Perceptron Neural Network Research Articles

Optimizing pervious concrete with machine learning: Predicting permeability and compressive strength using artificial neural networks

This study makes a significant contribution to the field of pervious concrete by using machine learning to innovatively predict both mechanical and hydraulic performance. Unlike existing methods that rely on labor-intensive trial-and-error experiments, our proposed approach leverages a multilayer perceptron network. To develop this approach, we compiled a comprehensive dataset comprising 271 sets and 3,252 experimental data points. Our methodology involved evaluating 22,246 network configurations, employing Monte Carlo cross-validation over 20 iterations, and using 4 training algorithms, resulting in a total of 1,779,680 training iterations. This results in an optimized model that integrates diverse mix design parameters, enabling accurate predictions of permeability and compressive strength even in the absence of experimental data, achieving R² values of 0.97 and 0.98, respectively. Sensitivity analyses validate the model's alignment with established principles of pervious concrete behavior. By demonstrating the efficacy of machine learning as a complementary tool for optimizing pervious concrete mix designs, this research not only addresses current methodological limitations but also lays the groundwork for more efficient and effective approaches in the field.

Comparison of enhanced neural network and response surface models in predicting bio-dissolution of aluminum and vanadium from bauxite residue by isolated Aspergillus niger strains

Background: Recovery of Al and V from red mud, a hazardous residue of the Bayer process, using bioleaching is a nature-based waste management solution that simultaneously improves environment protection and metallurgical recoveries.Methods: For the first time, a novel hybrid multilayer Perceptron (MLP) network enhanced by an Imperialist Competitive Algorithm (ICA), and a response surface developed based on a factorial experiment design methodology (RSM) were employed and compared for, prediction optimization and optimization prediction of Al and V bioleaching by strains of A. Niger microorganisms isolated from pistachio shells and grape skins. The controlling variables were fungi source (A), adoption strategy (B), solid activation (C), solid percent (D), and bioleaching time. Considering the stochastic nature of ICA, a multi-criteria-ranking system based on accuracy and error indices was developed to select the best MLP. Probability value at 95% confidence interval, lack-of-fit, analysis of variance, R2, Adj. R2, and predicted R2 were the judges for the determination of the developed model's statistical significance.Significant Findings: Based on ANOVA, between A, B, C, and D, the effectiveness orders of A > D > C>AC>AD on Al, and A>AD>C>AC>D on V bioleaching were obtained. The superiority of the developed hybrid ICA-MLP model over the developed RSM model in predicting Al and V dissolution was determined by NSE, RSME, MAE, and MEDAE. Parameters optimization and consequently evaluating optimum condition repeatability through three repeating experiments resulted in maximum dissolution recoveries of RSM: 96.5% for Al and 91.2% for V, and ICA-MLP: 97.1% for Al and 90.3% for V. Considering the lower relative errors of repeating validation tests, it can be concluded that, although both models provide reasonable results, but the ICA-MLP methodology is more reliable (relative error <1.8%).

Fast radiance field reconstruction from sparse inputs

Neural Radiance Field (NeRF) has emerged as a powerful method in data-driven 3D reconstruction because of its simplicity and state-of-the-art performance. However, NeRF requires densely captured calibrated images and lengthy training and rendering time to realize high-resolution reconstruction. Thus, we propose a fast radiance field reconstruction method from a sparse set of images with silhouettes. Our approach integrates NeRF with Shape from Silhouette, a traditional 3D reconstruction method that uses silhouette information to fit the shape of an object. To combine NeRF’s implicit representation with Shape from Silhouette’s explicit representation, we propose a novel explicit–implicit radiance field representation consisting of voxel grids with confidence and feature embedding for geometry and a multilayer perceptron network to decode view-dependent color emission for appearance. We propose to make the reconstructed geometry compact by taking advantage of silhouette images, which can avoid the majority of artifacts in sparse input scenarios and speed up training and rendering. We also apply voxel dilating and pruning to refine the geometry prediction. In addition, we impose a total variation regularization on our model to encourage a smooth radiance field. Experiments on the DTU and the NeRF-Synthetic datasets show that our algorithm surpasses the existing baselines in terms of efficiency and accuracy.

An Ensemble Spectral Prediction (ESP) model for metabolite annotation.

