Transfer Learning Approach Research Articles

Breast histopathological imaging using ultra-fast fluorescence confocal microscopy to identify cancer lesions at early stage.

Ultrafast fluorescent confocal microscopy is a hypothetical approach for breast cancer detection because of its potential to achieve instantaneous, high-resolution images of cellular-level tissue features. Traditional approaches such as mammography and biopsy are laborious, invasive, and inefficient; confocal microscopy offers many benefits over these approaches. However, confocal microscopy enables the exact differentiation of malignant cells, the expeditious examination of extensive tissue sections, and the optical sectioning of tissue samples into tiny slices. The primary goal should be to prevent cancer altogether, although detecting it early can help achieve that objective. This research presents a novel Breast Histopathology Convolutional Neural Network (BHCNN) for feature extraction and recursive feature elimination method for selecting the most significant features. The proposed approach utilizes full slide images to identify tissue in regions affected by invasive ductal carcinoma. In addition, a transfer learning approach is employed to enhance the performance and accuracy of the models in detecting breast cancer, while also reducing computation time by modifying the final layer of the proposed model. The results showed that the BHCNN model outperformed other models in terms of accuracy, achieving a testing accuracy of 98.42% and a training accuracy of 99.94%. The confusion matrix results show that the IDC positive (+) class achieved 97.44% accuracy and 2.56% inaccurate results, while the IDC negative (-) class achieved 98.73% accuracy and 1.27% inaccurate results. Furthermore, the model achieved less than 0.05 validation loss. RESEARCH HIGHLIGHTS: The objective is to develop an innovative framework using ultra-fast fluorescence confocal microscopy, particularly for the challenging problem of breast cancer diagnosis. This framework will extract essential features from microscopy and employ a gradient recurrent unit for detection. The proposed research offers significant potential in enhancing medical imaging through the provision of a reliable and resilient system for precise diagnosis of breast cancer, thereby propelling the progression of state-of-the-art medical technology. The most suitable feature was determined using BHRFE optimization techniques after retrieving the features by proposed model. Finally, the features chosen are integrated into a proposed methodology, which is then classified using a GRU deep model. The aforementioned research has significant potential to improve medical imaging by providing a complex and reliable system for precise evaluation of breast cancer, hence advancing the development of cutting-edge medical technology.

Small molecule inhibitors of IL-1R1/IL-1β interaction identified via transfer machine learning QSAR modelling

The human interleukin-1 receptor I (IL-1R1) is a cytokine receptor recognized by interleukin 1β (IL-1β), among other cytokines. Over activation of IL-1R1 has been implicated in various inflammatory conditions. This research aims to identify small-molecule inhibitors targeting the hIL1R1/IL1β interaction, employing a multi-task transfer learning approach for quantitative structure-activity relationship (QSAR) modelling. A comprehensive bioactivity dataset from functionally related proteins was utilised to build a robust ensemble machine learning model for predicting IC50 values against the target protein. Despite the availability of antibody-based therapies, the absence of orally available small-molecule inhibitors necessitates their development. By combining model predictions with docking and simulation approaches, the interleukin-1 receptor inhibitor (IRI-1) emerged as a lead compound. It potently inhibited human IL1-R1 with micromolar activity in THP-1 and Saos-2 cells and demonstrated good biocompatibility. Western blot analysis revealed that IRI-1 inhibits IL-1β-mediated phosphorylation of IL1-R1, JNK, IRAK-4, and ERK in THP-1 cells. Furthermore, molecular dynamics simulations confirmed the structural stability of the protein-ligand complexes. This study highlights the effectiveness of multi-task transfer learning approaches for building robust QSAR models against novel proteins or those with limited bioactivity data, such as hIL-1β/IL-1R1 protein.

