Digital Twin-based Parameter Optimization Method and System for Samarium Iron Nitrogen Injection Molding Process

Abstract

This invention relates to the field of samarium iron nitride (SFeNi) permanent magnet material manufacturing technology, and discloses a method and system for optimizing SFeNi injection molding process parameters using digital twins. This method achieves dynamic optimization by integrating a digital twin with the physical production line. Real-time sensor data from the production line is collected and preprocessed. Based on this, a digital twin containing a particle layer, a melt layer, and a magnetic pole layer is constructed and updated. Using this twin, the agglomeration and orientation of SFeNi particles are predicted in the particle layer, temperature, shear, and magnetic flux distribution are simulated in the melt layer, and magnetic performance indicators are evaluated in the magnetic pole layer. Based on these predictions and evaluations, the material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are optimized using an artificial intelligence proxy model. Finally, the optimized parameters are sent to the production line for execution via a programmable logic controller (PLC). This method achieves online adaptive adjustment of process parameters, improving the quality and efficiency of magnet molding.

Description

Technical Field

[0001] This invention relates to the field of samarium iron nitrogen permanent magnet material manufacturing technology, specifically to a method and system for optimizing samarium iron nitrogen injection molding process parameters using digital twins. Background Technology

[0002] Samarium iron nitride (SFIN) permanent magnets are widely used in precision motors, sensors, and other fields due to their excellent magnetic properties and cost advantages. Injection molding is a key technology for fabricating complex-shaped SFIN magnets, and its process parameters, such as material temperature, injection speed, and magnetic field waveform, directly affect the final density, orientation, and magnetic properties of the magnet. Current technologies for optimizing the injection molding process mainly rely on two types of methods. One type is computer-aided engineering simulation based on a single physics field, such as simulating only melt flow or temperature field distribution. This type of method treats SFIN as a homogeneous medium, ignoring the crucial role of its particle characteristics in the molding process. The other type is empirical models or statistical process control based on production data, which guide parameter setting by analyzing the correlation between historical process parameters and final product quality.

[0003] Single-physics simulations cannot accurately describe the multi-scale, multi-physics evolution of samarium iron nitrogen particles from solid to molten state and then to solidification. The particle aggregation behavior within the material, its orientation behavior under flow and magnetic field influences, and the interaction between these microstructural evolutions and macroscopic temperature, shear, and magnetic fields are severely simplified and ignored. This leads to significant deviations between simulation results and actual production, making it difficult to effectively predict high-precision magnetic properties. Empirical models based on historical data heavily rely on large amounts of high-quality finished product testing data, resulting in long model development cycles and a lack of responsiveness to dynamic changes during production. When there are minor fluctuations in raw material batches or equipment conditions, the original model may become invalid, requiring the re-accumulation of data, demonstrating poor adaptability.

[0004] Existing methods lack a high-fidelity model that can span multiple scales—particles, melts, and magnetic poles—and map the physical production line state in real time. Conventional process optimization either focuses only on macroscopic melt flow or only on post-process quality statistics, failing to dynamically capture and predict changes in the material's microstructure during manufacturing. Therefore, there is an urgent need for a method that can reflect the intrinsic relationship between the microstructure evolution of samarium iron nitride (SMR) materials during molding and macroscopic process parameters and final magnetic properties, thereby achieving precise and dynamic optimization of process parameters. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for optimizing process parameters of samarium iron nitride injection molding using digital twins, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for optimizing process parameters of samarium iron nitride injection molding using digital twins, the method comprising:

[0007] The method achieves dynamic optimization of process parameters through the integration of a digital twin with an injection molding production line, and includes the following stages:

[0008] Collect and preprocess real-time sensor data from the injection molding production line;

[0009] A digital twin is constructed and updated based on preprocessed real-time sensor data. The digital twin includes a particle layer, a melt layer, and a magnetic pole layer.

[0010] The digital twin is used to predict the agglomeration and orientation of samarium iron nitrogen particles in the particle layer, simulate the temperature distribution, shear distribution and magnetic flux distribution in the melt layer, and evaluate the magnetic performance indicators in the magnetic pole layer.

[0011] Based on the aforementioned aggregation degree, orientation degree, temperature distribution, shear distribution, magnetic flux distribution, and magnetic performance indicators, the material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are optimized using an artificial intelligence proxy model.

[0012] The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are sent to the injection molding production line for execution via a programmable logic controller.

[0013] Preferably, the step of collecting and preprocessing real-time sensor data from the injection molding production line includes:

[0014] Real-time sensor data from the injection molding production line is cleaned and standardized in format to obtain purified sensor data;

[0015] Feature extraction is performed on the purified sensor data to obtain particle feature data, melt feature data, and magnetic pole feature data.