A key challenge in metabolomics is annotating measured spectra from a biological sample with chemical identities. Currently, only a small fraction of measurements can be assigned identities. Two complementary computational approaches have emerged to address the annotation problem: mapping candidate molecules to spectra, and mapping query spectra to molecular candidates. In essence, the candidate molecule with the spectrum that best explains the query spectrum is recommended as the target molecule. Despite candidate ranking being fundamental in both approaches, limited prior works incorporated rank learning tasks in determining the target molecule. We propose a novel machine learning model, Ensemble Spectral Prediction (ESP), for metabolite annotation. ESP takes advantage of prior neural network-based annotation models that utilize multilayer perceptron (MLP) networks and Graph Neural Networks (GNNs). Based on the ranking results of the MLP- and GNN-based models, ESP learns a weighting for the outputs of MLP and GNN spectral predictors to generate a spectral prediction for a query molecule. Importantly, training data is stratified by molecular formula to provide candidate sets during model training. Further, baseline MLP and GNN models are enhanced by considering peak dependencies through label mixing and multi-tasking on spectral topic distributions. When trained on the NIST 2020 dataset and evaluated on the relevant candidate sets from PubChem, ESP improves average rank by 23.7% and 37.2% over the MLP and GNN baselines, respectively, demonstrating performance gain over state-of-the-art neural network approaches. However, MLP approaches remain strong contenders when considering top five ranks. Importantly, we show that annotation performance is dependent on the training dataset, the number of molecules in the candidate set and candidate similarity to the target molecule. The ESP code, a trained model, and a Jupyter notebook that guide users on using the ESP tool is available at https://github.com/HassounLab/ESP.

Enhancing Customer Satisfaction Analysis Using Advanced Machine Learning Techniques in Fintech Industry

Customer satisfaction (CSAT) is vital in service and marketing, indicating how well products or services meet customer expectations. Traditional CSAT methods like the American Customer Satisfaction Index (ACSI) and Net Promoter Score (NPS) face challenges such as survey fatigue and low response rates. This study introduces a novel framework using advanced machine learning (ML) and deep learning (DL) techniques, specifically Bidirectional Encoder Representations from Transformers (BERT), to classify customer feedback into distinct CSAT drivers. Integrating term frequency-inverse document frequency (TF-IDF) methods with BERT-based embeddings, the framework significantly improves prediction accuracy. Using a proprietary dataset of 5,943 customer feedback responses from 39 companies across 13 industries, the fine-tuned BERT model achieved an F1 score of 0.84, surpassing traditional methods like TF-IDF and support vector machine (SVM) with an F1 score of 0.47, and TF-IDF with multi-layer perceptron (MLP) networks at 0.50. A hybrid approach combining BERT and TF-IDF embeddings with MLP networks yielded an F1 score of 0.71. The results show the transformative potential of DL techniques, particularly fine-tuned BERT models, in enhancing CSAT prediction accuracy. This research bridges the gap between traditional and advanced text mining methods, setting a new standard for CSAT modeling and offering a robust framework for extracting actionable insights from customer feedback. It highlights the importance of adopting advanced ML and DL models for strategic decision-making and improving customer satisfaction measurement.

Time series imaging-based deep learning method for inventory control

ABSTRACT Data-driven newsvendor methods offer promising solutions for estimating replenishment quantity within complex markets by leveraging demand-related features. However, most features utilized by these methods rely on manual selection, and some important unobservable features cannot be thoroughly utilized. These issues might result in the loss of some crucial information, thus incurring considerable cost in these decision-making results, meaning that there is still much room for improvement. To remedy this, we propose a time series imaging-based deep learning method, which automatically extracts crucial features alongside those selected manually. This method maximizes the advantages of time series imaging, convolutional neural networks, bidirectional long short-term memory networks, and multilayer perceptron networks. Experiments with four diverse product time series and real-world data show that our approach reduces the average newsvendor problem cost by 16.33% to 33.43% compared to the best benchmark, highlighting its robustness. Additionally, each component within our method significantly contributes to decision-making.