Abstract B010: Improving pancreatic cancer image classification with transfer learning: CD45 vs. PANCK

Abstract Pancreatic ductal adenocarcinoma (PDAC) is known for its high aggressiveness and is the fourth most common cause of death of cancer-related death in the United States. It is highly malignant and often goes undiagnosed until advanced stages due to its late-onset symptoms, resulting in poor prognosis. Using quantitative analyses of histological features, such as the expression of prognostic biomarker, can provide valuable insight to guide treatment decisions, monitor disease progression and predict outcomes. Previous researchers have found that CD45 and PANCK play a huge role in characterizing the tumor microenvironment of PDAC. One study using single-cell RNA sequencing compared the microenvironments of healthy pancreata and pancreatic tumors, revealing differences in cellular composition. Tumor samples showed a higher proportion of immune cells, including CD45+ cells, compared to heathy samples and utilized PANCK to differentiate between epithelial and non-epithelial cells. This study aims to enhance the accuracy of classifying pancreatic cancer pathological images based on prognostic markers using transfer learning techniques. We curated and pre-processed a dataset of pancreatic cancer pathological images with different prognostic markers. Using the pre-trained convolutional neural network (CNN) model VGG16, we fine-tuned it with our dataset to better differentiate between two prognostic markers, CD45 and PANCK. The model's performance was evaluated using a similarity matrix to illustrate distinctions among the two prognostic markers. Our results demonstrate that the transfer learning approach significantly improves classification accuracy, achieving confusion matrix validation results of 90% between CD45 and PANCK prognostic markers. SSIM values were relatively low, highlighting the ability of our model to differentiate between the two prognostic markers. The low SSIM value also shows variability within the two prognostic markers, indicating heterogeneity in structural changes due to pancreatic cancer. These findings indicate that transfer learning can be a powerful tool for improving the classification of pathological images and that there is notable variability in the structural features associated with different prognostic markers in pancreatic cancer. Citation Format: Miracle Thomas. Improving pancreatic cancer image classification with transfer learning: CD45 vs. PANCK [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Tumor-body Interactions: The Roles of Micro- and Macroenvironment in Cancer; 2024 Nov 17-20; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2024;84(22_Suppl):Abstract nr B010.

Neural network–based transfer learning to improve stiffness modeling of industrial robots with small experimental data sets

AbstractStiffness modeling is an essential subject for the composition of robot control. Accurate stiffness modeling is helpful for improving the control accuracy of industrial robots, particularly under dynamic load circumstances. The classic virtual joint modeling (VJM) method is challenging in predicting the deformation of the end-effector throughout the full workspace due to the nonlinear deformation of the robot joint and its serial articulated structure. This paper proposes a full-space stiffness modeling method for robots based on the integration of a multi-layer perceptual (MLP) model and VJM. To provide enough training data for the MLP model, VJM is used to build a stiffness model with a small set of experimental data to generate 106,400 training data. A model-based transfer learning approach is proposed to improve the model’s accuracy and generalization regarding the difference between generated training data and actual experimental data. The VJM stiffness model is compared with the MLP stiffness model and the existing CNN-based transfer learning model based on the same experimental data. Considering the deformation prediction in the three directions in Cartesian space, the mean absolute error, standard deviation, and maximum error of the MLP model are decreased by at least 24.90%, 14.20%, and 8.50%, respectively, than the VJM. These prediction results demonstrate that the proposed modeling technique can significantly increase the accuracy of robot stiffness modeling, which is essential for position compensation in precise motion control of robots under dynamic load.

Adaptive Compression-Aware Split Learning and Inference for Enhanced Network Efficiency

The growing number of AI-driven applications in mobile devices has led to solutions that integrate deep learning models with the available edge-cloud resources. Due to multiple benefits such as reduction in on-device energy consumption, improved latency, improved network usage, and certain privacy improvements, split learning, where deep learning models are split away from the mobile device and computed in a distributed manner, has become an extensively explored topic. Incorporating compression-aware methods (where learning adapts to compression level of the communicated data) has made split learning even more advantageous. This method could even offer a viable alternative to traditional methods, such as federated learning techniques. In this work, we develop an adaptive compression-aware split learning method (“deprune”) to improve and train deep learning models so that they are much more network-efficient, which would make them ideal to deploy in weaker devices with the help of edge-cloud resources. This method is also extended (“prune”) to very quickly train deep learning models through a transfer learning approach, which tradesoff little accuracy for much more network-efficient inference abilities. We show that the “deprune” method can reduce network usage by 4× when compared with a split-learning approach (that does not use our method) without loss of accuracy, while also improving accuracy over compression-aware split-learning by up to 4 percent. Lastly, we show that the “prune” method can reduce the training time for certain models by up to 6× without affecting the accuracy when compared against a compression-aware split-learning approach.