[0016] Preferably, the construction and updating of the digital twin based on preprocessed real-time sensor data includes:

[0017] Based on the particle feature data, a samarium iron nitrogen anisotropic constitutive model and an oxide coating resistance model are coupled in the particle layer digital twin model to establish a particle layer digital twin.

[0018] Based on the melt characteristic data, a computational fluid dynamics model coupling rheology and magnetic field is embedded in the digital twin model of the melt layer to establish a digital twin of the melt layer;

[0019] Based on the magnetic pole feature data, an artificial intelligence agent model is trained in the magnetic pole layer digital twin model to establish a magnetic pole layer digital twin.

[0020] Preferably, the method of using the digital twin to predict the aggregation degree and orientation degree of samarium iron nitrogen particles in the particle layer includes:

[0021] The particle layer digital twin is used to predict the aggregation behavior and orientation state of samarium iron nitrogen particles to obtain the predicted aggregation degree and orientation degree.

[0022] The digital twin of the melt layer is used to simulate the temperature field, shear field and magnetic flux field during the melt flow process in real time, so as to obtain temperature distribution data, shear distribution data and magnetic flux distribution data;

[0023] The magnetic performance parameters are calculated using the digital twin of the magnetic pole layer to obtain remanence data and maximum magnetic energy product data.

[0024] Preferably, the optimization of material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve using an artificial intelligence agent model includes:

[0025] The predicted aggregation degree, predicted orientation degree, temperature distribution data, shear distribution data, magnetic flux distribution data, remanence data, and maximum energy product data are input into the artificial intelligence proxy model;

[0026] The AI ​​agent model optimizes the strategy using aggregation degree, orientation degree, remanence, and maximum magnetic energy product as reward functions to generate optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve.

[0027] Preferably, the use of the artificial intelligence agent model includes:

[0028] The input data is feature-encoded to obtain the encoded feature vector;

[0029] The encoded feature vectors are processed by a policy network to obtain the process parameter adjustment strategy;

[0030] Action sampling was performed on the process parameter adjustment strategy to obtain optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve.

[0031] Preferably, the step of sending the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve to the injection molding production line via a programmable logic controller includes:

[0032] The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are received by the programmable logic controller.

[0033] The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are converted into control commands.

[0034] Control commands are sent to the actuators in the injection molding production line to adjust the barrel temperature, injection screw speed, magnetic field generator, and holding pressure.

[0035] Preferably, the feature encoding of the input data includes:

[0036] Spatial features of temperature distribution data, shear distribution data, and magnetic flux distribution data are extracted using a convolutional neural network to generate feature maps.

[0037] By processing time series of clustering and orientation predictions using long short-term memory networks, dynamic change patterns can be captured.

[0038] The feature map and time series features are concatenated into an encoded feature vector, and an attention mechanism is applied to weight the important features.

[0039] Preferably, the step of converting the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve into control commands includes:

[0040] The optimized material temperature is mapped to the heater power setpoint, and the temperature value is converted into an analog voltage signal by a lookup table method;

[0041] The injection speed is converted into a servo motor speed command, and a drive signal is generated using pulse width modulation technology.

[0042] The pulsed magnetic field waveform parameters are parsed into frequency and amplitude control words for the magnetic field generator and written into the register;

[0043] The pressure holding curve is discretized into a time-pressure sequence, and a smooth pressure control curve is generated by an interpolation algorithm.

[0044] Preferably, the present invention also includes a digital twin samarium iron nitride injection molding process parameter optimization system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the above-described digital twin samarium iron nitride injection molding process parameter optimization method.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] By constructing a multi-level digital twin comprising a particle layer, a melt layer, and a magnetic pole layer, a multi-scale, high-fidelity mapping of the injection molding process was achieved. The digital twin predicts the agglomeration and orientation of samarium iron nitrogen (SMI) particles at the particle layer, which directly relates to the initial microscopic homogeneity and anisotropic potential of the material. Simultaneous simulation of temperature, shear, and magnetic flux distributions at the melt layer reveals the influence of macroscopic process conditions on the material's rheological properties and particle arrangement under magnetic field influence. The magnetic pole layer evaluates the magnetic properties of the final product. This comprehensive modeling approach, from microscopic particle behavior to macroscopic field distribution and final functional performance, overcomes the limitations of traditional single-physics simulations that only homogenize materials. Its effect lies in providing a deeper and more comprehensive digital representation of the intrinsic mechanism of the molding process, enabling process parameter adjustments based on predictions of changes in the material's microstructure, greatly improving the scientific rigor and accuracy of process optimization.

