Publications and works
Open-source tools
Thinking with AI on an infinite canvas.
Extended with researchers at the Max Planck Institute for Intelligent Systems in TraceVista.
Features and demos
Features and demos
Dissertation
The dissertation behind the first three alignments: decision-making process, methodological paradigm and interaction pattern alignment.
A · Decision-making process alignment
AFront-Loading Cost Estimation in Early-Stage BIM Design
Brings cost estimation forward into early-stage BIM design.
ABWorkflow-Oriented Convex Decomposition of 2D Building Geometry
Convex decomposition of 2D building geometry, organised around the engineering workflows that use it.
AEvaluation of Building Design Variants in Early Phases on the Basis of Adaptive Detailing Strategies
Evaluates building design variants in early phases, when information is incomplete, on the basis of adaptive detailing strategies.
Organises passive and active building design decisions into two stages and carries their uncertainty through both.
This manuscript introduces a recommendation system as an application extension practice for machine assistance in the sustainable building design domain. By involving multi-objective optimization processes, this system aims to be the bridge...
This research implements a practice of the integration between a data-driven model and knowledge-based methods within the same context in the early phases of building design. The objective is to enhance decision-making...
Drawing inspiration from the human nervous system’s estimation process, we present a key framework designed to revolutionize how we approach complex decision-making challenges across various domains. This framework seamlessly integrates uncertainty quantification,...
B · Methodological paradigm alignment (prior knowledge)
A heterogeneous dataset of European building energy models and simulation results, organised for surrogate modelling.
ABFrom 'What-is' to 'What-if' in Human-Factor Analysis: A Post-Occupancy Evaluation Case
In complex human-centric systems, we may lose ourselves optimizing for symptoms rather than causes. Most analytical methods are designed to answer "what-is" (descriptive) questions, yet we often apply them to "what-if" (interventional)...
BIntegrating symbolic neural networks with building physics: A study and proposal
Explores Kolmogorov–Arnold Networks (KAN) for symbolic learning in building physics, demonstrating improved interpretability, robustness under data scarcity, and potential for physics-informed knowledge discovery.
Introducing a groundbreaking intersection of systems engineering and deep learning: our component-based, data-driven model. This unique approach not only adapts to the inherent structure of systems engineering but also prioritizes domain knowledge,...
BComponent-based machine learning modeling for predicting building energy performance – An overview
An overview of component-based machine learning for predicting building energy performance.
In the engineering domain, most systems are composed of the same basic components with diverse compositions. By disentangling the compositionality from the system and embedding it into the model organization, we actually...
BMachine learning in proton exchange membrane water electrolysis — A knowledge-integrated framework
This paper demonstrates the general applicability of the knowledge-integration machine learning framework. It is showcased through real-world PEMWE case studies, highlighting its potential to drive sustainable energy solutions and pave the way...
BFeasibility Analysis of POD and Deep-autoencoder for Indoor Environment CFD Prediction
Compares proper orthogonal decomposition and deep autoencoders as reduced representations for indoor-environment CFD prediction.
Component-based machine learning to compare topology organisations of district heating systems.
BPathway toward prior knowledge-integrated machine learning in engineering
Proposing a paradigm to systematically embed prior knowledge and domain insights into data-driven models, addressing domain-specific challenges through three levels of knowledge integration: modeling knowledge for system description, inductive logic and disentanglement...
ABIntroducing causal inference in the energy-efficient building design process
In the design and decision-making process, distinguishing between correlation and causation is not just a philosophical mindset but a crucial methodological consideration. As we transition from intuition-driven decisions to data-driven methodologies, it’s...
BA hybrid-model forecasting framework for reducing the building energy performance gap
An intuitive way to combine knowledge-based simulation for data enrichment while introducing extra information to machine learning models. This hybrid-model framework ensures that both domain knowledge and data are well utilized, reducing...
ABA Dynamic Feedforward Control Strategy for Energy-efficient Building System Operation
Leveraging the confluence of simulation, first-principles models, and machine learning, we introduce a pioneering approach that marries the strengths of feedforward and feedback loops. Think of it as a reinforcement learning paradigm:...
BA hybrid-model forecasting framework for reducing the building energy performance gap
Introducing a cutting-edge approach to bridging the persistent energy performance gap in the building industry: By synergistically merging domain knowledge from building simulations with the prowess of machine learning, our hybrid-model forecasting...
Imagine building performance prediction as a dynamic Lego set, where each block represents a distinct component of the design. Just as Lego pieces can be assembled in countless ways to create varied...
C · Interaction pattern alignment (communication)
ACContext Repair for LLMs: Pilot v2
How a wrong claim keeps steering an LLM conversation after its source turn is deleted, and which graph edits remove it, tested on nine model endpoints with exact-match scoring.
This research proposes a symbiosis pattern between human intelligence and machine capabilities with the importance of causal analysis in preventing biases in data-driven models and prior knowledge: the necessity of understanding the...
BCToward reliable signals decoding for electroencephalogram: A benchmark study to EEGNeX
Are the limits of our language the limits of our world? Just as we use words to share ideas, our brains have their own unique language, spoken through electrical signals. To better...
D · Fundamental mechanism alignment
DConstraint Breeds Generalization: Temporal Dynamics as an Inductive Bias
Proposes that physical constraints in biological systems shape dynamics to function not as limitations, but as a temporal inductive bias that breeds generalization. Through phase-space analysis, reveals a fundamental asymmetry between expansive...
DDynamical Alignment: A Principle for Adaptive Neural Computation
Reveals a universal adaptive computation principle in spiking neural networks, showing that matching input temporal structure to neuronal timescales produces two computational modes: dissipative (efficient) and expansive (high-performance), allowing fixed architectures to...