Abstract PR019: Transfer learning for accurate tissue of origin classification from cfDNA methylation

Abstract Accurately predicting the tissue of origin (TOO) is required for blood-based multi-cancer screening to guide follow-up imaging and care when cancer is detected. However, precisely determining the TOO using cell-free DNA (cfDNA) methylation remains challenging due to low and variable concentrations of circulating tumor DNA and cfDNA and the complex, poorly characterized nature of tissue-specific methylation patterns. To address these challenges, we developed a transfer learning approach that leverages methylation profiles learned from tissue biopsy samples to predict the TOO of cfDNA cancer samples. Our model achieved 89% balanced accuracy (BA) in classifying the TOO across 10 cancer groups and outperformed models trained solely on plasma cfDNA data. Unmatched tissue biopsy (N=517) and blood (N=1,155, NCT05435066) samples were collected from treatment-naïve cancer patients across 21 tumor types consolidated into 17 cancer groups. All samples were analyzed using a custom targeted bisulfite sequencing hybrid capture assay. Five unique metrics were used to quantify informative methylation signal at each TOO region of interest. A feed-forward neural network was first trained on biopsy methylation features, where the cancer signal is strongest. Subsequently, the model was fine-tuned using plasma cfDNA methylation features, which have significantly lower signal-to-noise ratio, allowing the model to adapt its predictions to the noisier data. The Adam optimizer was employed for both training and fine-tuning, with early stopping implemented to prevent overfitting. Dropout layers were incorporated to enhance model generalization and performance was evaluated using 10-fold cross validation on plasma cfDNA samples across 10 cancer groups for which at least 25 plasma samples were available (N=1,041). The same architecture was trained exclusively on plasma cfDNA samples for comparison. Results are reported for samples with tumor content greater than 0.1% (N=415). Our TOO prediction achieved 73% BA in top-one predictions, where the true label matched the highest probability prediction, and 89% BA in top-two predictions, where the true label matched the highest or the second-highest probability prediction. Multi-class classification revealed our method significantly outperforms a model trained only on plasma cfDNA data. The plasma only model achieved 69% and 82% BA in top-one and top-two predictions, respectively. Accuracy per indication revealed that training on biopsy data significantly improves the detection of tumor types that shed DNA at low rates compared to training on plasma data alone (breast cancer +14%, prostate cancer +12%, and uterine cancer +20% points). In conclusion, our transfer learning approach, utilizing tissue biopsy methylation data, significantly improves cfDNA-based TOO detection accuracy. This approach highlights the capability of transfer learning to enable automated feature extraction from signal-rich data to enhance the analysis of more challenging, heterogeneous cfDNA samples, advancing non-invasive cancer diagnostics. Citation Format: Shiva Farashahi, Amirali Kia, Dorna Kashef, Esther Brown, Feras Hantash, Kieran I Chacko. Transfer learning for accurate tissue of origin classification from cfDNA methylation [abstract]. In: Proceedings of the AACR Special Conference: Liquid Biopsy: From Discovery to Clinical Implementation; 2024 Nov 13-16; San Diego, CA. Philadelphia (PA): AACR; Clin Cancer Res 2024;30(21_Suppl):Abstract nr PR019.

Multisource representation learning for pediatric knowledge extraction from electronic health records.

