Traditional Simulation Methods Research Articles

Wing ice accretion prediction based on conditional generation adversarial network

The ice accretion on the aircraft's surface under low temperatures and high humidity is crucial for flight safety. With respect to the limitation of traditional icing simulation methods, it is very hard to predict exact ice profiles, which can extremely affect the flight performance of an aircraft. A conditional generative adversarial network (CGAN) is utilized to rapidly predict ice accretion and reconstruct three-dimensional ice patterns along the leading edge of a wing. The CGAN is trained using experimental data obtained from a wing with varying sweep angles. The results indicate that the CGAN achieves a high level of accuracy, specifically 97.5%, in predicting the similarity of ice shapes in the test set. When assessing the sample feature capture and prediction capability of the predictive model, it is shown that the CGAN exhibits superior predictive performance across different sample sizes.

Clock mesh synthesis through dynamic programming with physical parameters consideration

In response to the evolving technological landscape, the traditional clock network architecture faces challenges in meeting the complexities of modern System-on-Chip (SoC) designs. While the clock mesh topology offers resilience against On-Chip Variation (OCV) fluctuations, its manual implementation leaves room for advancements in methodology and swift analytical techniques. This paper introduces an innovative clock mesh synthesis approach, leveraging dynamic programming algorithms and emphasizing compliance with critical physical implementation parameters. Our experimental results demonstrate a significant 26.6% reduction in power consumption compared to baseline methodologies. Moreover, it achieves an impressive average runtime reduction of 78.0% when contrasted with traditional simulation methods. These findings underscore the potential of our methodology to enhance the efficiency and power management of clock mesh designs.

Leveraging Visual Language Model and Generative Diffusion Model for Zero-Shot SAR Target Recognition

Simulated data play an important role in SAR target recognition, particularly under zero-shot learning (ZSL) conditions caused by the lack of training samples. The traditional SAR simulation method is based on manually constructing target 3D models for electromagnetic simulation, which is costly and limited by the target’s prior knowledge base. Also, the unavoidable discrepancy between simulated SAR and measured SAR makes the traditional simulation method more limited for target recognition. This paper proposes an innovative SAR simulation method based on a visual language model and generative diffusion model by extracting target semantic information from optical remote sensing images and transforming it into a 3D model for SAR simulation to address the challenge of SAR target recognition under ZSL conditions. Additionally, to reduce the domain shift between the simulated domain and the measured domain, we propose a domain adaptation method based on dynamic weight domain loss and classification loss. The effectiveness of semantic information-based 3D models has been validated on the MSTAR dataset and the feasibility of the proposed framework has been validated on the self-built civilian vehicle dataset. The experimental results demonstrate that the first proposed SAR simulation method based on a visual language model and generative diffusion model can effectively improve target recognition performance under ZSL conditions.

An AI-assisted terahertz reconfigurable metamaterial in standard 180-nm CMOS

For large-scale terahertz (THz) reconfigurable metamaterials (RM), each unit element needs to be controlled by an independent voltage source. This leads to a huge solution space for control words to achieve the required RM performance specification. In this paper, we propose a fast AI-assisted control-word generation scheme to reduce the computation cost of electromagnetic (EM) simulations (Forward Model) and to speed up iterations for control word searches (Inverse Model). We demonstrate the effectiveness of our models by designing and fabricating a 12 × 9 RM array with a Bi-layer Bi-Split-Ring Resonator structure (B-BSRR) in standard 180 nm CMOS technology. The terahertz RM can be electrically reconfigured, providing active metamaterials with an amplitude modulation depth of around 0–1.7 dB at 0.49 THz. The results show that our Forward Model can predict the S-parameters between 100 GHz and 800 GHz with a mean absolute error of 0.61 dB. Our method is more than 200 times faster than traditional full-wave simulation methods. The Inverse Model can generate required control words within a 1.3 dB error margin in less than 200 iterations.

Research on digital twin architecture of ship maneuverability simulation for system model

Abstract With the development of civil and military requirements, ship maneuverability simulation systems tend to be complicated, and the traditional simulation methods are difficult to meet in terms of the development needs of digitalization. The joint simulation technology based on the FMI standard can realize the digital twin of complex system models, which can solve the difficult problem of multidisciplinary and multi-sub-system integration simulation. This paper analyzes the current domestic and foreign research content on joint simulation technology, digital twin technology, and complex system modeling, takes a typical Z-type maneuverability test as an example, establishes a multidisciplinary and cross-platform subsystem model, and designs and develops a ship maneuverability joint simulation algorithm and system software for the complex system model. The simulation results verify the effectiveness of the system modeling technique and the joint simulation technique based on FMI. Based on this system, the whole framework of the ship maneuverability simulation system, which is based on digital twins, is designed. According to the development requirements of ship maneuverability simulation, the main design framework, key technology, and implementation approach of this research method are given, which provides basic methods and ideas for the digital research of ship maneuverability simulation.

