Multilayer Perceptron Neural Network Research Articles

Machine Learning and Deep Learning-Based Multi-Attribute Physical-Layer Authentication for Spoofing Detection in LoRaWAN

The use of wireless sensor networks (WSNs) in critical applications such as environmental monitoring, smart agriculture, and industrial automation has created significant security concerns, particularly due to the broadcasting nature of wireless communication. The absence of physical-layer authentication mechanisms exposes these networks to threats like spoofing, compromising data authenticity. This paper introduces a multi-attribute physical layer authentication (PLA) scheme to enhance WSN security by using physical attributes such as received signal strength indicator (RSSI), battery level (BL), and altitude. The LoRaWAN join procedure, a key risk due to plain text transmission without encryption during initial communication, is addressed in this study. To evaluate the proposed approach, a partially synthesized dataset was developed. Real-world RSSI values were sourced from the LoRa at the Edge Dataset, while BL and altitude columns were added to simulate realistic sensor behavior in a forest fire detection scenario. Machine learning (ML) models, including Logistic Regression (LR), Random Forest (RF), and K-Nearest Neighbors (KNN), were compared with deep learning (DL) models, such as Multi-Layer Perceptron (MLP) and Convolutional Neural Networks (CNN). The results showed that RF achieved the highest accuracy among machine learning models, while MLP and CNN delivered competitive performance with higher resource demands.

Astronomical Seeing Prediction with Machine Learning

Abstract We investigate the use of supervised machine learning models to predict current atmospheric seeing at the Observatorio del Roque de Los Muchachos, La Palma. The prediction is made from current, local meteorological conditions as measured by an on-site weather station. The model was trained over a six-month period from 2024 February 8 to 2024 July 3. We found that a simple random forest method provides better prediction than an multilayer perceptron neural network. The most important predictor is the maximum wind speed over a one-minute interval (0.23), followed by air pressure (0.18), wind speed standard deviation (0.17), relative humidity (0.16), air temperature (0.13), wind direction (0.10) and average wind speed (0.06). All other features had importance <0.05.

Optimization and modeling of sulfur removal from liquid fuel using carbon-based adsorbents through synergistic application of RSM and machine learning

This research investigates the adsorption desulfurization of liquid fuels using artificial neural networks (ANN) and response surface methodology (RSM) approaches. The effectiveness of sulfur removal was predicted by analyzing five important factors: temperature, concentration, surface area, fuel/adsorbent, and time. We employed radial basis function (RBF) and multilayer perceptron (MLP) algorithms for ANN modeling. The optimal MLP configuration, utilizing the Levenberg–Marquardt (Trainlm) algorithm, consisted of three hidden layers with 20, 17, and 9 neurons, respectively, while the optimal RBF network contained 43 neurons. The MLP network’s determination coefficient (R2) was 0.98 over 30 epochs, and its mean squared error (MSE) was 0.0028. The RBF network also obtained an R2 of 0.98 and an MSE of 0.0026 over 40 epochs. A two-factor interaction design served as the basis for the RSM model, which produced an R2 of 0.91. A comparison of the RSM, MLP, and RBF models, using the average absolute relative deviation, indicated that the ANN models, particularly the RBF model, produced more accurate predictions than the RSM model. The findings show that temperature and concentration were the two most significant factors influencing sulfur removal efficiency. Overall, artificial neural networks outperformed the RSM approach in predicting desulfurization performance, providing a more reliable modeling tool for optimizing the sulfur removal process.

Alzheimer’s Prediction Methods with Harris Hawks Optimization (HHO) and Deep Learning-Based Approach Using an MLP-LSTM Hybrid Network

Background/Objective: Alzheimer’s disease is a progressive brain syndrome causing cognitive decline and, ultimately, death. Early diagnosis is essential for timely medical intervention, with MRI medical imaging serving as a primary diagnostic tool. Machine learning (ML) and deep learning (DL) methods are increasingly utilized to analyze these images, but accurately distinguishing between healthy and diseased states remains a challenge. This study aims to address these limitations by developing an integrated approach combining swarm intelligence with ML and DL techniques for Alzheimer’s disease classification. Method: This proposal methodology involves sourcing Alzheimer’s disease-related MRI images and extracting features using convolutional neural networks (CNNs) and the Gray Level Co-occurrence Matrix (GLCM). The Harris Hawks Optimization (HHO) algorithm is applied to select the most significant features. The selected features are used to train a multi-layer perceptron (MLP) neural network and further processed using a long short-term (LSTM) memory network in order to classify tumors as malignant or benign. The Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset is utilized for assessment. Results: The proposed method achieved a classification accuracy of 97.59%, sensitivity of 97.41%, and precision of 97.25%, outperforming other models, including VGG16, GLCM, and ResNet-50, in diagnosing Alzheimer’s disease. Conclusions: The results demonstrate the efficacy of the proposed approach in enhancing Alzheimer’s disease diagnosis through improved feature extraction and selection techniques. These findings highlight the potential for advanced ML and DL integration to improve diagnostic tools in medical imaging applications.