Electronic Health Record (EHR) systems are particularly valuable in pediatrics due to high barriers in clinical studies, but pediatric EHR data often suffer from low content density. Existing EHR code embeddings tailored for the general patient population fail to address the unique needs of pediatric patients. To bridge this gap, we introduce a transfer learning approach, MUltisource Graph Synthesis (MUGS), aimed at accurate knowledge extraction and relation detection in pediatric contexts. MUGS integrates graphical data from both pediatric and general EHR systems, along with hierarchical medical ontologies, to create embeddings that adaptively capture both the homogeneity and heterogeneity between hospital systems. These embeddings enable refined EHR feature engineering and nuanced patient profiling, proving particularly effective in identifying pediatric patients similar to specific profiles, with a focus on pulmonary hypertension (PH). MUGS embeddings, resistant to negative transfer, outperform other benchmark methods in multiple applications, advancing evidence-based pediatric research.

NIMG-33. HIGH-RESOLUTION SODIUM MRI OF HUMAN GLIOMAS AT 3T USING PHYSICS-BASED GENERATIVE AI

Abstract BACKGROUND Sodium neuroimaging is a promising technique for diagnosing and monitoring brain tumors, providing insights into microenvironment and metabolism. However, at 3T, it is limited by low signal-to-noise ratio and resolution, resulting in long acquisition times and low-quality images. To address these limitations, we developed a physics-informed generative adversarial network (GAN) approach for high-resolution sodium neuroimaging of brain tumors at 3T. METHODS 5,078 anatomical sequences from 1,330 brain tumor patients undergoing routine proton MRI were used to create a synthetic dataset with physics-based simulated artifacts, including B0 magnetic susceptibility artifacts, chemical shift artifacts, Nyquist aliasing artifacts, Gibbs artifacts, and low-resolution images, with Rician noise added during training to augment the dataset. By using multiple proton MRI contrasts, we hypothesized that the GAN model would learn to correct artifacts while preserving the inherent contrast information. We employed a transfer learning approach with a modified Pix2PixGAN architecture and an Attention-R2UNet generator. Twenty glioma patients consented to participate in this IRB-approved prospective study. Sodium and proton MRIs were acquired in the same sessions using a 3T Siemens scanner using a dual-tuned head coil, with a 3D spoiled gradient-echo sequence optimized for short TE. Sodium-proton exchanger (NHE1) expression was tested with immunohistochemistry from image-guided surgical biopsies. RESULTS High-resolution synthetic-sodium MR images were generated using the GAN model. The proposed method successfully reconstructed high-resolution synthetic-sodium MR images with improved SNR compared to traditional sodium images. Synthetic and native measurements were strongly correlated (R2=0.8539, P<0.0001). A significant correlation was seen between synthetic-sodium MR measurements and relative NHE1 tissue expression (ρ=0.5817, P=0.0036). CONCLUSION Results suggest that high-resolution synthetic-sodium MRI retains much of the inherent information from native sodium MR while accurately representing NHE1 expression. This approach could make high-resolution sodium neuroimaging at 3T feasible in the clinical environment, potentially improving diagnosis, monitoring, and treatment of brain tumors.

Pep2TCR: Accurate prediction of CD4 T cell receptor binding specificity through transfer learning and ensemble approach

Pep2TCR: Accurate prediction of CD4 T cell receptor binding specificity through transfer learning and ensemble approach

Defect classification of composite materials using transfer learning methods

ABSTRACT Nowadays, composite materials have become prevalent across various sectors, particularly finding usage in large-scale applications such as spaceships, automobiles, and aircrafts. The accurate detection of the defects in these materials is crucial, yet traditional methods often rely on human inspection, which is susceptible to errors. Recent advancements in machine learning have enabled defect detection using ultrasonic non-destructive testing methods. This paper introduces a new dataset named UNDT, which is obtained from the scans of 60 different composite materials, generating a total of 1150 images depicting both defective and non-defective areas. Several transfer learning methods are applied on the newly introduced UNDT dataset as well as the publicly available USimgAIST ultrasonic dataset. Comparative performance assessments illustrate the significance of utilising the transfer learning approach for defect classification on ultrasonic inspection images. Furthermore, the research emphasises the substantial benefits of employing these transfer learning methods. Notably, the DenseNet121 and VGG19 models achieve the highest accuracy rates, with 98.8% and 98.6% on the UNDT and USimgAIST datasets, respectively.

