Multi-scale electronic component service life prediction method based on physical deep learning

Abstract

The invention discloses a multi-scale electronic component service life prediction method based on physical deep learning, and relates to the field of intelligent prediction of electronic component service life. The objective of the invention is to solve the problem that a machine learning model adopted during life prediction is limited in multi-scale multi-source data cooperative processing capability and is difficult to accurately characterize the performance change of the whole life cycle of a device. Constructing a threshold voltage model with the simultaneous influence of the interface state and the oxide trap charge based on the dominating factor; considering the threshold voltage model with the simultaneous influence of the interface state and the oxide trapped charge, and constructing a variable quantity model of the on-resistance; establishing a comprehensive index model including drain current, gate leakage current, transconductance, capacitance, switching time and power consumption and thermal resistance of the tested electronic component; and taking the output results of the threshold voltage model, the on-resistance variable quantity model and the comprehensive index model as the input of a deep learning model, and predicting through the deep learning model to obtain the service life of the tested electronic component.

Description

Technical Field

[0001] The invention belongs to the field of intelligent prediction of the life of electronic components. Background Art

[0002] In recent years, with the rapid development and deployment of electronic components towards high-density integration, high performance and multi-function, the reliability and remaining useful life (RUL) prediction problems faced by devices in multiple complex environments such as humidity, temperature and radiation have become increasingly prominent. Existing technologies often rely mainly on empirical models or big data-driven machine learning models when predicting life, ignoring the coupling of multi-scale mechanisms, lacking the system connection from device mechanisms to macroscopic scales, and only analyzing device failures at a single limited level, making it difficult to accurately capture the true degradation process under multiple stress environments; secondly, although machine learning models can mine complex patterns in massive data, they lack effective physical failure mechanism constraints, and the models are prone to fall into "black boxes", and there are large errors in the model extrapolation ability under complex environments. In practical applications, the failure process of electronic components is often affected by the coupling of multiple stresses and multiple measurement signals. Existing deep learning or traditional models have limited collaborative processing capabilities for multi-scale and multi-source data, making it difficult to fully utilize data information, lack multi-scale quantification of key parameters, and make it difficult to accurately characterize the performance changes of devices throughout their life cycle. Summary of the invention

[0003] The present invention aims to solve the problems that the machine learning model used in the prior art for life prediction has limited collaborative processing capabilities for multi-scale and multi-source data, is difficult to fully utilize data information, lacks multi-scale quantification of key parameters, and is difficult to accurately characterize the performance changes of the device throughout its life cycle. A multi-scale electronic component life prediction method based on physical deep learning is now provided.

[0004] Multi-scale electronic component life prediction method based on physical deep learning, including:

[0005] During the operation of the electronic components under test, a dominant factor that combines reaction temperature, humidity and radiation dose and can affect the reliability degradation mechanism of the electronic components under test is introduced. Based on this dominant factor, a threshold voltage model with the simultaneous influence of interface states and oxide trap charges is constructed.

[0006] Considering the threshold voltage model with the simultaneous influence of interface states and oxide trap charges, a model of the change in on-resistance is constructed;

[0007] Establish a comprehensive indicator model including drain current, gate leakage current, transconductance, capacitance, switching time, power consumption and thermal resistance of the electronic components under test;

[0008] Taking the output results of the threshold voltage model, the change amount model of the on-resistance, and the comprehensive index model as the input of the deep learning model, the lifetime of the electronic component under test is predicted through the deep learning model.

[0009] Further, the governing factor expression that comprehensively reflects temperature, humidity, and radiation dose and can affect the reliability degradation mechanism of the electronic component under test is as follows:

[0010]

[0011] In the formula, T is temperature, H is humidity, D is radiation dose, T0, H0, and D0 respectively represent the nominal values of T, H, and D, and α, β, and γ all represent empirical exponential parameters;

[0012] Define the threshold Δ th1 = 1 and Δ th2 = 1.5.

[0013] When Δ(T, H, D) < Δ th1 , the electronic component under test is affected by interface state trap charges;

[0014] When Δ th1 < Δ(T, H, D) < Δ th2 , the electronic component under test is affected by both interface state trap charges and oxide trap charges;

[0015] When Δ(T, H, D) > Δ th2 , the electronic component under test is affected by oxide trap charges.

[0016] Further, the above-mentioned threshold voltage model based on the governing factor and affected by both interface states and oxide trap charges includes:

[0017] Construct a threshold voltage model affected by interface state trap charges;

[0018] Construct a threshold voltage model affected by oxide trap charges;

[0019] Combining the threshold voltage model affected by interface state trap charges with the threshold voltage model affected by oxide trap charges to obtain a threshold voltage model affected by both interface states and oxide trap charges.