Using Machine Learning Methods for Contactless Monitoring of Abnormality in Breathing Patterns

<p dir="ltr"><span>The objective of the present work was to develop and evaluate different methods for measuring and detecting abnormalities in breathing patterns using a pulsed coherent radar sensor. This study represents a crucial initial step toward developing a life-saving hardware product. The contactless monitoring of breathing abnormalities offers significant advantages over traditional methods using contact sensors, particularly in medical scenarios such as post-stroke recovery. Experiments were conducted at three different distances (0.35 − 0.45m, 0.75 − 0.90m, and 1.20 − 1.40m) using a pulsed coherent radar system. Machine learning methods, including kNN (k-Nearest Neighbors) and ANN-MLP (Artificial Neural Networks Multilayer Perceptron), were employed to distinguish whether the monitored individual exhibited normal or abnormal breathing patterns. Our analysis revealed that employing ML models and data signals from the radar sensor holds promise for classifying breathing pattern abnormalities. This paper covers the steps performed to accomplish this: a) Equipment and Data Processing; b) Measurement Trials and Data Collection; c) Study involving k-Nearest Neighbors method and Feed-Forward Neural Networks.</span></p>

Research on supercritical CO2-based filled type U-pipe solar evacuated tubular collector: A thermodynamically parametric optimization and implementation machine learning modeling

Using supercritical CO2 as a working fluid in a heat and power system operating on solar energy is an advantageous approach that achieves higher efficiency, carbon capture, and emissions reduction. Therefore, the present work focused on researching the effectiveness of developing a supercritical CO2-based U-pipe solar evacuated tubular collector (ETC) by a filled layer that includes contributions to thermodynamic parametric optimization and machine learning modeling to forecast the outlet temperature intelligently. Results indicated that the filled type increase the outlet temperature by 24.29 % compared with the conventional type. The evaluation of the optimization research findings shows that an optimal mass flow rate of 8.0 kg/hr achieves a maximum exergy efficiency of 13.70 % to 14.51 %. Furthermore, the results recommended that arranging 13 tubes with a length of 1800 mm is optimal for exergy efficiency within a mass flow rate range of 6.0–12.0 kg/hr. Based on machine learning implementations, Gradient Boosting, Support Vector Machine, and Multi-Layer Perceptron neural networks can intelligently predict with R2|Tout = 1.000, 0.997, and 0.995, respectively. A polynomial-based GMDH/ML with R2|Tout = 0.986 was presented to intelligently forecast the outlet temperature using five input (Tamb,ṁ,ηopt,Ltube,It) parameters.

Detection and Quantification of Water Contamination in Fuel Oil Using Substrate Integrated Waveguide Resonator-Based Microwave Sensors Coupled with Multilayer Perceptron Neural Networks

Detection and Quantification of Water Contamination in Fuel Oil Using Substrate Integrated Waveguide Resonator-Based Microwave Sensors Coupled with Multilayer Perceptron Neural Networks

Artificial Neural Network Multi-layer Perceptron Models to Classify California's Crops using Harmonized Landsat Sentinel (HLS) Data

Advances in remote sensing and machine learning are enhancing cropland classification, vital for global food and water security. We used multispectral Harmonized Landsat 8 Sentinel-2 (HLS) 30-m data in an artificial neural network (ANN) multi-layer perceptron (MLP) model to classify five crop classes (cotton, alfalfa, tree crops, grapes, and others) in California's Central Valley. The ANN MLP model, trained on 2021 data from the United States Department of Agriculture's Cropland Data Layer, was validated by classifying crops for an independent year, 2022. Across the five crop classes, the overall accuracy was 74%. Producer's and user's accuracies ranged from 65% to 87%, with cotton achieving the highest accuracies. The study highlights the potential of using deep learning with HLS time series data for accurate global crop classification.