A Transfer Learning Model for Rolling Bearing Fault Diagnosis Based on Texture Loss Strategy and Nuclear Norm Regularization

Abstract In recent years, transfer learning (TL) approaches have seen extensive application in diagnosing bearing faults due to their exceptional performance. However, mechanical noise, equipment aging, and wear lead to notable disparities and differences in the multi-level feature distributions across the source and target domain signals. The issue is addressed by proposing a transfer learning model based on a texture loss strategy and nuclear norm regularization method. First, a Feature-Enhanced Network (MSRnet) is designed, which significantly improves the ability to capture local details and long-range dependencies by combining a multi-scale feature extraction module with a dilated residual module. Next, a texture loss strategy is proposed to align multi-scale features across domains by minimizing the Gram matrix of signal features. Finally, a nuclear norm regularization method is proposed to perform low-rank approximation on the signal matrix, facilitating the extraction of more robust feature data and mitigating the risk of overfitting. The experimental results demonstrate that the proposed method achieved an average accuracy of 98.58% on the University of Ottawa bearing fault dataset and 98.11% on the Jiangnan University (JNU) bearing dataset, surpassing eight other algorithms in bearing fault diagnosis.

Transfer Learning Approaches for Brain Metastases Screenings

Background: In this study, we examined the effectiveness of transfer learning in improving automatic segmentation of brain metastases on magnetic resonance imaging scans, with potential applications in preventive exams and remote diagnostics. Methods: We trained three deep learning models on a public dataset from the ASNR-MICCAI Brain Metastasis Challenge 2024, fine-tuned them on a small private dataset, and compared their performance to models trained from scratch. Results: Results showed that models using transfer learning performed better than scratch-trained models, though the improvement was not statistically substantial. The custom Tversky and Binary Cross-Entropy loss function helped manage class imbalance and reduce false negatives, limiting missed tumor regions. Medical experts noted that, while fine-tuned models worked well with larger, well-defined tumors, they struggled with tiny, scattered tumors in complex cases. Conclusions: This study highlights the potential of transfer learning and tailored loss functions in medical imaging, while also pointing out the models’ limitations in detecting very small tumors in challenging cases.

Equitable Skin Disease Prediction Using Transfer Learning and Domain Adaptation

In the realm of dermatology, the complexity of diagnosing skin conditions manually necessitates the expertise of dermatologists. Accurate identification of various skin ailments, ranging from cancer to inflammatory diseases, is paramount. However, existing artificial intelligence (AI) models in dermatology face challenges, particularly in accurately diagnosing diseases across diverse skin tones, with a notable performance gap in darker skin. Additionally, the scarcity of publicly available, unbiased datasets hampers the development of inclusive AI diagnostic tools. To tackle the challenges in accurately predicting skin conditions across diverse skin tones, we employ a transfer-learning approach that capitalizes on the rich, transferable knowledge from various image domains. Our method integrates multiple pre-trained models from a wide range of sources, including general and specific medical images, to improve the robustness and inclusiveness of the skin condition predictions. We rigorously evaluated the effectiveness of these models using the Diverse Dermatology Images (DDI) dataset, which uniquely encompasses both underrepresented and common skin tones, making it an ideal benchmark for assessing our approach. Among all methods, Med-ViT emerged as the top performer due to its comprehensive feature representation learned from diverse image sources. To further enhance performance, we conducted domain adaptation using additional skin image datasets such as HAM10000. This adaptation significantly improved model performance across all models.

