Panagiotis Karapiperis during his final presentation.
Program: MSc Integrated Building Systems (MIBS), ETH Zurich
Thesis duration: 6 months / 30 ECTS
Total number of students supervised: 4 students (main supervisor), 2 students (co-supervisor)
About the thesis. The Master’s Thesis concludes the MSc in Integrated Building Systems at ETH Zurich and provides students with the opportunity to conduct an independent research project on a current topic in the field. The interdisciplinary program brings together architecture, building science, engineering, energy, and computational methods, with a strong emphasis on sustainable and integrated approaches to buildings and urban systems.
My role: As a supervisor, I supported students throughout the development of their thesis projects, from defining the research question and methodology to developing computational workflows, analysing results, and communicating their findings. Depending on the project, my supervision focused particularly on environmental simulation, urban microclimate, and building performance analysis, computer vision, and data-driven methods for the built environment.
List of supervised projects and students:
1.
Adapting pretrained generative video models for dynamic urban wind simulations
Student: Janne Perini (currently at Gartenmann Engineering AG as a Project Manager, Building Physics & Sustainability)
Duration, role: 6 months, co-supervisor
Designing urban spaces that provide pedestrian wind comfort and safety requires time-resolved Computational Fluid Dynamics (CFD) simulations, but their current computational cost makes extensive design exploration impractical. This thesis introduces WinDiNet (Wind Diffusion Network), a pretrained video diffusion model that is repurposed as a fast, differentiable surrogate for this task. Starting from LTX-Video, a 2B-parameter latent video transformer, the model is fine-tuned on 10,000 2D incompressible CFD simulations over procedurally generated building layouts. A systematic study of training regimes, conditioning mechanisms, and VAE adaptation strategies, including a physics-informed decoder loss, identifies a configuration that outperforms purpose-built neural PDE solvers. The resulting model generates full 112-frame rollouts in under a second. As the surrogate is end-to-end differentiable, it doubles as a physics simulator for gradient-based inverse optimization: given an urban footprint layout, building positions are optimized directly through backpropagation to improve wind safety as well as pedestrian wind comfort. Experiments on single- and multi-inlet layouts show that the optimizer discovers effective layouts even under challenging multi-objective configurations, with all improvements confirmed by ground-truth CFD simulations.
.
Wind speed magnitude at t = 24 for inlet speeds from 5 to 29 m/s
(left to right). Out-of-distribution inlet speeds (uin > 20 m/s) are highlighted
with a cyan border. Prepared by Janne Perini.
2.
Augmenting urban building energy models with image data using computer vision methods
Student: Panagiotis Karapiperis (currently at vyzn as an AI Engineer)
Duration, role: 6+3 months, supervisor
Urban Building Energy Modeling (UBEM) is increasingly critical for supporting sustainable urban development and achieving climate targets, yet its adoption is often hindered by a lack of high-resolution building-level input data. This thesis addresses this limitation by introducing a visual-data-driven framework for automated feature extraction from street-level imagery (SVI) to enhance UBEM workflows. Central to this work is the Zurich Facades Dataset, comprising 11,156 building facades annotated with construction age and primary use, alongside an augmented version enriched with natural language descriptions generated through a collaboration between human annotators and generative AI. Extensive model evaluation demonstrates the effectiveness of both state-of-the-art vision models and multimodal large language models (MLLMs), achieving high predictive performance: 81% exact accuracy and 93% +–1 accuracy for building age, and 81% accuracy for primary use classification. Furthermore, a real-world case study involving 1,055 buildings in Zurich shows that the proposed workflow estimates heating demand with only a 3% deviation from true values, while maintaining high spatial resolution of predictions. The findings confirm that SVI-augmented UBEM is a viable and scalable approach, enabling automated, accurate, and spatially detailed energy modeling. This work demonstrates the potential of integrating visual and AI-driven data pipelines into UBEM, paving the way for more precise urban energy assessments to support the decarbonization of the built environment.
.
Comparison of all models in the construction period (GBAUP) prediction results. (Upper) Confusion matrices for the best-performing models. (lower) Evaluation metrics: Accuracy, +–1 Accuracy, Weighted Accuracy. Prepared by Panagiotis Karapiperis.
3.
A catalogue of building-integrated photovoltaic facades
Student: Kai Marti (currently working in London as an Architect)
Duration, role: 3 months, supervisor
Building Integrated Photovoltaic (BIPV) facades can be an important aspect of sustainable architecture, yet their application often remains confined to rooftops. Key obstacles mentioned by stakeholders are limited knowledge and the diversity of products, which hinder their broader adoption. This study analyzes current BIPV facade practices by building a dataset of 406 projects, providing insights into the diversity of applications and their architectural impact. The dataset was classified using a taxonomy that consists of four main categories, including general information, technology, module design, and architectural integration. 13 distinct facade compositions were identified, highlighting how BIPV influences architectural expression. It has been shown that the majority of compositions can be applied to non-PV facades as well; however, compositions such as "ZigZag", "Grid," and "High Tech" have been shown to be more relevant in the context of building with PV as an integral part of architecture. Analysis of the dataset shows an energy-driven design, with a majority of the projects focusing on system efficiency. In general, a similar design practice can be observed between new construction and retrofit buildings when it comes to BIPV applications. The final outcome is a catalogue featuring 24 projects, which show the broad diversity in BIPV facade applications. The projects are classified by the developed taxonomy and act as an inspiration for stakeholders, and reduce the barrier of adoption. The proposal calls for a more diverse integration of PV, placing equal emphasis on the integration into architecture, as well as on its integration into the fields of energy and construction.