[0047] By leveraging this multi-level digital twin-driven AI proxy model for parameter optimization, multi-objective collaborative dynamic optimization of process parameters was achieved. The AI ​​proxy model takes material temperature, injection speed, pulsed magnetic field waveform, and holding pressure curve as inputs, and learns and iterates based on the aggregation degree, orientation degree, various field distributions, and magnetic performance indicators predicted by the digital twin. This method combines a high-fidelity physical mechanism model with an efficient data-driven optimization algorithm, enabling rapid evaluation of the potential effects of massive parameter combinations in virtual space and selection of the optimal solution. Its effectiveness lies in changing the traditional model that relies on manual trial and error or offline static optimization. It can dynamically adjust parameters according to the real-time status of the production line, significantly shortening the parameter optimization cycle and adapting to fluctuations in raw materials or the environment, continuously improving the consistency and stability of the product's magnetic properties. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the working principle of the digital twin-based method for optimizing process parameters in samarium iron nitrogen injection molding as described in this invention.

[0049] Figure 2 A flowchart for building and updating digital twins;

[0050] Figure 3 A flowchart for the use of digital twins;

[0051] Figure 4 A dynamic comparison diagram of the orientation degree of samarium iron nitrogen particles before and after optimization during the molding process;

[0052] Figure 5 This is a histogram showing the particle orientation distribution during the samarium iron nitrogen injection molding process. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides a digital twin-based method for optimizing process parameters in samarium iron nitrogen (SMR) injection molding. The method includes collecting and preprocessing real-time sensor data from the injection molding production line; constructing and updating a digital twin based on the preprocessed real-time sensor data; the digital twin comprising a particle layer, a melt layer, and a magnetic pole layer; predicting the agglomeration and orientation of SMR particles in the particle layer using the digital twin; simulating temperature, shear, and magnetic flux distribution in the melt layer; evaluating magnetic performance indicators in the magnetic pole layer; and optimizing material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve using an artificial intelligence proxy model based on the agglomeration, orientation, temperature, shear, magnetic flux distribution, and magnetic performance indicators. The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are then sent to the injection molding production line for execution via a programmable logic controller (PLC).

[0055] Example 1: See Figure 2 The process involves collecting and preprocessing real-time sensor data from the injection molding production line. This includes data cleaning and format standardization to obtain purified sensor data, followed by feature extraction to obtain particle feature data, melt feature data, and magnetic pole feature data. Based on the preprocessed real-time sensor data, a digital twin is constructed and updated. This includes coupling a samarium-iron-nitrogen anisotropic constitutive model and an oxide coating resistance model into a particle layer digital twin model based on particle feature data; embedding a computational fluid dynamics model coupling rheology and magnetic fields into a melt layer digital twin model based on melt feature data; and training an artificial intelligence agent model in a magnetic pole layer digital twin model based on magnetic pole feature data to establish a magnetic pole layer digital twin.

[0056] In practice, real-time sensor data from the injection molding production line is collected and preprocessed. This data originates from thermocouples, pressure sensors, magnetometers, and high-speed cameras installed on the barrel, mold, and magnetic field generator. Data cleaning addresses impulse noise and steady-state deviation in sensor readings using a combination of median filtering and moving average filtering. Format standardization converts current and voltage signals with different sampling frequencies and dimensions into standardized physical quantity values. Feature extraction calculates time-domain and frequency-domain statistics from the purified sensor data, such as texture features of particle images, fluctuation variance of melt pressure, and harmonic components of magnetic field signals. This separates particle feature data characterizing particle state, melt feature data characterizing melt flow state, and magnetic pole feature data characterizing magnetic field action.

[0057] The construction and updating of the digital twin based on preprocessed real-time sensor data relies on particle feature data. In specific implementations, a samarium-iron-nitrogen anisotropic constitutive model and an oxide-coated resistance model are coupled. The samarium-iron-nitrogen anisotropic constitutive model describes the mechanical response of particles under the influence of magnetic and flow fields, while the oxide-coated resistance model quantifies the influence of the oxide layer on the particle surface on the interfacial impedance. This coupling process can be understood as achieving the simultaneous solution of the particle force balance equations and the circuit network equations. One expression used to calculate the interparticle magnetization force can be written as:

[0058] ;

[0059] in: This represents the magnetizing force vector between particle p and particle q. It is the vacuum permeability. and These are the magnetic moment vectors of particles p and q, respectively. It is the displacement vector connecting the centers of the two particles. This refers to its modulus. The construction of the digital twin model of the melt layer is based on melt feature data. In specific implementation, a computational fluid dynamics model coupling rheology and magnetic field is embedded. This model introduces a magnetic volume force term into the Navier-Stokes equations and considers the Joule heating effect in the energy equation. The construction of the digital twin model of the magnetic pole layer is based on magnetic pole feature data. In specific implementation, an artificial intelligence agent model is trained. This artificial intelligence agent model uses the outputs of the aforementioned particle layer and melt layer as input features and uses measured magnetic performance indicators as supervision signals for supervised learning.

