Umeå University's logo

The advancement of artificial intelligence (AI) in material design and engineering has led to significant improvements in predictive modeling of material properties. However, the lack of interpretability in machine learning (ML)-based material informatics presents a major barrier to its practical adoption. This study proposes a novel quantitative computational framework that integrates ML models with explainable artificial intelligence (XAI) techniques to enhance both predictive accuracy and interpretability in material property prediction. The framework systematically incorporates a structured pipeline, including data processing, feature selection, model training, performance evaluation, explainability analysis, and real-world deployment. It is validated through a representative case study on the prediction of high-performance concrete (HPC) compressive strength, utilizing a comparative analysis of ML models such as Random Forest, XGBoost, Support Vector Regression (SVR), and Deep Neural Networks (DNNs). The results demonstrate that XGBoost achieves the highest predictive performance ((Formula presented.)), while SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-Agnostic Explanations) provide detailed insights into feature importance and material interactions. Additionally, the deployment of the trained model as a cloud-based Flask-Gunicorn API enables real-time inference, ensuring its scalability and accessibility for industrial and research applications. The proposed framework addresses key limitations of existing ML approaches by integrating advanced explainability techniques, systematically handling nonlinear feature interactions, and providing a scalable deployment strategy. This study contributes to the development of interpretable and deployable AI-driven material informatics, bridging the gap between data-driven predictions and fundamental material science principles.

The utilization of immersive virtual environments (IVEs) has emerged as a pivotal tool in enhancing observation of occupant-building interaction (OBI) in non-existing and pre-operational buildings (e.g., buildings under-designed, renovated, and retrofitted). The data derived from IVEs are critical in developing Building Predictive Models (BPMs) that prioritize occupant comfort and optimize building performance. Nevertheless, a persistent challenge is the collection of sufficiently large sample sizes from IVEs, often resulting in data sets inadequate for creating accurate and dependable BPMs. To address the gap, the generation of synthetic data is one promising solution. Mega-trend diffusion (MTD) is particularly adept at managing the nuances of small, mixed-type, and imbalanced data sets aligning with the natures of the IVE data sets. This study explores MTD-based algorithms such as baseline MTD, baseline MTD with class probability function, and k-Nearest Neighbors MTD (kNNMTD), all of which are adept at addressing the inherent data challenges. Various small data sets associated with OBI in IVEs were used to test these algorithms. The fidelity of the synthetic data sets is assessed using the Pairwise Correlation Difference (PCD) and accuracy of Artificial Neural Networks (ANNs) trained on the synthetic data sets with several modeling structures. A variety of findings indicated strength and limitations of the algorithms, where some areas need further investigation. At this stage, the evaluation based on this study found that the kNNMTD produced synthetic data sets that were closest to the experimental data set (i.e., the smallest PCD), contributing to the most accurate ANN models.

As part of Sweden's commitment to carbon neutrality, various municipalities have established energy efficiency targets. Achieving these targets requires decision-making knowledge at the local level, particularly concerning the energy retrofitting of existing building stocks. It includes understanding of the thermal properties (U-values) of building stock, their impact on energy use, and the potential energy retrofits. Our research focuses on assessing the thermal properties performance and their impact on energy use of residential building stocks in Umeå, Sweden. We employ explainable artificial intelligence (XAI) integrated with machine learning regression framework to elucidate how different building thermal features influence the building's energy use and also the correlations among these features. The findings highlight the significant impact of building floor area on energy use, followed by location, age, etc. Among thermal properties, the exterior walls have high impact and attic floor has the lowest impact on the energy use of Umeå's residential building stock. Ultimately, this study provides municipality-level decision-making insights for planning energy retrofitting initiatives for Umeå building stock.

We've devised an interpretable stochastic integrated machine learning method for predicting thermal conductivity in Polymeric Graphene-Enhanced Composites (PGECs) across different scales. Our approach, integrating techniques like Regression Trees methods (Gradient Boosting machine), handles uncertain parameters via a bottom-up multi-scale framework using Representative Volume Elements in Finite Element Modeling (RVE-FEM). To understand factors affecting predictions, we use the SHapley Additive exPlanations (SHAP) algorithm. This method aligns well with training data, offering promise for efficiently designing new composite materials for thermal management.

Energy-efficient renovations significantly affect how people use buildings, and these occupant behaviors, in turn, influence the effectiveness of building renovations. Exploring interactions between occupants and renovations is essential for implementing building energy-efficient renovation. However, physical experiments for this purpose require extensive setups in the laboratory to observe occupant behaviors under various renovations. Immersive virtual environment (IVE) experiments as an emerging method still need to adequately incorporate thermal stimuli, essential for studying occupant behaviors, building renovations, and related interactions. Therefore, this study proposes a novel approach that integrates virtual and physical environments in experiments to explore occupant behaviors and their interactions with building renovations. The interactive and immersive capabilities of IVE experiments allow for effective simulation of various renovations and occupant behaviors. By incorporating thermal stimuli from physical experiments, this approach overcomes previous limitations in studying thermal-related occupant behaviors. In a field study, an office building looking for renovation is used to explore occupant behaviors and their interactions with building renovations. It is found that energy-efficient renovation impacts personal heater use and door opening behaviors, but not clothing behaviors; such changes in heater use subsequently impact the energy performance of building renovation. In further analysis, obvious correlations are revealed between personal heater use and renovation scenarios, thermal perception, and times of day. The proposed approach is validated as a novel method to engage the occupants in achieving occupant-centric building energy-efficient transitions.