Dynamic prediction of fracture propagation in horizontal well hydraulic fracturing: A data-driven approach for geo-energy exploitation

Accurate and efficient prediction of hydraulic fracturing fracture propagation is of utmost importance for unconventional reservoir development. Traditional numerical simulation methods require specialized domain knowledge and technical expertise, and machine learning approach can be an alternative predictive tool for fracture propagation. Additionally, existing neural networks cannot be directly applied to hydraulic fracturing scenarios influenced by multiple coupled factors. Thus, this study presents an innovative approach, “AE-ATT-ConvLSTM”, that integrates an additional Convolutional Layer, Autoencoder, and an Attention Mechanism Feature Fusion Layer into the architecture of a Convolutional Long Short-Term Memory (ConvLSTM) sequential image prediction network. This approach aims to predict fracture propagations in different fracturing stages of horizontal wells. The proposed model is trained on a sample dataset consisting of 5000 instances of hydraulic fracturing in horizontal wells. The dataset includes crucial data such as fracture propagation images, wellbore perforation images, natural fracture images, pumping schedules, and reservoir properties. The model achieves remarkable results, with an MSE less than 15 × 10−5, a maximum SSIM of 0.93, and an average FMAE less than 48. The research findings demonstrate that this method significantly enhances prediction efficiency and provides new insights for predicting fracture morphology and optimizing fracturing design.

Flow field prediction in bed configurations: A parametric spatio-temporal convolutional autoencoder approach

The use of deep learning methods for modeling fluid flow has drawn a lot of attention in the past few years. Here we present a data-driven reduced-order model (ROM) for predicting flow fields in a bed configuration of hot particles. The ROM consists of a parametric spatio-temporal convolutional autoencoder. The neural network architecture comprises two main components. The first part resolves the spatial and temporal dependencies present in the input sequence, while the second part of the architecture is responsible for predicting the solution at the subsequent timestep based on the information gathered from the preceding part. We also propose the utilization of a post-processing non-trainable output layer following the decoding path to incorporate the physical knowledge, e.g. no-slip condition, into the prediction. The ROM is evaluated by comparing its predicted solution with the high-fidelity counterpart. In addition, proper orthogonal decomposition (POD) is employed to systematically analyze and compare the dominant structures present in both sets of solutions. The assessment of the ROM for a bed configuration with variable particle temperature showed accurate results at a fraction of the computational cost required by traditional numerical simulation methods.

Integrating mixed reality preparation into acute coronary syndrome simulation for nursing students: a single-group pretest-posttest study

BackgroundTimely and effective intervention within the ‘golden hour’—the critical first 90 min after the symptom onset—is crucial for initiating life-saving treatment and reducing mortality in acute coronary syndrome (ACS). This highlights the need for nursing students to be proficient in ACS care, emphasizing the importance of preparatory training. This study enhanced traditional simulation methods by integrating a mixed reality (MR) preparation step, offering a more immersive learning experience. We aimed to evaluate the effectiveness of integrating MR preparation into ACS simulation education, focusing on enhancements in knowledge, self-confidence in learning, and self-efficacy in learning. Additionally, we examined performance, practice immersion, and satisfaction to comprehensively evaluate the MR application.MethodsOne-group pretest-posttest design was implemented in a convenience sample of thirty-nine senior nursing students from a university in South Korea in August 2022. We developed a simulation program integrating MR preparation into ACS simulation (IMRP-ACSS), which was validated through expert review for content validity. The students participated in the simulation program over six hours across two days, including a 40-minute individual session of MR-based simulation preparation using head-mounted displays (the HoloLens 2). Individual changes in knowledge, self-confidence in learning, and self-efficacy in learning evaluated by the survey were analyzed using paired t-tests. Additionally, group performance assessed using the checklist was analyzed. Immersion and satisfaction were measured with a tool and a 10-point Likert scale, respectively.ResultsIndividually, participants demonstrated significantly increased knowledge (t = 11.87, p < .001), self-confidence in learning (t = 7.17, p < .001), and self-efficacy in learning (t = 4.70, p < .001) post-education. Group performance yielded a mean score of 56.43/70 ± 7.45. Groups scored higher in electrocardiogram interpretation, patient safety, and heparin administration. Participants reported a practice immersion level of 37.82/50 ± 9.13 and expressed satisfaction with the program, achieving an average score of 8.85/10 ± 1.35.ConclusionIntegrating MR preparation into ACS simulation enhanced nursing students’ knowledge, self-confidence in learning, and self-efficacy in ACS care, providing a replicable and immersive learning experience. This method is an effective addition to nursing education, preparing students through comprehensive, technology-enhanced training.