Free-Phish: detecting phishing websites hosted on free web hosting domains

Phishing attacks have changed over time, and a family of phishing websites hosted over free web hosting domains (FHDs) are increasing. These can be created and maintained at scale with minimal effort while also effectively evading prevalent anti-phishing detection and resisting website takedowns. Their coverage time is also slower; hence, it's challenging for anti- phishing tools to catch them. From August 2022 to April 2023, more than 5,000 of these FHD URLs were observed on phish- tank.com and openphish.com by 41 different FHD providers. This paper investigates a family of phishing websites hosted on free web hosting domains (FHDs). We have developed an updated dataset containing both FHD and non-FHD URLs. While we leveraged some pre-existing features, we also introduced two novel features unique to this study. This work provides the first comprehensive study focused on distinguishing FHD and non-FHD URLs, filling a notable gap in the literature. Our dataset comprises 4,922 recently collected FHD URLs and 5,010 non-FHD URLs. By employing a multilayer perceptron neural network with 32 features, we achieved an impressive accuracy rate of 96.5% for identifying FHD URLs. Additionally, using a random forest classifier, we achieved an accuracy of 97.7% for non-FHD URLs.

Log-sigmoid Activation Function based MLP Network for Aggregate Classification

Mechanical sifting and manual grading have conventionally been utilised to assess the grade of aggregates. Nonetheless, such evaluations require a range of mechanical, chemical, and physical examinations, typically conducted manually, resulting in a process that is tedious, subjective, and labour-intensive. This research aims to provide an image-based classification system for the categorisation of aggregates. An artifi cial neural network (ANN) has been used to analyse the acquired images and categorise their shapes. The composite images are obtained and utilised as the input parameter for prediction prior to the thresholding step. The Log-sigmoid (Logsig) activation function, utilised in a Multilayer Perceptron (MLP) network, exhibits a lower mean square error (MSE) and superior regression performance relative to the Pureline activation functions. The Logsig-based network has a MSE of 1.7473 and a regression capability of 0.9521.

Improving the accuracy of soil texture determination using pH and electro conductivity values with ultrasound penetration-based digital soil texture analyzer

Soil texture analysis is critical for advancing agricultural productivity, ensuring environmental sustainability, and maintaining ecosystem balance. Traditional sedimentation-based methods, such as the hydrometer technique, are fast and practical but prone to inaccuracies due to the effects of water-soluble substances. This study focuses on the practical framework of integrating pH (potential of hydrogen) and EC (electrical conductivity), as indicators of dissolved substances that influence soil texture estimation. Using the Ultrasound Penetration-based Digital Soil Texture Analyzer (USTA), this research combined ultrasound time series data with pH and EC measurements to predict sand, silt, and clay ratios through machine learning methods—support vector regression (SVR), Random Forest (RF), and multi-layer perceptron neural network (MLPNN). Simulations showed that RF yielded the best results, improving R2 values to 0.52, 0.33, and 0.31 for sand, silt, and clay, respectively. The enhanced model performance demonstrates the viability of integrating pH and EC with advanced machine learning techniques to improve soil texture analysis accuracy. These findings suggest that automated systems like USTA, with modular pH and EC sensors, can provide cost-effective, efficient alternatives to traditional methods, offering practical implications for soil management and agricultural optimization.

Phishing Website Detection: An In‐Depth Investigation of Feature Selection and Deep Learning

ABSTRACTCloud and fog computing technologies benefit from integrating AI‐driven phishing detection as it enhances security, scalability, real‐time reaction, and privacy. Nowadays, there is a noticeable rise in illegal activity taking place online. One of the illicit cybersecurity practices is phishing, in which hackers trick consumers by pretending to be authentic websites and spoofing them to obtain sensitive user information. Phishing attacks, regrettably, have increased dramatically in recent years, according to research. Machine learning (ML) and deep learning (DL) techniques have shown encouraging progress in thwarting these attacks. Consequently, we employed DL and ML techniques to identify phishing websites in this study. This article presents four scenarios in both ML and DL models. Two are proposed in ML, while the others are employed in DL. The outcomes of four scenarios were contrasted to determine which algorithm performed better at distinguishing between legal and illicit websites. Many popular ML techniques were used, including K‐nearest neighbour, random forest (RF), decision trees, and SVMs. PCA and Importance Features are implemented in both ML scenarios to find the best features. RF successfully reached an accuracy of 97.82% using the Importance Feature technique. However, the PCA method failed to improve the performance of ML algorithms. As a result of ML‐based scenarios, 98 features are selected for the final deep learning scenarios. In DL‐based scenarios, algorithm architectures are essential to avoid overfitting and bias due to various hyperparameters. Thus, in the third scenario, our aim focuses on DL architecture design. Multilayer perceptron and convolutional neural networks (CNNs) are employed to detect phishing websites. Finally, our proposed 1D CNN model, using stratified k‐fold cross‐validation, outperformed the classical ML algorithm, achieving 98.94% accuracy and 0.99 AUC‐ROC score in detecting phishing websites.