[0020] Further, the above-mentioned construction of the threshold voltage model affected by interface state trap charges includes:

[0021] When the electronic component under test is affected by interface state trap charges, calculate the interface state trap charge ΔC IS :

[0022] ΔCIS = qD IS ,

[0023] where q is the elementary charge, and D IS is the total number of interface state trap charge captures;

[0024] The oxide capacitance R CA represents the voltage change ΔV IS caused by the interface state trap charge ΔC IS :

[0025]

[0026] The oxide capacitance ε CA is the dielectric constant of the oxide, C is the length of the oxide, K is the width of the oxide, and T CA is the thickness of the gate oxide layer of the electronic component under test;

[0027] Establish a relational expression that correlates the total number of interface state trap charge captures D IS with T, H, and D:

[0028]

[0029] where A IS and B IS both represent fitting parameters of the interface state trap charge under complex environmental factors;

[0030] Substitute the oxide capacitance R CA , the relational expression that correlates the total number of interface state trap charge captures D IS with T, H, and D, and the expression of the interface state trap charge ΔC IS into the expression of the voltage change ΔV IS to obtain the threshold voltage model affected by the interface state trap charge:

[0031]

[0032] where ΔV TH1 is the threshold voltage affected by the interface state trap charge;

[0033] The construction of the threshold voltage model affected by the oxide trap charge includes:

[0034] Under the influence of the oxide trap charge on the electronic component under test, calculate the oxide trap charge ΔC OS :

[0035] ΔC OS = qD OS ,

[0036] Where D OS is the total number of oxide trap charge captures;

[0037] The voltage change ΔV CA caused by the oxide trap charge ΔC OS is represented by the oxide capacitance R OS :

[0038]

[0039] Establish a relational expression for the total number of oxide trap charge captures D OS associated with T, H, and D:

[0040]

[0041] Where A OS and B OS both represent fitting parameters of the oxide trap charge under complex environmental factors;

[0042] Substitute the oxide capacitance R CA , the relational expression for the total number of oxide trap charge captures D OS associated with T, H, and D, and the expression of the oxide trap charge ΔC OS into the expression of the voltage change ΔV OS to obtain a threshold voltage model affected by the oxide trap charge:

[0043]

[0044] where ΔV TH2 is the threshold voltage affected by the oxide trap charge;

[0045] The expression of the threshold voltage model affected by the simultaneous presence of interface states and oxide trap charges is:

[0046]

[0047] where ΔV TH3 is the threshold voltage affected by the simultaneous presence of interface states and oxide trap charges.

[0048] Furthermore, the expression of the change amount model of the on-resistance is:

[0049]

[0050] where ΔR DS(ON) represents the change amount of the on-resistance of the measured electronic component, and K G and K L respectively represent the widths of the channel and the structural accumulation region of the measured electronic component, and C G and CL represents the length of the channel and the structure accumulation region of the electronic component under test, Q G and Q L represent the electron mobility of the channel and the structure accumulation region of the electronic component under test, R CA is the oxide capacitance, ΔV TH3 is the threshold voltage affected by the simultaneous presence of interface states and oxide trap charges, V GS represents the gate voltage of the electronic component under test, R C represents the correction factor, V TH represents the initial threshold voltage of the electronic component under test.

[0051] Furthermore, the above-mentioned establishment of a comprehensive index model including drain current, gate leakage current, transconductance, capacitance, switching time, as well as power consumption and thermal resistance of the electronic component under test, includes:

[0052] Establish mathematical parameter expressions {x1, x2, x3, x4, x5, x6, x7} for the drain current, gate leakage current, transconductance, capacitance, switching time, and power consumption and thermal resistance parameters of the device;

[0053] Normalize each type of mathematical parameter respectively and establish a comprehensive index model H(t):

[0054]

[0055] where i = 1,…,7, x′ i (t) represents the result after normalization of the i-th type of mathematical parameter, w i (t) represents the parameter x i 's adaptive weight.

[0056] Furthermore, the above-mentioned normalization of each type of mathematical parameter respectively includes:

[0057] Perform inverse mapping normalization on the transconductance and capacitance,

[0058] Perform forward mapping normalization on the drain current, gate leakage current, switching time, and power consumption and thermal resistance of the device.

