solely those of the individual author(s) and
contributor(s) and do not reﬂect the views of the Architecture, Buildings, Construction and Cities (ABC2) Journal and/or its
editor(s). DFBI Journal and/or its editor(s) disclaim any responsibility for any injury to people or property resulting from any
ideas, methods, instructions, or products referred to in the content.

Proceedings of Smart and Sustainable Built Environment Conference Series

Volume xx, Issue xx