[0060] Optionally, the update mechanism for the granular layer digital twin involves comparing the particle distribution predicted by the granular layer digital twin model with actual images acquired by a high-speed camera, and using an error backpropagation algorithm to adjust some empirical parameters in the samarium-iron-nitrogen anisotropic constitutive model. Optionally, the update of the melt layer digital twin involves assimilating measured temperature and pressure data at key points within the mold cavity and using an ensemble Kalman filter algorithm to correct the initial field and boundary conditions of the computational fluid dynamics simulation in real time. It is understandable that the update cycle for the magnetic pole layer digital twin is relatively long; typically, new samples are used to incrementally train the artificial intelligence proxy model only after completing one injection molding cycle and measuring the actual magnetic properties of the magnetic poles.

[0061] Example 2: See Figure 3 The method utilizes digital twins to predict the agglomeration and orientation of samarium iron nitrogen (SMM) particles in the particle layer. This includes using a particle layer digital twin to predict the agglomeration behavior and orientation state of SMM particles to obtain predicted agglomeration and orientation values; using a melt layer digital twin to perform real-time simulation of the temperature field, shear field, and magnetic flux field during melt flow to obtain temperature distribution data, shear distribution data, and magnetic flux distribution data; and using a magnetic pole layer digital twin to calculate magnetic performance indicators to obtain remanence data and maximum energy product data.

[0062] In practice, a particle layer digital twin is used to predict the aggregation behavior and orientation state of samarium iron nitrogen particles to obtain predicted aggregation and orientation values. The particle layer digital twin receives particle distribution feature data from an image sensor as input. Specifically, the prediction of aggregation behavior is achieved by analyzing the collision probability of particles under the coupling of flow and magnetic fields using the discrete element method and the adhesion energy calculated based on van der Waals forces and magnetic dipole forces. The prediction of orientation state is achieved by solving the rotational dynamics equations of each particle (e.g., considering the balance of fluid torque and magnetic torque) to track the angle of deviation of its principal axis relative to the direction of the external magnetic field. The orientation prediction value can be understood as a statistic used to quantify the consistency of the orientation of the entire particle group. Its calculation involves fitting a statistical distribution of the deviation angles of all particles. An expression for calculating the orientation degree can be written as:

[0063] ;

[0064] in: This represents the predicted value of orientation. It represents the total number of samarium iron nitrogen particles within the simulation domain. It is the angle between the easy magnetization axis of the i-th particle and the direction of the applied external magnetic field. The predicted aggregation degree is obtained by calculating the size distribution and spatial density of the particle clusters. For example, after the simulation domain is meshed, the number of particles in each grid is counted. Grids exceeding a set threshold are considered as aggregation regions. The aggregation degree can be defined as the ratio of the total volume of the aggregation region to the volume of the simulation domain.

[0065] A digital twin of the melt layer is used to simulate the temperature, shear, and magnetic flux fields during melt flow in real time to obtain temperature, shear, and magnetic flux distribution data. The digital twin receives temperature, pressure, and flow rate data from sensors inside the barrel and mold as boundary conditions. In some embodiments, the temperature field simulation is based on unsteady-state heat conduction equations and considers heat source terms generated by polymer melt viscous dissipation and Joule heating effects generated by magnetic field-induced eddy currents. In some embodiments, the shear field simulation is performed by calculating the velocity gradient tensor in the flow field and obtaining its second invariant, while the magnetic flux field simulation is achieved by solving Maxwell's equations and considering the influence of samarium iron nitrogen particle magnetization on the local magnetic field. Specifically, the coupling between the flow field and the magnetic field is achieved by adding a magnetic volume force term (J×B) to the momentum conservation equation, where J is the induced current density and B is the magnetic flux density. The boundary conditions required for the simulation, such as the velocity curve at the injection port and the temperature distribution on the mold wall, are provided by real-time sensor data or interpolation results based on sensor data.