Modelling the performance of building stocks is crucial in facilitating the renovation at the building stock level. Bottom-up building stock modelling begins by detailing individual buildings and then aggregates them into stock level. Its primary advantage lies in capturing the inherent heterogeneity among distinct buildings, which enables tailored retrofitting. Naturally, this approach requires a comprehensive dataset with detailed building information such as geometry and envelope thermal properties. However, a common challenge is the incompleteness of available data in individual datasets. To address this, previous bottom-up studies have filled the missing data with representative or statistical data. Such practice could lead to homogeneous modelling of distinct buildings within the same statistical group. This limits the utilization of key ability of bottom-up building stock modelling in capturing heterogeneity, such as tailored retrofitting to explore potential retrofitting areas and strategies. To address this challenge of homogeneous modelling, we utilize data fusion framework for bottom-up building stock modelling, employing probabilistic record linkage and inverse modelling techniques to integrate multiple incomplete building performance datasets. This framework fills the missing data in one dataset with information from another, thus capturing inherent heterogeneity in the building stock. An empirical study was conducted in Umeå, Sweden, to investigate the framework's effectiveness by modelling building stock with various retrofitting strategies. This study contribution lies in enhancing bottom-up building stock modelling by capturing inherent heterogeneity, to provide tailored retrofitting solutions.

Temperature extreme weather is rapid changes in heat or cold temperatures that occur suddenly and persists for days to weeks, having an important impact on inhabitants' indoor living comfort and health. During temperature extreme weather, inhabitants may experience indoor overheating or overcooling and engage in self-adaptation to build their own resilience against these threats. However, self-adaptation alone may not adequately address the climate change challenge, external support is necessary in this process. This study aims to configure an experimental platform in the Intelligent Human-Buildings Interactions lab (IHBI) at Umeå University, to create a well-controlled experimental environment to investigate how inhabitants adapt to temperature extreme weather. The IHBI lab is developed by a hybrid virtual-physical framework, with the ability to simulate extreme weather and observe inhabitants' adaptive behaviors, facilitating the development of behavioral guidelines, interventions, and other external supports for inhabitants' adaptations. A pilot study validated the accuracy and authenticity of the developed experimental platform.

Polyurethane (PU) possesses excellent thermal properties, making it an ideal material for thermal insulation. Incorporating Phase Change Materials (PCMs) capsules into Polyurethane has proven to be an effective strategy for enhancing building envelopes. This innovative design substantially enhances indoor thermal stability and minimizes fluctuations in indoor air temperature. To investigate the thermal conductivity of the Polyurethane-Phase Change Materials foam composite, we propose a hierarchical multi-scale model utilizing Physics-Informed Neural Networks (PINNs). This model allows accurate prediction and analysis of the material’s thermal conductivity at both the meso-scale and macro-scale. By leveraging the integration of physics-based knowledge and data-driven learning offered by Physics-Informed Neural Networks, we effectively tackle inverse problems and address complex multi-scale phenomena. Furthermore, the obtained thermal conductivity data facilitates the optimization of material design. To fully consider the occupants’ thermal comfort within a building envelope, we conduct a case study evaluating the performance of this optimized material in a detached house. Simultaneously, we predict the energy consumption associated with this scenario. All outcomes demonstrate the promising nature of this design, enabling passive building energy design and significantly improving occupants’ comfort. The successful development of this Physics-Informed Neural Networks-based multi-scale model holds immense potential for advancing our understanding of Polyurethane-Phase Change Material’s thermal properties. It can contribute to the design and optimization of materials for various practical applications, including thermal energy storage systems and insulation design in advanced building envelopes.

We introduce an interpretable stochastic integrated machine learning based multiscale approach for the prediction of the macroscopic thermal conductivity in Polymeric graphene-enhanced composites (PGECs). This method encompasses the propagation of uncertain input parameters from the meso to macro scale, implemented through a foundational bottom-up multi-scale framework. In this context, Representative Volume Elements in Finite Element Modeling (RVE-FEM) are employed to derive the homogenized thermal conductivity. Besides, we employ two sets of techniques: Regression-tree-based methods (Random Forest and Gradient Boosting Machine) and Neural networks-based approaches (Artificial Neural Networks and Deep Neural Networks). To ascertain the relative influence of factors on output estimations, the SHapley Additive exPlanations (SHAP) algorithm is integrated. This interpretable machine learning methodology demonstrates strong alignment with published experimental data. It holds promise as an efficient and versatile tool for designing new composite materials tailored to applications involving thermal management.

Evaluating and comparing the performances of different strategies is critical for energy-efficient building retrofitting. Data-driven modelling based on large building performance datasets is an effective method for such evaluations. However, it could be challenging to apply this approach to buildings from data-scarce areas where local building performance datasets have not been well-established, which means the data falls short of the high demand for building retrofitting on a global level. To address this, a transfer learning approach is proposed in this study that can evaluate the performance of buildings without local well-established building performance datasets. The proposed approach is applied in the Swedish-Chinese empirical study that relies on the Swedish dataset to transfer and predict the building performance in China without well-established datasets. It was achieved by applying fuzzy C-means clustering and a neural network (FCM-BRBNN) to pre-train the evaluation model based on the Swedish dataset. Then, the proposed approach collects a small sample of Chinese buildings in the data-scarce area and transfers the model to local building performance prediction. The results show that the transfer learning approach can reliably predict the performance of building retrofitting in data-scarce areas with only hundreds of local building samples. As such, this study provides a novel methodology that can support the evaluation and comparison of retrofitting strategies in data-scarce regions and countries with only limited local data. It could efficiently assist designers in optimizing energy-efficient designs in the pre-retrofit stage. Crucially, the methodology enables the transfer of knowledge regarding building performance across different countries and regions, being pivotal for the international collaboration required to stimulate the global energy-efficiency transformation.