Knowledge Graph Modeling-Driven Large Language Model Operating System (LLM OS) for Task Automation in Process Engineering Problem-Solving

We present the Process Engineering Operations Assistant (PEOA), an AI-driven framework designed to solve complex problems in the chemical and process industries. The framework employs a modular architecture orchestrated by a meta-agent, which serves as the central coordinator, managing an action generator and instruction-tuned small-scale language models (expert models). The action generator decomposes complex problems into sub-tasks and identifies suitable expert models to execute each, delivering precise solutions for multi-step problem-solving. Key techniques include advanced knowledge modeling using property graphs for improved information retrieval, facilitating more accurate and contextually relevant solutions. Additionally, the framework utilizes a teacher-student transfer-learning approach with GPT-4 (Omni) to fine-tune the action generator and expert models for domain adaptation, alongside an iterative problem-solving mechanism with sophisticated error handling. Custom datasets were developed to evaluate the framework against leading proprietary language models on various engineering tasks. The results demonstrate the framework’s effectiveness in automating calculations, accelerating prototyping, and providing AI-augmented decision support for industrial processes, marking a significant advancement in process engineering capabilities.

Advanced semantic segmentation of aircraft main components based on transfer learning and data-driven approach

Abstract The implementation of Smart Airport and Airport 4.0 visions relies on the integration of automation, artificial intelligence, data science, and aviation technology to enhance passenger experiences and operational efficiency. One essential factor in the integration is the semantic segmentation of the aircraft main components (AMC) perception, which is essential to maintenance, repair, and operations in aircraft and airport operations. However, AMC segmentation has challenges from low data availability, high-quality annotation scarcity, and categorical imbalance, which are common in practical applications, including aviation. This study proposes a novel AMC segmentation solution, employing a transfer learning framework based on a sophisticated DeepLabV3 architecture optimized with a custom-designed Focal Dice Loss function. The proposed solution remarkably suppresses the categorical imbalance challenge and increases the dataset variability with manually annotated images and dynamic augmentation strategies to train a robust AMC segmentation model. The model achieved a notable intersection over union of 84.002% and an accuracy of 91.466%, significantly advancing the AMC segmentation performance. These results demonstrate the effectiveness of the proposed AMC segmentation solution in aircraft and airport operation scenarios. This study provides a pioneering solution to the AMC semantic perception problem and contributes a valuable dataset to the community, which is fundamental to future research on aircraft and airport semantic perception. Graphical abstract

An improved transfer learning method for rotor unbalance position identification from simulated data to experimental data

In the study of dynamic balancing for flexible rotors operating at high speeds, determining the unbalanced position of the rotor has consistently posed significant challenges. Accurate identification of the unbalanced position enables low-speed dynamic balancing to serve as a viable alternative to high-speed methods, ultimately reducing costs and enhancing operational efficiency. Although deep learning techniques utilizing limited labeled data have shown promising results in identifying unbalanced positions, the challenge of gathering a sufficient amount of labeled data for effectively training diagnostic models remains substantial. To address this issue, we propose a Pre-Adaptive Transfer Learning (PATL) approach that employs sample reconstruction through frequency domain correlation analysis. This technique facilitates cross-domain deep transfer recognition of rotor unbalanced positions by transferring insights from simulated dynamic model data to experimental datasets. Compared to existing methodologies, the proposed approach demonstrates significant improvements in both the accuracy and generalization of unbalanced position identification. The novelty of this research lies in the introduction of pre-adaptive transfer learning, which effectively minimizes the disparity between experimental and simulation data, thereby enhancing the model's recognition capabilities. Experimental results indicate that the proposed method achieves effective balancing outcomes across various operating conditions.

Potato disease identification using transfer learning approaches

Potato crop is one of the prominent consumed foods by human beings. When potato crops are infected by diseases it affects farmers negatively and to run in a loss. Therefore, early detection of the potato crop disease can play a vital role in minimizing the loss of the farmers. Nowadays, artificial intelligence technologies, more specifically deep learning techniques, provide solutions to many crops disease-related problems. However, training deep learning models requires a high computational power and huge amount of data as they are data hungry models. Also, designing a custom CNN models a difficult task and there are some variations to be considered. To avoid these difficulties, we adopted two pretrained models of DenseNet121 and VGG19 through transfer learning approaches. The achieved accuracy for DenseNet121 and VGG19 models are 82.6% and 98.56% respectively. DenseNet121 model obtained the average precision, sensitivity, and F1-score of 88.19%, 82.53, and 82.04%, respectively. Whereas VGG19 yields 98.39% of precision, 98.39% of sensitivity, and 97.26% of F1-score in ternary 3-class classification (early blight vs healthy vs late blight).