Predicting air quality index and fine particulate matter levels in Bagdad city using advanced machine learning and deep learning techniques

Particulate matter pollution is recognized globally as one of the most hazardous forms of air pollution, profoundly impacting environmental integrity and public health. Key metrics for assessing this pollution include the Air Quality Index (AQI) and fine particulate matter with diameters ≤2.5 μm (PM2.5). These indicators are closely associated with severe health consequences, such as premature death from chronic exposure. While traditional statistical methods have been employed in some studies to evaluate AQI and PM2.5, the application of advanced machine learning techniques has been limited. This research employs deep learning and artificial neural networks (ANN) to forecast AQI and PM2.5 levels in Baghdad, Iraq. The study utilizes an extensive dataset from July 1, 2016, to December 12, 2021, comprising over 48,000 data points for AQI and PM2.5. Time serves as an independent input variable influencing these dependent variables. The analysis employs a diverse set of machine learning algorithms, including random forest, decision tree, K-Nearest Neighbors (KNN), Multi-Layer Perceptron (MLP), and Long Short-Term Memory networks (LSTM). The findings demonstrate that MLP and LSTM models outperform other methods, providing the most accurate predictions. The correlation coefficients were 0.977 and 0.983 for the prediction of AQI and 0.973 and 0.985 for the prediction of PM2.5 using MLP and LSTM, respectively. In addition, the outcomes showed that both AQI and PM2.5 were within the moderate to unhealthy ranges, and their distribution levels pointed to the need for addressing air quality in Baghdad city. Furthermore, this study contributes to the burgeoning field of machine learning applications in environmental science by establishing a robust and nuanced predictive framework for evaluating air quality. It highlights the potential of deep learning in public health applications and offers actionable insights for policymaking to mitigate air pollution and its adverse effects.

Enhancing left ventricular segmentation in echocardiography with a modified mixed attention mechanism in SegFormer architecture

Echocardiography is a key tool for the diagnosis of cardiac diseases, and accurate left ventricular (LV) segmentation in echocardiographic videos is crucial for the assessment of cardiac function. However, since semantic segmentation of video needs to take into account the temporal correlation between frames, this makes the task very challenging. This article introduces an innovative method that incorporates a modified mixed attention mechanism into the SegFormer architecture, enabling it to effectively grasp the temporal correlation present in video data. The proposed method processes each time series by encoding the image input into the encoder to obtain the current time feature map. This map, along with the historical time feature map, is then fed into a time-sensitive mixed attention mechanism type of convolution block attention module (TCBAM). Its output can serve as the historical time feature map for the subsequent sequence, and a combination of the current time feature map and historical time feature map for the current sequence. The processed feature map is then input into the Multilayer Perceptron (MLP) and subsequent networks to generate the final segmented image. Through extensive experiments conducted on two different datasets: Hamad Medical Corporation, Tampere University, and Qatar University (HMC-QU), Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (CAMUS) and Sunnybrook Cardiac Data (SCD), achieving a Dice coefficient of 97.92 % on the SCD dataset and an F1 score of 0.9263 on the CAMUS dataset, outperforming all other models. This research provides a promising solution to the temporal modeling challenge in video semantic segmentation tasks using transformer-based models and points out a promising direction for future research in this field.

Unsupervised heterogeneous domain adaptation for EEG classification

Objective. Domain adaptation has been recognized as a potent solution to the challenge of limited training data for electroencephalography (EEG) classification tasks. Existing studies primarily focus on homogeneous environments, however, the heterogeneous properties of EEG data arising from device diversity cannot be overlooked. This motivates the development of heterogeneous domain adaptation methods that can fully exploit the knowledge from an auxiliary heterogeneous domain for EEG classification. Approach. In this article, we propose a novel model named informative representation fusion (IRF) to tackle the problem of unsupervised heterogeneous domain adaptation in the context of EEG data. In IRF, we consider different perspectives of data, i.e. independent identically distributed (iid) and non-iid, to learn different representations. Specifically, from the non-iid perspective, IRF models high-order correlations among data by hypergraphs and develops hypergraph encoders to obtain data representations of each domain. From the non-iid perspective, by applying multi-layer perceptron networks to the source and target domain data, we achieve another type of representation for both domains. Subsequently, an attention mechanism is used to fuse these two types of representations to yield informative features. To learn transferable representations, the maximum mean discrepancy is utilized to align the distributions of the source and target domains based on the fused features. Main results. Experimental results on several real-world datasets demonstrate the effectiveness of the proposed model. Significance. This article handles an EEG classification situation where the source and target EEG data lie in different spaces, and what’s more, under an unsupervised learning setting. This situation is practical in the real world but barely studied in the literature. The proposed model achieves high classification accuracy, and this study is important for the commercial applications of EEG-based BCIs.