Efficient Error-Rate Estimation for Optical Transmission Systems Using Artificial Neural Networks

ABSTRACT With the advancement in modulation formats, optical transmission systems are becoming more adaptable and dynamic. Predicting bit errors of higher-order modulation formats is a challenging problem for the optimum design of communication systems. The existing simulation methods have persisted by employing iterative procedures that frequently incur high computing expenses and time consumption. On the other hand, deep learning (DL) algorithms have demonstrated remarkable efficacy as practical computing tools, presenting a viable approach to accelerate modulation simulations. In this paper, an artificial neural network (ANN) based bit error rate (BER) estimation scheme is proposed for the popular modulation forms including 56 Gbps 16-quadrature amplitude modulation (16QAM), 100 Gbps 32QAM, and 120 Gbps 64QAM optical system. Based on constellation diagrams (CDs) acquired with different launch power, laser linewidth, transmission distance, and OSNR, amplitude histograms (AHs) are generated via a manual preprocessing method. The properly trained ANN architecture exhibits a computational speed that surpasses traditional simulation methods by a factor exceeding 347. Moreover, the design considerations including the number of layers, nodes, activation functions, learning rate, optimizers, and evolution epochs are also investigated in detail. This research paves the way for optical transmission systems to use fast DL-based optimization strategies.

High-fidelity simulations of boundary layer transition on an underwater axisymmetric body with stability theory

During the cruise status of an underwater high-speed vehicle, the flow around the vehicle's head typically transitions from a laminar to a turbulent state, triggering flow noise that can interfere with the normal operation of sonar. In order to accurately investigate this flow noise through numerical simulation, a high-fidelity turbulent flow field solution is essential. Common traditional turbulence numerical simulation methods, such as unsteady Reynolds-averaged simulation and large eddy simulation (LES), struggle to capture high-frequency turbulent fluctuations accurately due to their inability to directly resolve small-scale eddy structures, which results in compromising the simulation accuracy of high-frequency flow noise. To address this issue, this paper employs direct numerical simulation (DNS) to achieve high-fidelity resolution of the turbulent flow field, thereby enabling a more accurate assessment of flow noise distribution on the vehicle's surface. Meanwhile, considering significant computational resources required to solve the entire flow field in an underwater high Reynolds number environment, this study also incorporates the fixed transition modeling method and stability theory to confine the DNS computational domain to the vicinity of the transitional zone to improve simulation efficiency. Comparative analysis of flow noise monitoring results in the laminar, transitional, and turbulent zones revealed that the flow noise source in the laminar zone exhibits the lowest amplitude across all frequencies, while the flow noise source in the transitional zone features the highest amplitude, approximately 10 dB higher than that in the turbulent zone. Moreover, significant amplitudes in high-frequency components (above 30 kHz) are detected in both the transition and turbulent zones. Additionally, this study employs LES with the Smagorinsky model to simulate the flow field within the same computational domain as DNS, demonstrating the limitations of the Smagorinsky model-based LES in capturing high-frequency flow noise.

Intelligent reconstruction of unsteady combustion flow field of scramjet based on physical information constraints

To alleviate the problem of high-fidelity data dependence and inexplicability in pure data-driven neural network models, physical informed neural networks (PINNs) provide a new learning paradigm. This study constructs an efficient, accurate, and robust PINN framework for predicting unsteady combustion flow fields based on Navier–Stokes (NS) equation constraints. To achieve fast prediction of a multi-physical field in a scramjet combustion chamber, we propose a U-shaped residual neural network model based on feature information fusion. The model uses a residual neural network module as the backbone, uses jump connection to improve model generalization, and uses the U-shaped structure to fuse the receptive field features with different scales to enhance the feature expression ability of the model. To prevent improper assumptions from leading to wrong method constraints, we consider the flow characteristic mechanism of each physical field to constrain the neural network and verify its accuracy through numerical simulation of the unsteady flow field in the scramjet combustor with Mach number (Ma) 2.0. This method can accurately predict the multi-physical field of unsteady turbulent combustion based on the time, space, Ma and turbulent eddy viscosity coefficients of a small number of samples. Specially, the proposed physical driven and data driven fusion proxy model can predict the unsteady combustion flow field in milliseconds. It has important reference value to solve the problem of low calculation efficiency of a traditional numerical simulation method of a combustion process.