Textile Fabric Defect Detection Using Enhanced Deep Convolutional Neural Network with Safe Human–Robot Collaborative Interaction

The emergence of modern robotic technology and artificial intelligence (AI) enables a transformation in the textile sector. Manual fabric defect inspection is time-consuming, error-prone, and labor-intensive. This offers a great possibility for applying more AI-trained automated processes with safe human–robot interaction (HRI) to reduce risks of work accidents and occupational illnesses and enhance the environmental sustainability of the processes. In this experimental study, we developed, implemented, and tested a novel algorithm that detects fabric defects by utilizing enhanced deep convolutional neural networks (DCNNs). The proposed method integrates advanced DCNN architectures to automatically classify and detect 13 different types of fabric defects, such as double-ends, holes, broken ends, etc., ensuring high accuracy and efficiency in the inspection process. The dataset is created through augmentation techniques and a model is fine-tuned on a large dataset of annotated images using transfer learning approaches. The experiment was performed using an anthropomorphic robot that was programmed to move above the fabric. The camera attached to the robot detected defects in the fabric and triggered an alarm. A photoelectric sensor was installed on the conveyor belt and linked to the robot to notify it about an impending fabric. The CNN model architecture was enhanced to increase performance. Experimental findings show that the presented system can detect fabric defects with a 97.49% mean Average Precision (mAP).

A hybrid mechanism-based and data-driven model for efficient indoor temperature distribution prediction with transfer learning

Efficient prediction of indoor temperature distribution (ITD) is crucial for various building-related applications, particularly in real-time thermal management. Despite efforts from both physics-based and data-driven perspectives, ITD prediction still faces challenges, particularly in balancing prediction accuracy with computational speed and addressing the need for extensive high-quality data in practical applications. To address these challenges, this study develops a hybrid temperature distribution prediction model (HTDPM) that integrates deep learning with physical mechanisms, enabling the effective capture of spatial-temporal dependencies and uncertainties with limited input parameters, achieving a balance between prediction accuracy and computational efficiency. Meanwhile, a novel transfer learning approach is applied to HTDPM (HTDPM-TL), significantly reducing data requirements and enhancing model generalization across various practical scenarios. Besides, orthogonal and randomized experiments are designed to simulate multiple real-world scenarios, thus constructing an extensive source domain dataset. The HTDPM-TL was tested in various real-world scenarios with several baseline methods, demonstrating an average RMSE of 0.794 ∘C, a computational time of 0.289 s, and limited data volume. The computational speed was improved by three orders of magnitude compared to CFD simulations, and the prediction resolution was enhanced threefold compared to traditional data-driven models. These results highlight the potential of HTDPM-TL to achieve an optimal trade-off between prediction accuracy, computational efficiency, and data requirement.

Updating surrogate models in early building design via tabular transfer learning

Surrogate models provide estimations of design performance objectives, which can yield rapid feedback in early building design. However, constructing parametric models and generating data to train surrogate models is time-consuming. Once established, designers are discouraged from modifying the structure of the design space, which may require additional simulation and retraining. To address this limitation, we propose a new workflow that incorporates a tabular transfer learning approach, coupled with a random walks sampling technique. This approach preserves the knowledge from the initial dataset, thus reducing the number of simulations required to update the surrogate model when new variables are introduced. Through a façade design space case study with a daylighting performance objective, we demonstrate that fewer samples are needed to update the surrogate model and achieve adequate model performance compared to classical interpretable classifiers when only a few new samples are possible or desired. As the number of new samples increases, the performances of some classical methods using traditional sampling improve to eventually meet or surpass the tabular transfer learning method.