Exploring deep fully convolutional neural networks for surface defect detection in complex geometries

In this paper, we propose a machine learning approach for detecting superficial defects in metal surfaces using point cloud data. We compare the performance of two popular deep learning architectures, multilayer perceptron networks (MLPs) and fully convolutional networks (FCNs), with varying feature sets. Our results show that FCNs (F1=0.94) outperformed MLPs (F1=0.52) in terms of precision, recall, and F1-score. We found that transfer learning with pre-trained models can improve performance when the amount of available data is limited. Our study highlights the importance of considering the amount and quality of training data in developing machine learning models for defect detection in industrial settings with 3D images.

Forecasting Robot Movement with Sensor Readings and Multi-Layer Perceptron Models

Classification of sensor data plays a crucial role in different fields, aiding in important tasks like detecting faults, recognizing events and predicting maintenance needs. This research paper thoroughly explores the use of machine learning methods, in specific Multi-Layer Perceptron (MLP) networks, to classify sensor data. The dataset used in this study includes raw measurements from all 24 ultrasound sensors alongside their corresponding class labels indicating the robots actions (such as moving or turning left). The study tackles challenges like imbalanced classes, noisy signals, and striving for classification through a case study employing MLP models. Through conducting experiments and analysis, we fine-tuned the MLP models setup to achieve a 93.04% accuracy on the test dataset. Additionally evaluation metrics like precision, recall and F1 score highlighted the models effectiveness across different classes. A comparison with Support Vector Machines (SVM) and Logistic Regression models highlighted the performance of the MLP model. These results not only show the effectiveness of MLP networks but also provide valuable insights into best practices, for classifying sensor data. Through examining the intricacies of sensor data analysis and classification this study contributes to enhancing our knowledge of applying machine learning to real world challenges.

Analyzing speed-difference impact on freeway joint injury severities of Leading-Following vehicles using statistical and data-driven models

Rear-end (RE) crashes are notably prevalent and pose a substantial risk on freeways. This paper explores the correlation between speed difference among the following and leading vehicles (Δν) and RE crash risk. Three joint models, comprising both uncorrelated and correlated joint random-parameters bivariate probit (RPBP) approaches (statistical methods) and a cross-stitch multilayer perceptron (CS-MLP) network (a data-driven method), were estimated and compared against three separate models: Support Vector Machines (SVM), eXtreme Gradient Boosting (XGBoost), and MLP networks (all data-driven methods). Data on 15,980 two-vehicle RE crashes were collected over a two-year period, from January 1, 2021, to December 31, 2022, considering two possible levels of injury severity: no injury and injury/fatality for both drivers of following and leading vehicles. The comparative performance analysis demonstrates the superior predictive capability of the CS-MLP network over the uncorrelated/correlated joint RPBP model, SVM, XGBoost, and MLP networks in terms of recall, F-1 Score, and AUC. Significantly, numerous shared variables influence the injury severity outcomes for the following and leading vehicles across both statistical and data-driven approaches. Among these factors, the following vehicle (a truck) and the leading vehicle (a passenger car) demonstrate contrasting effects on the injury severity outcomes for both vehicles. Furthermore, the SHapley Additive exPlanations (SHAP) values from the CS-MLP network visually show the relationship between Δν and injury severity, revealing non-linear trends unlike the average effects shown by statistical methods. They indicate that the least injury outcomes for both following and leading vehicles occurs at a Δν of 0 to 10 mph, matching observed patterns in RE crash data. Additionally, a marked variation in the trend of SHAP values for the two vehicles is noted as the speed difference increases. Therefore, the findings affirm the superior performance of joint model development and substantiate the non-linear impacts of speed difference on injury outcomes. The adoption of dynamic speed control measures is recommended to mitigate the injury outcomes involved in two-vehicle RE crashes.

A study on the application of artificial neural network to predict k-eff and peaking factor of a small modular PWR

Machine learning (ML) using artificial neural network (ANN) methods is being applied to predict required parameters for nuclear reactors based on learning from big data sets. The ML models usually give faster calculation speed while the accuracy is good agreement with physical simulators. In this work, a multi-layer perceptron network was built and trained to predict k-eff and peaking factor of a small modular pressurized water reactor (PWR). The results are compared with those obtained by using a reactor physics code system, i.e. SRAC2006. The comparison shows good agreement accuracy and higher performance of the ML models.