[0066] Optionally, a magnetic pole layer digital twin is used to calculate magnetic performance indicators to obtain remanence data and maximum energy product data. The magnetic pole layer digital twin receives the final orientation and agglomeration prediction values ​​from the particle layer digital twin, and the post-solidification magnetic flux distribution data from the melt layer digital twin as input. Optionally, the remanence data is calculated based on the Stoner-Wohlfarth model to simulate the hysteresis behavior of individual grains, considering the influence of intergranular magnetic static interactions and agglomeration structure on the demagnetization process. Specifically, the particle orientation distribution obtained from the particle layer is used to set the easy axis orientation distribution of the Stoner-Wohlfarth model set, and the agglomeration information is used to adjust the effective demagnetization field coefficient. It can be understood that the maximum energy product data is calculated by numerically integrating the simulated demagnetization curve to find the maximum value of the product of magnetic flux density B and magnetic field strength H. This calculation process considers the geometry of the magnetic poles and the inhomogeneity of the material. The magnetic flux distribution data provided by the melt layer after solidification is used to assess the possible magnetic flux non-uniformity regions inside the magnetic poles, which may affect the overall shape of the demagnetization curve.

[0067] Example 3: Optimizing material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve using an artificial intelligence agent model. This involves inputting predicted values ​​of agglomeration degree, orientation degree, temperature distribution data, shear distribution data, magnetic flux distribution data, remanence data, and maximum magnetic energy product data into the artificial intelligence agent model. The artificial intelligence agent model then uses agglomeration degree, orientation degree, remanence, and maximum magnetic energy product as reward functions to optimize the strategy and generate optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve.

[0068] In practice, predicted values ​​of aggregation, orientation, temperature, shear distribution, magnetic flux distribution, remanence, and maximum energy product are input into the AI ​​proxy model, serving as its state space. Before input, the data undergoes normalization to map the different dimensions and orders of magnitude of these predicted values ​​to a unified numerical range. The AI ​​proxy model employs a deep reinforcement learning framework, with the number of neurons in its input layer matching the dimension of the state vector.

[0069] An AI-powered proxy model is used to optimize the process parameters—agglomeration degree, orientation degree, remanence, and maximum energy product—as reward functions to generate optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve. The optimization process aims to find the combination of process parameters that maximizes the reward function. In some embodiments, the reward function design treats agglomeration degree as a negative indicator to be minimized, while orientation degree, remanence, and maximum energy product are treated as positive indicators to be maximized. An exemplary construction is as follows:

[0070] ;

[0071] in: Indicates the instant reward value. This represents the normalized remanence data. This represents the normalized maximum magnetic energy product data. This represents the normalized predicted orientation degree. This represents the normalized predicted clustering value. , , , These are weighting coefficients used to adjust the relative importance of various indicators. In some embodiments, policy optimization employs a proximal policy optimization algorithm, which updates the policy network parameters through multiple iterations, enabling the AI ​​agent model to explore a more promising process parameter space.

[0072] Optionally, the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve serve as the action output of the AI ​​agent model. Material temperature is a scalar value, injection speed is a discrete sequence of points on a curve that varies over time, the pulsating magnetic field waveform is defined by parameters such as frequency, amplitude, and duty cycle, and the holding pressure curve represents the relationship between holding pressure and time. Optionally, the strategy optimization process is performed after each injection molding cycle. The AI ​​agent model calculates a reward value based on the difference between the magnetic performance indicators predicted by the current digital twin and the desired target, and updates its decision-making strategy accordingly. It can be understood that generating the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve is a continuous learning and adjustment process. The AI ​​agent model continuously improves the process parameter settings through interaction with the environment (i.e., the digital twin) until the reward function converges or the preset process target is reached.

[0073] Example 4: The use of an AI proxy model includes feature encoding of the input data to obtain an encoded feature vector, processing the encoded feature vector using a policy network to obtain a process parameter adjustment strategy, and sampling the process parameter adjustment strategy to obtain optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve. Feature encoding of the input data includes using a convolutional neural network to extract spatial features of temperature distribution data, shear distribution data, and magnetic flux distribution data and generating a feature map. A long short-term memory network is used to process the time series of aggregation degree prediction values ​​and orientation degree prediction values ​​to capture dynamic change patterns. The feature map and time series features are concatenated into an encoded feature vector, and an attention mechanism is applied to weight important features.

[0074] In practical implementation, the use of the AI ​​proxy model involves feature encoding of the input data to obtain encoded feature vectors. The input data includes predicted aggregation, predicted orientation, temperature distribution, shear distribution, magnetic flux distribution, remanence, and maximum energy product. This data is preprocessed before input to unify timestamps and spatial resolution. In the implementation, the feature encoding process converts multimodal heterogeneous data into a unified numerical representation. The encoded feature vectors serve as input to the policy network, which employs a deep neural network structure. Its hidden layers use the ReLU activation function for nonlinear transformation, and the output layer uses different activation functions depending on the type of process parameter; for example, the sigmoid function is used for normalized material temperature values, and the tanh function is used for the rate of change of injection speed.