Development of deep reinforcement learning-based fault diagnosis method for actuator faults in unmanned aerial vehicles

Abstract Actuator faults in unmanned aerial vehicles (UAVs) can have significant and potentially adverse effects on their safety and performance, highlighting the critical importance of fault diagnosis in UAV design. Ensuring the reliability of these systems in various applications often requires the use of advanced diagnostic algorithms. Artificial intelligence methods, such as deep learning and machine learning techniques, enable fault diagnosis through sample-based learning without the need for prior knowledge of fault mechanisms or physics-based models. However, UAV fault datasets are typically small due to stringent safety standards, which presents challenges for achieving high-performance fault diagnosis. To address this, deep reinforcement learning (DRL) algorithms offer a unique advantage by combining deep learning’s automatic feature extraction with reinforcement learning’s interactive learning approach, improving both learning capabilities and robustness. In this study, we propose and evaluate two DRL-based fault diagnosis models, which demonstrate remarkable accuracy in fault diagnosis, consistently exceeding $99{\rm{\% }}$ . Notably, under small sample scenarios, the proposed models significantly outperform traditional classifiers such as decision trees, support vector machines, and multilayer perceptron neural networks. These findings suggest that the integration of DRL enhances fault diagnosis performance, particularly in data-limited environments.

Forecasting Solid Waste Generation in Rodriguez, Rizal Using Artificial Neural Network (Ann) and Regression Analysis: An Input to Municipality’s Solid Waste Management Plan

Accurate and reliable forecasting of metropolitan solid devastate is extremely significant for an effective solid plan for waste management. Every local government is having a constant review of measure in the implementation of its waste management program to ensure its continuous viability and importance. The aim of this paper is to recognize the several influential variables and expand an effective model for authentic forecasting of ‘MSW generation’ and offer strategic recommendations that will improve several practices that are associated with waste management practices and several policy making in Rodriguez, Rizal.Solid waste collection of the municipality from 2010-2022, the population, different households, the GDP, commercial establishments and services (CES), and the tourist arrivalfrom 2010 to 2022 were gathered. Two forecasting methods, the “Artificial Neural Network” (ANN) and the multivariable simple regression with the use of Principal Component Analysis (PCA) have been tested that is for their ability for predicting the annual waste production within the municipality. Among the five components, population, Household, and Commercial Establishments have the highest eigenvalues and it account for almost 86% of the total variance in the original data. Furthermore, these components, present the lowestp-values; to which regression model was developed. Artificial Neural Network (ANN) model was has been established through using Multilayer-perceptron Neural Network. The same factors with normalized importance were identified, the Population, Household and Commercial Establishments. Result showed that “ANN” that is Artificial Neural Network has outperformed regression analysis in predicting the solid waste generation having less quantity in terms of the root RMSE. It was also including “mean error” (ME). At the same time, it also include “mean absolute deviation” (MAD) along with “mean percentage error” (MPE). It was in terms of mean “absolute percentage error” (MAPE).A strategic measure has been recommended to enhance waste management practices and policy making of the municipality of Rodriguez that will improve the accuracy as well as effectiveness of its waste management practices and also reduce environmental impacts.

Exploration of transfer learning techniques for the prediction of PM10

Modelling of pollutants provides valuable insights into air quality dynamics, aiding exposure assessment where direct measurements are not viable. Machine learning (ML) models can be employed to explore such dynamics, including the prediction of air pollution concentrations, yet demanding extensive training data. To address this, techniques like transfer learning (TL) leverage knowledge from a model trained on a rich dataset to enhance one trained on a sparse dataset, provided there are similarities in data distribution. In our experimental setup, we utilize meteorological and pollutant data from multiple governmental air quality measurement stations in Graz, Austria, supplemented by data from one station in Zagreb, Croatia to simulate data scarcity. Common ML models such as Random Forests, Multilayer Perceptrons, Long-Short-Term Memory, and Convolutional Neural Networks are explored to predict particulate matter in both cities. Our detailed analysis of PM10 suggests that similarities between the cities and the meteorological features exist and can be further exploited. Hence, TL appears to offer a viable approach to enhance PM10 predictions for the Zagreb station, despite the challenges posed by data scarcity. Our results demonstrate the feasibility of different TL techniques to improve particulate matter prediction on transferring a ML model trained from all stations of Graz and transferred to Zagreb. Through our investigation, we discovered that selectively choosing time spans based on seasonal patterns not only aids in reducing the amount of data needed for successful TL but also significantly improves prediction performance. Specifically, training a Random Forest model using data from all measurement stations in Graz and transferring it with only 20% of the labelled data from Zagreb resulted in a 22% enhancement compared to directly testing the trained model on Zagreb.