Understanding Childhood Victimization Experiences and Mental Health Outcomes in a Sample of Prison Inmates in Northern Chile

In this paper, the weak points have been investigated, the proposal to solve them, the application of the solution and the comparison of the results in the mode of simulation and experimental testing have been discussed. For this purpose, a four-wheel mobile robot has been considered for simulation and implementation, and a linear quadratic regulator controller and a nonlinear model predictive controller have been used to control it. But the combination of these classic and modern controllers with machine learning can greatly help to make these controllers work more accurately; As a result, in order to increase the accuracy of the performance of these controllers, by training neural networks of multilayer perceptrons, the controllers have been made intelligent. Controllers with cost function have coefficients as weighting to the matrix of system state variables and control input, which are greatly affected by changing these two weighting matrices of problem solving and optimization. For this reason, it is necessary to extract these two matrices for each separate path in order to improve the performance of the controller by trial and error. But by applying the proposed network which is trained with a new algorithm, not only the performance accuracy has increased, but the network extracts these two matrices without the need to spend human energy. Also, in order to reduce the existing time delays, especially in the implementation of the nonlinear controller on the robot in the experimental mode, by training other neural networks to optimally extract the benefit of the forecast horizon, reducing the calculations and increasing the speed of the solution has been achieved. In the hardware part, by examining and using operators such as the pixy camera and the U2D2 interface, which are faster than the usual method, the solution time has been reduced.

Variational temporal convolutional networks for I-FENN thermoelasticity

Machine learning (ML) has been used to solve multiphysics problems like thermoelasticity through multi-layer perceptron (MLP) networks. However, MLPs have high computational costs and need to be trained for each prediction instance. To overcome these limitations, we introduced an integrated finite element neural network (I-FENN) framework to solve transient thermoelasticity problems in Abueidda and Mobasher (2024). This approach used a physics-informed temporal convolutional network (PI-TCN) within a finite element scheme for solving transient thermoelasticity problems. In this paper, we introduce an I-FENN framework using a new variational TCN model trained to minimize the thermoelastic variational form rather than the strong form of the energy balance. We mathematically prove that the I-FENN setup based on minimizing the variational form of transient thermoelasticity still leads to the same solution as the strong form. Introducing the variational form to the ML model brings the advantages of lower requirement for the differentiability of the basis function and, thus, lower memory requirement and higher computational efficiency. Also, it automatically satisfies zero Neumann boundary conditions, thus reducing the complexity of the loss function. The formulation based on the variational form complies with thermodynamic requirements. The proposed loss function reduces the difference between predicted and target data while minimizing the variational form of thermoelasticity equations, combining the benefits of both data-driven and variational methods. In addition, this study uses finite element shape functions for spatial gradient calculations and compares their performance against automatic differentiation. Our results reveal that models leveraging shape functions exhibit higher accuracy in capturing the behavior of the thermoelasticity problem and faster convergence. Adding the variational term and using shape functions for gradient calculations ensure better adherence to the underlying physics. We demonstrate the capabilities of this I-FENN framework through multiple numerical examples. Additionally, we discuss the convergence of the proposed variational TCN model and the impact of hyperparameters on its performance. The proposed approach offers a well-founded and flexible platform for solving fully coupled thermoelasticity problems while retaining computational efficiency, where the efficiency scales proportional to the model size.

Ultra-fast GST-based optical neuron for the implementation of integrated photonic neural networks

A fast-integrated optical neuron is proposed based on a micro-ring resonator structure. The GST phase change material is used in the MRR structure, and neuron switching is achieved by the electrical actuation of this material. The nonlinear activation function of the neuron is extracted through three-dimensional electrical and optical simulations. The neuron is then employed in a three-layer perceptron network using the extracted activation function, and its performance is examined in the classification of handwritten digits. The overall performance of the proposed neuron is compared with the previous optical neural networks. Simulations show an ultra-fast switching time of 200fs, which is one order of magnitude faster than the best-reported value, and an accuracy of 98.9% (in the classification of MNIST handwritten digits dataset) which is among the best results, while the estimated neuron footprint is relatively small.

Antiviral drugs preformulation: thermal energy effects by multilayer perceptron network approach and Hirshfeld surface properties

Antiviral drugs preformulation: thermal energy effects by multilayer perceptron network approach and Hirshfeld surface properties

Fast Prediction of Airfoil Aerodynamic Characteristics Based on a Combined Autoencoder