[0075] In some embodiments, the encoded feature vector is processed by a policy network to obtain a process parameter adjustment strategy. The policy network outputs a probability distribution or deterministic value of the process parameter adjustment amount. The process parameter adjustment strategy includes the increment of material temperature, the correction coefficient of the injection speed curve, the offset of the pulsed magnetic field waveform parameters, and the adjustment vector of the control point of the holding pressure curve. In some embodiments, the policy network is trained using a policy gradient method, which updates the network weights by maximizing the expected cumulative reward. The reward signal is calculated based on the difference between the magnetic performance index predicted by the digital twin and the target value.

[0076] Optionally, action sampling is performed on the process parameter adjustment strategy to obtain optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve. The action sampling process involves random sampling based on the probability distribution output by the strategy network. For example, multinomial sampling is used for discrete pulsating magnetic field waveform selection, while Gaussian sampling is used for continuous holding pressure adjustment. Optionally, a convolutional neural network is used in feature encoding to extract spatial features from temperature distribution data, shear distribution data, and magnetic flux distribution data. The input to the convolutional neural network is two-dimensional grid data, with each grid point storing physical field values. The convolutional layers use 3x3 convolutional kernels for feature extraction, and the pooling layers use max pooling to reduce feature dimensionality. The generated feature map is a three-dimensional tensor. It can be understood that a long short-term memory network processes the time series of aggregation and orientation prediction values. The input sequence length of the long short-term memory network is fixed, corresponding to a complete injection molding cycle. The number of hidden layer units is set to 64, and the output is the hidden state of the last time step as the time series feature.

[0077] In practice, the feature map and time-series features are concatenated into an encoded feature vector. Before the concatenation operation, the feature map tensor is flattened into a one-dimensional vector, and the time-series features are directly concatenated. An attention mechanism is applied to weight important features, calculating the weight coefficient for each feature dimension. The weighted sum is then used to obtain the final encoded feature vector. The attention weights are calculated based on the dot product of a trainable query vector and the feature vector, and then normalized using the softmax function. A formula for calculating the attention weights can be written as:

[0078] ;

[0079] in: This represents the attention weight of the nth feature dimension. It is a trainable query vector. It is the feature value of the nth dimension in the encoded feature vector. It is the total dimension of the encoded feature vector. This represents the dot product operation. See Table 1, which shows some of the input data and their corresponding processing output dimensions during the feature encoding process.

[0080] Table 1: Input Data and Output Dimensions of Feature Encoding

[0081]

[0082] See Figure 4 The results show the dynamic changes in particle orientation before optimization (purple-red curve) and after optimization (cyan-blue curve) within the molding time step. Specifically, the optimized orientation degree remained at a relatively high level above 0.85, with a significantly narrowed fluctuation range; while the average orientation degree before optimization was about 0.7, and showed a significant decrease (as low as 0.6) in some time steps (such as around step 60). This difference stems from the process parameter optimization driven by the digital twin: after optimization, parameters such as material temperature and injection speed were dynamically adjusted through an artificial intelligence proxy model, suppressing the disturbance of melt flow on particle orientation and enabling the particles to maintain a more consistent directional arrangement throughout the molding process. In terms of parameter correlation, the stable high value of orientation degree corresponds to the synergistic control effect of the particle layer constitutive model and the melt layer flow-magnetic field coupled simulation in the digital twin, which is a direct manifestation of the process parameter optimization method at the microstructure control level.

[0083] Example 5: The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are sent to the injection molding production line via a programmable logic controller (PLC). This includes receiving the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve through the PLC; converting these parameters into control commands; and sending these control commands to the actuators of the injection molding production line to adjust the barrel temperature, injection screw speed, magnetic field generator, and holding pressure. Converting the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve into control commands includes mapping the optimized material temperature to a heater power setpoint and converting the temperature value into an analog voltage signal using a lookup table; converting the injection speed into a servo motor speed command and generating a drive signal using pulse width modulation (PWM); parsing the pulsating magnetic field waveform parameters into frequency and amplitude control words for the magnetic field generator and writing them into a register; and discretizing the holding pressure curve into a time-pressure sequence and generating a smooth pressure control curve using an interpolation algorithm.