Deep Double Towers Click Through Rate Prediction Model with Multi-Head Bilinear Fusion

The click-through rate (CTR) forecast is among the mainstream research directions in the domain of recommender systems, especially in online advertising suggestions. Among them, the multilayer perceptron (MLP) has been extensively utilized as the cornerstone of deep CTR prediction models. However, current neural network-based CTR prediction models commonly employ a single MLP network to capture nonlinear interactions between high-order features, while disregarding the interaction among differentiated features, resulting in poor model performance. Although studies such as DeepFM have proposed dual-branch interaction models to learn complex features, they still fall short of achieving more nuanced feature fusion. To address these challenges, we propose a novel model, the Deep Double Towers model (DDT), which improves the accuracy of CTR prediction through multi-head bilinear fusion while incorporating symmetry in its architecture. Specifically, the DDT model leverages symmetric parallel MLP networks to capture the interactions between differentiated features in a more structured and balanced manner. Furthermore, the multi-head bilinear fusion layer enables refined feature fusion through symmetry-aware operations, ensuring that feature interactions are aligned and symmetrically integrated. Experimental results on publicly available datasets, such as Criteo and Avazu, show that DDT surpasses existing models in improving the accuracy of CTR prediction, with symmetry contributing to more effective and balanced feature fusion.

SDUST2023BCO: a global seafloor model determined from a multi-layer perceptron neural network using multi-source differential marine geodetic data

Abstract. Seafloor topography, as a fundamental marine spatial geographic information, plays a vital role in marine observation and science research. With the growing demand for high-precision bathymetric models, a multi-layer perceptron (MLP) neural network is used to integrate multi-source marine geodetic data in this paper. A new bathymetric model of the global ocean, spanning 180° E–180° W and 80° S–80° N, known as the Shandong University of Science and Technology 2023 Bathymetric Chart of the Oceans (SDUST2023BCO), has been constructed, with a grid size of 1′ × 1′. The multi-source marine geodetic data used include gravity anomaly data released by the Shandong University of Science and Technology, the vertical gravity gradient and the vertical deflection data released by the Scripps Institution of Oceanography (SIO), and the mean dynamic topography data released by Centre National d'Etudes Spatiales (CNES). First, input and output data are organized from the multi-source marine geodetic data to train the MLP model. Second, the input data at interesting points are fed into the MLP model to obtain prediction bathymetry. Finally, a high-precision bathymetric model with a resolution of 1′ × 1′ has been constructed for the global marine area. The validity and reliability of the SDUST2023BCO model are evaluated by comparing with shipborne single-beam bathymetric data and GEBCO_2023 and topo_25.1 models. The results demonstrate that the SDUST2023BCO model is accurate and reliable, effectively capturing and reflecting global marine bathymetric information. The SDUST2023BCO model is available at https://doi.org/10.5281/zenodo.13341896 (Zhou et al., 2024).

Predicting and Mitigating Delays in Cross-Dock Operations: A Data-Driven Approach

Cross-docking operations are highly dependent on precise scheduling and timely truck arrivals to ensure streamlined logistics and minimal storage costs. Predicting potential delays in truck arrivals is essential to avoiding disruptions that can propagate throughout the cross-dock facility. This paper investigates the effectiveness of deep learning models, including Convolutional Neural Networks (CNN), Multilayer Perceptrons (MLPs), and Recurrent Neural Networks (RNNs), in predicting late arrivals of trucks. Through extensive comparative analysis, we evaluate the performance of each model in terms of prediction accuracy and applicability to real-world cross-docking requirements. The results highlight which models can most accurately predict delays, enabling proactive measures for handling deviations and improving operational efficiency. Our findings support the potential for deep learning models to enhance cross-docking reliability, ultimately contributing to optimized logistics and supply chain resilience.