Aircraft airfoils are classified into two main categories: symmetrical and asymmetrical. Both types of airfoils have a significant impact on the flight performance and safety of the aircraft. The fast prediction of the aerodynamic coefficients and pressure distributions of airfoils is crucial for the design of aircraft. The traditional wind tunnel test and CFD methods have the disadvantages of high test cost and high time consumption. To solve these problems, a combined autoencoder (CAE) network is proposed in this paper, which can achieve the fast prediction of airfoil aerodynamic coefficients and pressure distributions. The network consists of an airfoil shape autoencoder (AE) network and a multilayer perceptron (MLP) network. Firstly, an autoencoder network reflecting the characteristics of the airfoil shape is established, and the effects of different latent variables on the performance of the autoencoder network are investigated. Then, the latent variables obtained from the autoencoder are concatenated with the inflow conditions such as the Reynolds number and the angle of attack to be used as inputs to the MLP network, and the aerodynamic coefficients of different airfoils in different inflow conditions are predicted. The effects of various latent variable inputs, as well as the direct input of the airfoil shape into the MLP network, on the prediction performance of aerodynamic coefficients are compared and analyzed. The optimal aerodynamic coefficient prediction network is then obtained. Finally, the CAE network is also applied to predict the pressure distributions of different airfoils in different inflow conditions and the effects of different latent variables and input conditions on the prediction performance of the pressure distributions are analyzed and compared with the advantages and disadvantages of the CAE network and the conditional variational autoencoder (CVAE) network. The results demonstrate that the proposed method is capable of accurately predicting aerodynamic characteristics in a shorter time, offering a valuable reference for the fast and efficient design of aircraft airfoils.

Analyzing the effects of streetscape and land use on urban accidents and predicting future accidents by using machine learning algorithms (case study: Mashhad)

In general, land use and layout of streets can have a significant impact on the behavior of drivers and pedestrians. In particular, streetscape has often been overlooked that recognizing the role of streetscape on street accident in urban areas is important. The aim of this research is to investigate the influence of streetscape and land use on urban accidents that occurred in Mashhad between the years 2017 and 2021. To achieve this objective, the study focused on analyzing accidents in three different urban zones. It also considered the land use types adjacent to both closed and open streets, including residential, commercial, and mixed land uses. The research employed various surveys to gather the necessary data and insights related to the targeted areas. Statistics on accident in three zones show that among the mentioned land uses, commercial areas have experienced the highest number of accidents, with their share being approximately three times that of accidents in residential areas. Additionally, 75 % of all accidents took place in areas with open streetscape, whereas accidents in areas with enclosed view accounted for one third of the number of accidents in open streetscape areas. In this research, analysis and modeling were conducted using machine learning algorithms implemented in the Python programming language. Several models were employed, and the best models were selected based on their performance and accuracy, which include Random Forest Regression (RFR), Multilayer Neural Network Perceptron Regression (MLP) and Extreme Boost Gradient Regression (XGBoost). The accuracy of the machine learning models which successfully predicted future outcomes was as follows: Random Forest Regression (RFR) achieved 85 % accuracy, Extreme Boost Gradient Regression (XGBoost) achieved 81 % accuracy, and finally, Neural Network Multilayer Perceptron Regression (MLP) achieved 75 % accuracy.

Prediction of BLEVE-induced response of road tunnel using Transformer network with modified self-attention (SAMT)

Road tunnels might be exposed to Boiling Liquid Expansion Vapor Explosion (BLEVE) due to the transportation of liquified gas tankers passing through road tunnels. An efficient and accurate prediction of the response of road tunnels under internal BLEVEs can facilitate the reliable BLEVE-resistant design and risk assessment of road tunnels. This study introduces an advanced deep-learning model that employs a Transformer-based architecture with a modified self-attention mechanism, termed as Self-Attention Modified Transformer (SAMT), to predict BLEVE-induced support rotation of tunnel structure, which is a common criterion in assessing reinforced concrete structure damage to blast loads. Unlike the Transformer with the traditional self-attention mechanism, the proposed SAMT effectively aggregates global information across all variables while mitigating undue dependencies among uncorrelated variables. Consequently, the proposed SAMT is better suited for processing tabular data with uncorrelated variables. The feasibility and advantages of the proposed SAMT are verified by extensive data generated using calibrated numerical models of box-shaped road tunnels subjected to internal BLEVEs. By comparing the performance of the proposed SAMT with the non-modified Transformer network (FT-Transformer) as well as two other typical deep learning networks, i.e., Multi-layer Perceptron (MLP) and Residual Network (ResNet), it is found that the SAMT offers higher prediction accuracy and robustness than the other three models in predicting BLEVE-induced support rotations of box-shaped road tunnels. The study demonstrates that the proposed SAMT is an effective tool for the prediction of BLEVE-induced support rotations of road tunnels.