Other
欧盟-德国建筑碳中和前沿; Frontiers of carbon neutrality in EU-German building sector
A thorough overview of carbon neutrality in the EU regarding policy framework, technology support, economics foundation, and recycling strategy.
Under review
A heterogeneous dataset of European building energy models and simulation results, organised for surrogate modelling.
2026
ACContext Repair for LLMs: Pilot v2
How a wrong claim keeps steering an LLM conversation after its source turn is deleted, and which graph edits remove it, tested on nine model endpoints with exact-match scoring.
AFront-Loading Cost Estimation in Early-Stage BIM Design
Brings cost estimation forward into early-stage BIM design.
ABWorkflow-Oriented Convex Decomposition of 2D Building Geometry
Convex decomposition of 2D building geometry, organised around the engineering workflows that use it.
ABFrom 'What-is' to 'What-if' in Human-Factor Analysis: A Post-Occupancy Evaluation Case
In complex human-centric systems, we may lose ourselves optimizing for symptoms rather than causes. Most analytical methods are designed to answer "what-is" (descriptive) questions, yet we often apply them to "what-if" (interventional)...
AEvaluation of Building Design Variants in Early Phases on the Basis of Adaptive Detailing Strategies
Evaluates building design variants in early phases, when information is incomplete, on the basis of adaptive detailing strategies.
2025
DConstraint Breeds Generalization: Temporal Dynamics as an Inductive Bias
Proposes that physical constraints in biological systems shape dynamics to function not as limitations, but as a temporal inductive bias that breeds generalization. Through phase-space analysis, reveals a fundamental asymmetry between expansive...
DDynamical Alignment: A Principle for Adaptive Neural Computation
Reveals a universal adaptive computation principle in spiking neural networks, showing that matching input temporal structure to neuronal timescales produces two computational modes: dissipative (efficient) and expansive (high-performance), allowing fixed architectures to...
BIntegrating symbolic neural networks with building physics: A study and proposal
Explores Kolmogorov–Arnold Networks (KAN) for symbolic learning in building physics, demonstrating improved interpretability, robustness under data scarcity, and potential for physics-informed knowledge discovery.
2024
Introducing a groundbreaking intersection of systems engineering and deep learning: our component-based, data-driven model. This unique approach not only adapts to the inherent structure of systems engineering but also prioritizes domain knowledge,...
BComponent-based machine learning modeling for predicting building energy performance – An overview
An overview of component-based machine learning for predicting building energy performance.
In the engineering domain, most systems are composed of the same basic components with diverse compositions. By disentangling the compositionality from the system and embedding it into the model organization, we actually...
BMachine learning in proton exchange membrane water electrolysis — A knowledge-integrated framework
This paper demonstrates the general applicability of the knowledge-integration machine learning framework. It is showcased through real-world PEMWE case studies, highlighting its potential to drive sustainable energy solutions and pave the way...
The dissertation behind the first three alignments: decision-making process, methodological paradigm and interaction pattern alignment.
Organises passive and active building design decisions into two stages and carries their uncertainty through both.
2023
This research proposes a symbiosis pattern between human intelligence and machine capabilities with the importance of causal analysis in preventing biases in data-driven models and prior knowledge: the necessity of understanding the...
BCToward reliable signals decoding for electroencephalogram: A benchmark study to EEGNeX
Are the limits of our language the limits of our world? Just as we use words to share ideas, our brains have their own unique language, spoken through electrical signals. To better...
BFeasibility Analysis of POD and Deep-autoencoder for Indoor Environment CFD Prediction
Compares proper orthogonal decomposition and deep autoencoders as reduced representations for indoor-environment CFD prediction.
Component-based machine learning to compare topology organisations of district heating systems.
BPathway toward prior knowledge-integrated machine learning in engineering
Proposing a paradigm to systematically embed prior knowledge and domain insights into data-driven models, addressing domain-specific challenges through three levels of knowledge integration: modeling knowledge for system description, inductive logic and disentanglement...
This manuscript introduces a recommendation system as an application extension practice for machine assistance in the sustainable building design domain. By involving multi-objective optimization processes, this system aims to be the bridge...
2022
ABIntroducing causal inference in the energy-efficient building design process
In the design and decision-making process, distinguishing between correlation and causation is not just a philosophical mindset but a crucial methodological consideration. As we transition from intuition-driven decisions to data-driven methodologies, it’s...
This research implements a practice of the integration between a data-driven model and knowledge-based methods within the same context in the early phases of building design. The objective is to enhance decision-making...
BA hybrid-model forecasting framework for reducing the building energy performance gap
An intuitive way to combine knowledge-based simulation for data enrichment while introducing extra information to machine learning models. This hybrid-model framework ensures that both domain knowledge and data are well utilized, reducing...
ABA Dynamic Feedforward Control Strategy for Energy-efficient Building System Operation
Leveraging the confluence of simulation, first-principles models, and machine learning, we introduce a pioneering approach that marries the strengths of feedforward and feedback loops. Think of it as a reinforcement learning paradigm:...
Drawing inspiration from the human nervous system’s estimation process, we present a key framework designed to revolutionize how we approach complex decision-making challenges across various domains. This framework seamlessly integrates uncertainty quantification,...
欧盟-德国建筑碳中和前沿; Frontiers of carbon neutrality in EU-German building sector
A thorough overview of carbon neutrality in the EU regarding policy framework, technology support, economics foundation, and recycling strategy.
2021
BA hybrid-model forecasting framework for reducing the building energy performance gap
Introducing a cutting-edge approach to bridging the persistent energy performance gap in the building industry: By synergistically merging domain knowledge from building simulations with the prowess of machine learning, our hybrid-model forecasting...
Imagine building performance prediction as a dynamic Lego set, where each block represents a distinct component of the design. Just as Lego pieces can be assembled in countless ways to create varied...