Simulating fish autonomous swimming behaviours using deep reinforcement learning based on Kolmogorov–Arnold Networks

The study of fish swimming behaviours and locomotion mechanisms holds significant scientific and engineering value. With the rapid advancements in artificial intelligence, a new method combining deep reinforcement learning (DRL) with computational fluid dynamics has emerged and been applied to simulate the fish's adaptive swimming behaviour, where the complex fish behaviour is decoupled to focus on the fish's response to the hydrodynamic field, and the simulation is driven by reward-based objectives to model the fish's swimming behaviour. However, the scale of this cross-disciplinary method is directly affected by the efficiency of the DRL model. To promote it to more general application scenarios, there is a pressing need for further research on more efficient and economical network architectures to address the challenge of approximating state-value function in high-dimensional, dynamic, and uncertain environments. Building upon a previously proposed computational platform for the simulation of fish autonomous swimming behaviour, we integrated Kolmogorov-Arnold Networks(KANs) and tested their performance in point-to-point swimming and Kármán gait swimming environments. Experimental results demonstrated that, compared to long short-term memory Networks(LSTMs) and multilayer perceptron networks(MLPs), the introduction of KANs significantly enhanced the perception and decision-making abilities of the intelligent fish in complex fluid environments. With a smaller network scale, in the point-to-point swimming case, KANs effectively approximated the state-value function, achieving average reward improvements of up to 88.0% and 94.1% over MLPs and LSTMs networks, respectively, and increased by 766.7% and 105.6% in the Kármán gait swimming case. Under comparable network sizes, the intelligent fish with KANs exhibited faster learning capabilities and more stable swimming performance in complex fluid settings.

Efficient and accurate commissioning and quality assurance of radiosurgery beam via prior-embedded implicit neural representation learning.

Dosimetric commissioning and quality assurance (QA) for linear accelerators (LINACs) present a significant challenge for clinical physicists due to the high measurement workload and stringent precision standards. This challenge is exacerbated for radiosurgery LINACs because of increased measurement uncertainty and more demanding setup accuracy for small-field beams. Optimizing physicists' effort during beam measurements while ensuring the quality of the measured data is crucial for clinical efficiency and patient safety. To develop a radiosurgery LINAC beam model that embeds prior knowledge of beam data through implicit neural representation (NeRP) learning and to evaluate the model's effectiveness in guiding beam data sampling, predicting complete beam dataset from sparse samples, and verifying detector choice and setup during commissioning and QA. Beam data including lateral profile and tissue-phantom-ratio (TPR), collected from CyberKnife LINACs, were investigated. Multi-layer perceptron (MLP) neural networks were optimized to parameterize a continuous function of the beam data, implicitly defined by the mapping from measurement coordinates to measured dose values. Beam priors were embedded into network weights by first training the network to learn the NeRP of a vendor-provided reference dataset. The prior-embedded network was further fine-tuned with sparse clinical measurements and used to predict unacquired beam data. Prospective and retrospective evaluations of different beam data samples in finetuning the model were performed using the reference beam dataset and clinical testing datasets, respectively. Model prediction accuracy was evaluated over 10 clinical datasets collected from various LINACs with different manufacturing modes and collimation systems. Model sensitivity in detecting beam data acquisition errors including inaccurate detector positioning and inappropriate detector choice was evaluated using two additional datasets with intentionally introduced erroneous samples. Prospective and retrospective evaluations identified consistent beam data samples that are most effective in fine-tuning the model for complete beam data prediction. Despite of discrepancies between clinical beam and the reference beam, fine-tuning the model with sparse beam profile measured at a single depth or with beam TPR measured at a single collimator size predicted beam data that closely match ground truth water tank measurements. Across the 10 clinical beam datasets, the averaged mean absolute error (MAE) in percentage dose was lower than 0.5% and the averaged 1D Gamma passing rate (1%/0.5mm for profile and 1%/1mm for TPR) was higher than 99%. In contrast, the MAE and Gamma passing rates were above 1% and below 95% between the reference beam dataset and clinical beam datasets. Model sensitivity to beam data acquisition errors was demonstrated by significant model prediction changes when fine-tuned with erroneous versus correct beam data samples, as quantified by a Gamma passing rate as low as 18.16% between model predictions. A model for small-field radiosurgery beam was proposed that embeds prior knowledge of beam properties and predicts the entire beam data from sparse measurements. The model can serve as a valuable tool for clinical physicists to verify the accuracy of beam data acquisition and promises to improve commissioning and QA reliability and efficiency with substantially reduced number of beam measurements.