[0084] In practice, the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are sent to the injection molding production line for execution via a programmable logic controller (PLC). The PLC communicates with the upper-level optimization system via the industrial Ethernet protocol, receiving the optimal set of process parameters calculated by the artificial intelligence agent model. In practice, the data packets received by the PLC contain structured parameter information, such as the material temperature setpoint, the time-speed table of the injection speed curve, the frequency-amplitude-phase parameter set of the pulsating magnetic field waveform, and multiple pressure-time setpoints of the holding pressure curve.

[0085] The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are converted into control commands. This conversion process relies on pre-set mapping relationships and algorithms within the programmable logic controller (PLC). In some embodiments, the optimized material temperature is mapped to a heater power setpoint. The mapping process is based on the material's heat capacity and heat loss model, using a lookup table to convert the temperature value into an analog voltage signal. The temperature-voltage correspondence used in the lookup table is stored in the PLC's non-volatile memory. In some embodiments, the injection speed is converted into a servo motor speed command. This conversion process needs to consider the mechanical transmission ratio of the injection screw and the system inertia, employing pulse width modulation (PWM) technology to generate the drive signal. The duty cycle of the PWM is linearly related to the target speed.

[0086] Optionally, the pulsed magnetic field waveform parameters are analyzed into frequency and amplitude control words for the magnetic field generator. This analysis involves a digital signal synthesis algorithm, and the generated control words are written to the register of the magnetic field generator control unit via a parallel data bus. Optionally, the holding pressure curve is discretized into a time-pressure sequence. The number of discrete points is determined based on the control precision of the injection molding machine. A smooth pressure control curve is generated using a cubic spline interpolation algorithm to ensure a smooth pressure transition during the holding pressure stage. It can be understood that the generation of control commands needs to follow a strict time sequence to ensure precise timing coordination between barrel heating, injection, magnetic field application, and the holding pressure process.

[0087] Control commands are sent to the actuators in the injection molding line to adjust the barrel temperature, injection screw speed, magnetic field generator, and holding pressure. Command transmission is accomplished through the digital and analog output modules of the programmable logic controller (PLC). In practice, the barrel temperature is controlled using a proportional-integral-derivative (PID) control algorithm, with an analog voltage signal driving a solid-state relay to adjust the power input of the heating coil. The injection screw speed control command is sent to the servo driver, which precisely controls the rotation of the servo motor based on the speed command. After the control word for the magnetic field generator is written, its internal direct digital frequency synthesizer chip generates a corresponding pulsed current waveform. The holding pressure is controlled by adjusting the opening of the proportional pressure valve using an analog voltage signal; closed-loop pressure control ensures consistency between the actual pressure and the set curve. The timing synchronization of the control commands is coordinated by a unified time base manager, and its synchronization error Δt is calculated using the following formula:

[0088] ;

[0089] in: This indicates the deviation between the actual time of the action and the expected time. It is the timestamp of the actual action fed back by the actuator. It is the expected action timestamp set by the programmable logic controller (PLC) program. The time base manager continuously monitors it. The system response delay is compensated by adjusting the lead time of command transmission.

[0090] See Figure 5 This figure presents the statistical distribution characteristics of the orientation degree of samarium iron nitrogen (SFeNi) particles predicted by the particle layer digital twin. Specifically, the horizontal axis represents the orientation degree (ranging from 0.60 to 0.90), and the vertical axis represents the frequency of occurrence of each orientation degree interval. The histogram visually shows the distribution density of different orientation degrees: the 0.65 orientation degree interval occurs 3 times, representing the peak interval in the figure, while the intervals of 0.70 and 0.75 occur once, and the intervals of 0.60 and 0.80 occur twice. This distribution result is the quantitative output of the particle orientation state prediction after the particle layer digital twin is coupled with the SFeNi anisotropic constitutive model and the oxide coating resistance model. This data will serve as one of the input features of the artificial intelligence proxy model, participating in the optimization decisions of process parameters such as material temperature and injection speed.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing process parameters of samarium iron nitride injection molding using digital twins, characterized in that, The method achieves dynamic optimization of process parameters through the integration of a digital twin with an injection molding production line, and includes the following stages: Collect and preprocess real-time sensor data from the injection molding production line; A digital twin is constructed and updated based on preprocessed real-time sensor data. The digital twin includes a particle layer, a melt layer, and a magnetic pole layer. The digital twin is used to predict the agglomeration and orientation of samarium iron nitrogen particles in the particle layer, simulate the temperature distribution, shear distribution and magnetic flux distribution in the melt layer, and evaluate the magnetic performance indicators in the magnetic pole layer. Based on the aforementioned aggregation degree, orientation degree, temperature distribution, shear distribution, magnetic flux distribution, and magnetic performance indicators, the material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are optimized using an artificial intelligence proxy model. The optimized material temperature, injection speed, pulsed magnetic field waveform, and holding pressure curve are sent to the injection molding production line for execution via a programmable logic controller. The method of using the digital twin to predict the aggregation degree and orientation degree of samarium iron nitrogen particles in the particle layer includes: The particle layer digital twin is used to predict the aggregation behavior and orientation state of samarium iron nitrogen particles to obtain the predicted aggregation degree and orientation degree. The digital twin of the melt layer is used to simulate the temperature field, shear field and magnetic flux field during the melt flow process in real time, so as to obtain temperature distribution data, shear distribution data and magnetic flux distribution data; The magnetic performance indicators are calculated using the magnetic pole layer digital twin to obtain remanence data and maximum energy product data. The magnetic pole layer digital twin receives the final orientation degree prediction value and agglomeration degree prediction value from the particle layer digital twin, as well as the solidified magnetic flux distribution data from the melt layer digital twin as input. The calculation of remanence data is based on the Stoner-Wohlfarth model to simulate the hysteresis behavior of individual grains, and considers the influence of intergranular magnetic-static interactions and agglomeration structure on the demagnetization process. Among them, the particle orientation distribution obtained from the particle layer is used to set the easy axis direction distribution of the Stoner-Wohlfarth model set, and the agglomeration degree information is used to adjust the effective demagnetization field coefficient.

2. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 1, characterized in that, The process of collecting and preprocessing real-time sensor data from the injection molding production line includes: Real-time sensor data from the injection molding production line is cleaned and standardized in format to obtain purified sensor data; Feature extraction is performed on the purified sensor data to obtain particle feature data, melt feature data, and magnetic pole feature data.

3. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 2, characterized in that, The construction and updating of the digital twin based on preprocessed real-time sensor data includes: Based on the particle feature data, a samarium iron nitrogen anisotropic constitutive model and an oxide coating resistance model are coupled in the particle layer digital twin model to establish a particle layer digital twin. Based on the melt characteristic data, a computational fluid dynamics model coupling rheology and magnetic field is embedded in the digital twin model of the melt layer to establish a digital twin of the melt layer; Based on the magnetic pole feature data, an artificial intelligence agent model is trained in the magnetic pole layer digital twin model to establish a magnetic pole layer digital twin.

4. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 3, characterized in that, The optimization of material temperature, injection speed, pulsed magnetic field waveform, and holding pressure curve using an artificial intelligence agent model includes: The predicted aggregation degree, predicted orientation degree, temperature distribution data, shear distribution data, magnetic flux distribution data, remanence data, and maximum energy product data are input into the artificial intelligence proxy model; The AI ​​agent model optimizes the strategy using aggregation degree, orientation degree, remanence, and maximum magnetic energy product as reward functions to generate optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve.

5. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 4, characterized in that, The use of the AI ​​agent model includes: The input data is feature-encoded to obtain the encoded feature vector; The encoded feature vectors are processed by a policy network to obtain the process parameter adjustment strategy; Action sampling was performed on the process parameter adjustment strategy to obtain optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve.

6. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 5, characterized in that, The process of sending the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve to the injection molding production line via a programmable logic controller includes: The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are received by the programmable logic controller. The optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve are converted into control commands. Control commands are sent to the actuators in the injection molding production line to adjust the barrel temperature, injection screw speed, magnetic field generator, and holding pressure.

7. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 6, characterized in that, The feature encoding of the input data includes: Spatial features of temperature distribution data, shear distribution data, and magnetic flux distribution data are extracted using a convolutional neural network to generate feature maps. By processing time series of clustering and orientation predictions using long short-term memory networks, dynamic change patterns can be captured. The feature map and time series features are concatenated into an encoded feature vector, and an attention mechanism is applied to weight the important features.

8. The method for optimizing samarium iron nitride injection molding process parameters using digital twins according to claim 7, characterized in that, The process of converting the optimized material temperature, injection speed, pulsating magnetic field waveform, and holding pressure curve into control commands includes: The optimized material temperature is mapped to the heater power setpoint, and the temperature value is converted into an analog voltage signal by a lookup table method; The injection speed is converted into a servo motor speed command, and a drive signal is generated using pulse width modulation technology. The pulsed magnetic field waveform parameters are parsed into frequency and amplitude control words for the magnetic field generator and written into the register; The pressure holding curve is discretized into a time-pressure sequence, and a smooth pressure control curve is generated by an interpolation algorithm.

9. A digital twin-based system for optimizing samarium iron nitride (SMR) injection molding process parameters, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for optimizing process parameters of samarium iron nitride injection molding based on digital twins as described in any one of claims 1 to 8.

Images