.
Page of BIPV Facade Catalogue, showing two projects with “Banner” Design Concept. Prepared by Kai Marti.
4.
An experimental study to quantify the effects of BIPV assemblies on outdoor thermal comfort
Student: Loukas Mettas (currently at Architecture and Building Systems, ETH Zürich as a Scientific Assistant)
Duration, role: 6 months, supervisor
Climate change necessitates the widespread adoption of renewable energy for carbon neutrality. Solar photovoltaic (PV) panels are among the most viable options, due to their low cost and high social acceptance. While the majority of current research regarding BIPV systems focuses on the energy yield, the implications of such assemblies on the local microclimate and outdoor thermal comfort are not fully understood and agreed upon by the literature. This master's thesis, building upon previous simulation work on the same topic, aims to experimentally quantify the effects of BIPV panels in urban conditions, addressing this research gap. To do so, BIPV prototypes that are a result of collaborative work between research institutions and industry are tested under real-world conditions with novel microclimate sensor assemblies. The results indicate that the influence of BIPVs is time and irradiance-dependent. During the daytime, a slight heating effect on the near-surface air temperatures is observed compared with ambient conditions. During nighttime, the opposite effect occurs, with BIPVs reaching slightly sub-ambient temperatures. To further expand on the experimental study, a CFD-based multiphysics model is developed and calibrated with the measured data. The model can be used as a parametric tool, upscaling the results of the experiment and giving additional insights into atmospheric conditions that could not be measured during the experimental campaign. It performs well on predicting the surface and cavity temperatures of the BIPV panels as well as the radiative heat exchange, but struggles to fully capture the convective flows that lead to the small increase in near-surface air temperature.
.
Overview of the final version of the experiment. In the middle, the weather station can be seen. All the sensors are mounted in the same horizontal plane, which corresponds to the half height of the BIPV elements. Captured by Loukas Mettas.
5.
Deep Learning-based Identification of Photovoltaic Facades
Student: Pedram Mirabian (currently at Infrastructure Management, ETH Zürich as a Scientific Assistant)
Duration, role: 6 months, supervisor
With the increasing global energy demand and negative environmental side effects of fossil fuels, the building sector – responsible for a significant portion of emissions – is a critical area for implementing renewable energy technologies. Building-integrated Photovoltaics have emerged as a promising method for addressing these issues and reducing the impact of buildings, especially for urban environments with limited roof area. In this thesis, the feasibility of using Google Street View and Deep Learning techniques in identifying and quantifying existing facade PV installations is assessed. The approach consists of a comprehensive dataset of PV and non-PV facade images, developing Deep Learning models for facade and PV identification by fine-tuning pre-trained models available on the Detectron2 platform, and combining the models along with rectification techniques using vanishing point estimations to calculate an estimated PV-to-surface ratio to provide a more accurate assessment of potential energy yield from PV installations. The results demonstrate the potential of this approach while highlighting issues of data diversity, manual data acquisition and labeling, and model accuracy in PV identification. The developed pipeline contributes to the broader field of sustainable urban development by offering an automated method for analyzing urban imagery. This lays the foundation for further studies investigating facade PV installations based on street-level imagery.
.
Cover image of the thesis report. Prepared by Pedram Mirabian.
6.
Machine-learning-based prediction of photovoltaic cell temperature using CFD simulations
Student: Pit Mathieu (currently at TBF + Partner AG as a Design Engineer)
Duration, role: 6+3 months, co-supervisor
Photovoltaic (PV) technology is integral to sustainable energy solutions, with cell temperature significantly affecting performance. Building Integrated Photovoltaic (BIPV) systems, which incorporate PV panels into the building envelope, face unique thermal challenges that are not well understood by most existing models. This research addresses these challenges by combining multiphysics computational fluid dynamics (CFD) simulations with machine learning (ML) models to predict BIPV cell temperature. A detailed three-dimensional, time-dependent CFD model is constructed using COMSOL Multiphysics, that should accurately capture the physical phenomena in BIPV installations. This model generates time-series cell temperature data under varying conditions. This simulation data is then used to train ML models, aiming for the accuracy of detailed simulations with the efficiency of simpler models. Results demonstrated that this hybrid method is a viable workflow to build PV cell temperature prediction models, highlighting the potential of integrating CFD and ML approaches to enhance the predictive accuracy and efficiency of such models. Additional considerations have given insight into the relevant aspects of the multiphysics model in terms of geometry and air cavity design. Moreover, the ML results provide further confirmation of their predictive performance and highlight the importance of properly representing possible weather conditions in the training data set. Future research should focus on refining these models and exploring their application in diverse environmental conditions and different BIPV configurations, as well as validation through empricical data.
.
Heat Map (XZ temperature distribution) in Kelvin and the corresponding velocity field. This example shows how high inlet velocity leads to a flow field not affected by buoyancy, resulting in a symmetrical temperature distribution on the PV boundary layer. Prepared by Pit Mathieu.