Multi-XGB: A multi-objective reliability evaluation approach for aeroengine turbine discs

Multiple failure sites often co-existed in aeroengine turbine discs, its reliability performance contains complex characteristics of high-nonlinearity and multiple-output, leading to the insufficient reliability evaluating accuracy/efficiency of traditional direct simulation or surrogate separate methods. To address this problem, a multi-XGB model is presented by fusing the linkage sampling technique to extract a multi-input multi-output dataset, the XGBoost series ensemble platform to build the modelling architecture, and the Bayesian optimization algorithm to find the optimal model hyperparameters. Moreover, a multi-XGB driven multi-objective reliability evaluation framework is presented as well. Regarding a typical aeroengine high-pressure turbine disc with multiple failure sites (i.e., disc core, disc rim) as an example, the probabilistic distributions, correlation relationships, and reliability degree for each/all frail sites in turbine disc are acquired by applying the presented multi-XGB model. Through the method comparisons, the presented method is verified to hold high computing accuracy and efficiency in evaluating the multi-objective reliability of turbine discs. The current efforts shed light on the theoretical development for reliability evaluation, design, and optimization of high-end equipment like aeroengine.

Evaluating solubility, stability, and inclusion complexation of oxyresveratrol with various β-cyclodextrin derivatives using advanced computational techniques and experimental validation

Oxyresveratrol (OXY), a natural stilbenoid in mulberry fruits, is known for its diverse pharmacological properties. However, its clinical use is hindered by low water solubility and limited bioavailability. In the present study, the inclusion complexes of OXY with β-cyclodextrin (βCD) and its three analogs, dimethyl-β-cyclodextrin (DMβCD), hydroxypropyl-β-cyclodextrin (HPβCD) and sulfobutylether-β-cyclodextrin (SBEβCD), were investigated using in silico and in vitro studies. Molecular docking revealed two binding orientations of OXY, namely, 4′,6′-dihydroxyphenyl (A-form) and 5,7-benzenediol ring (B-form). Molecular Dynamics simulations suggested the formation of inclusion complexes with βCDs through two distinct orientations, with OXY/SBEβCD exhibiting maximum atom contacts and the lowest solvent-exposed area in the hydrophobic cavity. These results corresponded well with the highest binding affinity observed in OXY/SBEβCD when assessed using the MM/GBSA method. Beyond traditional simulation methods, Ligand-binding Parallel Cascade Selection Molecular Dynamics method was employed to investigate how the drug enters and accommodates within the hydrophobic cavity. The in silico results aligned with stability constants: SBEβCD (2060 M−1), HPβCD (1860 M−1), DMβCD (1700 M−1), and βCD (1420 M−1). All complexes exhibited a 1:1 binding mode (AL type), with SBEβCD enhancing OXY solubility (25-fold). SEM micrographs, DSC thermograms, FT-IR and 1H NMR spectra confirm the inclusion complex formation, revealing novel surface morphologies, distinctive thermal behaviors, and new peaks. Notably, the inhibitory impact on the proliferation of breast cancer cell lines, MCF-7, exhibited by inclusion complexes particularly OXY/DMβCD, OXY/HPβCD, and OXY/SBEβCD were markedly superior compared to that of OXY alone.

Electromagnetic transient simulation of power system based on PSCAD calling ADPSS

Abstract As the power system gradually expands into a massive scale, time-varying complex system, the requirements for simulation analysis software are constantly improving, and electromagnetic transient analysis plays a crucial role in the power system. Therefore, to solve the possible limitations of this traditional simulation method, this paper proposes an electromagnetic transient simulation of power systems using PSCAD, called ADPSS. Firstly, according to the basic overview and advantages of the electromagnetic transient analysis program PSCAD/EMTDC and the mathematical model software package ADPSS, the interconnection principle and implementation method between PSCAD/EMTDC and ADPSS are studied, and a simple power system power supply model is established in PSCAD. Then, the principle of PSCAD-ADPSS interface technology is studied. The electromagnetic transient simulation is carried out in the interface environment, and the interface custom components are designed. Finally, by conducting a comparative analysis of the simulation outcomes, it is found that the simulation waveforms obtained in the two simulation environments are almost the same, which shows the effectiveness of PSCAD-ADPSS interface simulation technology in electromagnetic transient analysis and has the potential to enhance the precision and dependability of electromagnetic transient simulation.

General Aqueous System Simulation through an AI-Embedded Metaverse Chemistry Laboratory.

Recent decades have witnessed the rapid development of autonomous laboratories and artificial intelligence, where experiments can be automatically run and optimized. Although human work is reduced, the total time of experimental optimization is still consuming due to limitations of the current ab metaverse framework, which accurately predicts the future state of the system by receiving and analyzing in situ experimental data. To substitute for traditional simulation methods, we designed a physically endorsed deep learning model to predict the future system picture ranging from atomic image to bulk appearance, intensively using the correlations between properties of the system. Through this framework, we studied the general aqueous system, covering 100+ common ionic solutions. We can accurately simulate properties for a general aqueous system as well as predict the time of solvation of ionic compounds ahead of real experiments. In this way, the experiments can be optimized more efficiently without waiting for the end of a bad iteration. We hope our work offers a fresh direction for the digitization of chemical information, enhancing access to and use of experimental data in advancing the field of physical chemistry.

End-to-end dimensionality reduction and regression from 3D geological uncertainties to estimate oil reservoir simulations

Managing geological uncertainties in reservoir engineering involves significant challenges mainly due to the prohibitive computational costs of traditional simulation methods. These simulations, essential for generating geological models, often require extensive computational resources and can take days or weeks to complete. The computational load limits the number of viable models while the high dimensionality of properties compounds the challenge, and the abundance of producing wells, each associated with different objective functions, further complicates the problem. Current methodologies have focused on training an Artificial Neural Network (ANN) for each producer well to create a proxy model. Instead, we propose the development of a single ANN designed to simultaneously predict the behavior of multiple objective functions. This work proposes a novel end-to-end deep neural network that can handle 3D geological uncertainties and replicate the simulator’s response for several cumulative production curves. This approach leverages 3D convolutions to process spatial and depth dimensions within heterogeneous reservoirs, including diverse geostatistical realizations. The end-to-end solution was successfully implemented in a Brazilian offshore field and accurately replicated the simulator’s behavior and yielded results that outperformed state-of-the-art methods. The results indicate correlations that exceeded 0.95 between the proxy response and the numerical simulator. Additionally, our approach demonstrated robustness even when trained with a limited number of simulation models, and it notably reduced computational costs. The findings highlight our architecture’s capacity to integrate dimensional reduction and regression analysis within a unified framework, effectively predicting different fluid behaviors in the reservoir and showcasing robustness against high dimensionality and sparse data.

Design optimization of groundwater circulation well based on numerical simulation and machine learning

The optimal design of groundwater circulation wells (GCWs) is challenging. The key to purifying groundwater using this technique is its proficiency and productivity. However, traditional numerical simulation methods are limited by long modeling times, random optimization schemes, and optimization results that are not comprehensive. To address these issues, this study introduced an innovative approach for the optimal design of a GCW using machine learning methods. The FloPy package was used to create and implement the MODFLOW and MODPATH models. Subsequently, the formulated models were employed to calculate the characteristic indicators of the effectiveness of the GCW operation, including the radius of influence (R) and the ratio of particle recovery (Pr). A detailed collection of 3000 datasets, including measures of operational efficiency and key elements in machine learning, was meticulously compiled into documents through model execution. The optimization models were trained and evaluated using multiple linear regression (MLR), artificial neural networks (ANN), and support vector machines (SVM). The models produced by the three approaches exhibited notable correlations between anticipated outcomes and datasets. For the optimal design of circulating well parameters, machine learning methods not only improve the optimization speed, but also expand the scope of parameter optimization. Consequently, these models were applied to optimize the configuration of the GCW at a site in Xi’an. The optimal scheme for R (Q = 293.17 m3/d, a = 6.09 m, L = 7.28 m) and optimal scheme for Pr (Q = 300 m3/d, a = 3.64 m, L = 1 m) were obtained. The combination of numerical simulations and machine learning is an effective tool for optimizing and predicting the GCW remediation effect.

An industrial IoT-based deformation resistance prediction and thickness control method of cold-rolled strip in steel production systems

Different from traditional analytical and numerical simulation methods, data-driven methods can more effectively describe the effects of nonlinear and coupling factors. Although the fluctuations of deformation resistance and thickness for hot-rolled strips significantly influence the thickness accuracy in strip cold rolling, the traditional deformation resistance prediction model and thickness control mode don't involve that. Thus, based on the Industrial Internet of Things (IIOT) platform and data-driven technology, this work proposed a new method for deformation resistance prediction and a new mode of feed-forward control for strip thickness. IIOT platform can collect production data from the various levels to provide a complete data set. The deformation resistance prediction model is established by adopting the differential evolutionary algorithm to optimize the bi-directional long and short-term memory network. In addition, the feed-forward control (DFFC) strategy for strip thickness is proposed based on deformation resistance prediction. An application results of a 1420 mm production line of CR indicated that the established prediction model could provide a high accuracy prediction for deformation resistance, and compared to traditional thickness gauges-based automatic gauge control methods, the DDFC can capture the effect of the fluctuations of deformation resistance and thickness for the hot-rolled strip to improve thickness accuracy effectively.

Application of the fast 3D simplified simulation method for the large CAP1400 nuclear island evaporator based on the coupled source term method

As for the evaporator of the large CAP1400 nuclear power plant, it has a large volume with numerous tubes along with the complex flow heat transfer phenomenon. The traditional direct simulation method usually takes a lot of time and computing resources to obtain the flow heat transfer data. To solve the problem, a fast simulation method established on the porous medium source term was herein proposed. In this simplified simulation method, the complex evaporator model is replaced with the porous medium model, where the derived heat and momentum source terms are added into the governing equation to control the heat flows. Five kinds of simplified evaporator models with different numbers of pipes were set to be the research objects, the internal flows inside the evaporator were studied with the traditional direct simulation method and the newly proposed simplified simulation method. Via comparative analysis, it can be found that: 1) The newly proposed simplified simulation method accurately captures the distribution characteristics of pressure, temperature, and entropy production in the evaporator, serving as a viable alternative to traditional direct simulation methods. Specifically, the average simulation errors for inlet-outlet pressure difference and temperature difference are both less than 5 % and 7 %, respectively. 2) The newly proposed simplified simulation method can capture the mutual interference effects between the evaporator and the nuclear main pump. When the evaporator is connected in series with the main nuclear pump, the average head of the main pump increases by 0.34 m, while the efficiency decreases by an average of 0.34 %. The average simulation errors of the simplified simulation method are 0.06 m and 0.13 %, respectively. 3) The application of the simplified simulation method leads to a significant 85 % decrease in the average computational cost for evaporator calculations and a notable 36 % decrease for combined calculations of the evaporator and main pump. The study here is expected to provide technique support for the simulation and evaluation of the large nuclear island.

Simulation of thermal-fluid coupling in silicon single crystal growth based on gradient normalized physics-informed neural network

The thermal-fluid coupling phenomenon of silicon melt is significant in the growth process of silicon single crystals. Complex convection affects the temperature and concentration distribution of the silicon melt. Therefore, establishing and solving the thermal-fluid coupling model of silicon melts is crucial to optimizing the crystal growth process and improving crystal quality. Traditional numerical simulation methods have limitations in regard to optimization, control, and real-time monitoring. Physics-Informed Neural Network (PINN) does not require model discretization, after training, it can make predictions quickly, showing potential for industrial applications. However, when solving practical industrial coupling models, PINN often struggles to converge due to large parameter values and significant gaps between solution variables. Moreover, solving the thermal-fluid coupling model with PINN can be treated as a multitask problem, where the gradients of different equations interfere with each other, leading to gradient confusion, slow convergence, or even divergence. Therefore, this paper proposes a gradient normalized PINN (GNPINN) for solving the thermal-fluid coupling model of silicon melt. GNPINN balances the contribution of each task, ensuring a more equitable training speed between different tasks to stabilize the training process of the coupling model. This paper considers the thermal-fluid coupling model of silicon melt under different rotation conditions. GNPINN can accurately and comprehensively capture the complex temperature, velocity, and pressure distribution of silicon melt compared with other methods. Additionally, the experimental results uncover the flow and heat transfer properties of silicon melt, validating the effectiveness and industrial applicability of GNPINN.