[0059] Furthermore, the above-mentioned deep learning model includes an input layer, an LSTM layer, a linear layer, a multi-head attention layer, a dropout layer, a flattening layer, a multi-layer perceptron, and an output layer;

[0060] The update expression of the LSTM layer is:

[0061]

[0062] where, h tThe hidden state at time t, σ represents the sigmoid function, x t represents the input at a single time step t, tanh represents the hyperbolic tangent, W i 、W f 、W o and W c all represent the weight matrices of the trainable LSTM layer, α i 、α f and α o all represent the learnable scalars of the LSTM layer, β represents the trainable coefficient of the LSTM layer, γ ∈ [0, 1] represents the mixing ratio of the old and new memories in the output hidden state;

[0063] The expression of the multi - head self - attention layer is as follows:

[0064]

[0065] Q = xW Q + α Q q0

[0066] K = xW K + α K k0,

[0067] V = xW V + α V v0

[0068] where, Attention(Q, K, V) represents the output of the multi - head self - attention layer, x represents the input of the multi - head self - attention layer, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, W Q 、W K and W V all represent the weight matrices of the multi - head self - attention layer, α Q 、α K and α V all represent the trainable scalars of the multi - head self - attention layer, q0, k0, and v0 are all learnable factors, τ is the scaling factor, d head represents the dimension of a single attention head of the attention layer.

[0069] Furthermore, the loss function of the above - mentioned deep learning model is expressed as:

[0070]

[0071] In the formula, represents the change amount model loss error of the on - resistance, represents the threshold voltage model loss error, represents the comprehensive index model loss error, Denote the loss error of the deep learning model, w R , w V , w H and w D represent the loss weights corresponding to the loss error, and there is w R +w V +w H +w D = 1.

[0072] Furthermore, the change amount model loss error of the on-resistance The expression is:

[0073]

[0074] The threshold voltage model loss error The expression is:

[0075]

[0076] The comprehensive index model loss error The expression is:

[0077]

[0078] The loss error of the deep learning model The expression is:

[0079]

[0080] In the formula, n = 1, 2,..., N k , N k represents the total number of samples in a batch, β R , β V , β H and β D all represent scaling factors, γ D , γ R , γ V and γ H all represent translation factors, cosh represents the hyperbolic cosine function, x kn represents the nth input feature of the kth device, ΔR pred represents the on-resistance of the device predicted by the model, ΔR phy represents the on-resistance obtained from the physical model, ΔV pred represents the threshold voltage of the device predicted by the model, ΔV phy represents the threshold voltage obtained from the physical model, ΔH pred represents the comprehensive index predicted by the model, ΔH phy represents the comprehensive index obtained from the physical model, Y predRepresents the remaining useful life value of the device predicted by the deep learning model, Y actual Represents the actual remaining useful life value of the device, and θ represents the training parameters of the deep learning model.

[0081] The multi-scale electronic component life prediction method based on physics-enhanced deep learning according to the present invention has the following beneficial effects:

[0082] Compared with the life prediction models based on experience and statistics, the present invention realizes collaborative analysis in a complex multi-stress (electrical stress, irradiation, temperature, and humidity) environment, breaks through the limitation of single-stress modeling of traditional methods, and reduces the life prediction error to within 5%;

[0083] Due to cross-scale modeling from the charge scale to the macroscopic device scale, capturing the multi-scale interaction effects ignored by machine learning models, ensuring the physical consistency of predictions. In the actual test dataset, MAPE can be controlled below 0.5%, the model training time is reduced by 40%, the real-time prediction efficiency is improved to the millisecond level, and the computing efficiency is increased by 30%.

[0084] Supports the prediction of the life of different types of electronic components, such as MOSFETs, IGBTs, and diodes, with an adaptability coverage rate of over 95%. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 Is a flowchart framework diagram of a multi-scale physics-guided machine learning model;

[0086] Figure 2 Is a comparison diagram of the life prediction of the MPIDL model. DETAILED DESCRIPTION OF THE INVENTION

[0087] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other.

[0088] The deficiencies of the prior art have led to the difficulty of the electronic component life prediction technology in meeting the reliability application requirements of complex environments in terms of comprehensive accuracy, model interpretability, and applicability. Therefore, in view of the above problems existing currently, the present invention proposes a multi-scale electronic component life prediction method based on physics-enhanced deep learning, establishing a multi-scale physical model based on the device failure mechanism and charge accumulation relationship, and then establishing a modified loss function to achieve the fusion of multi-modal physics and deep learning, overcoming the deficiencies of the prior art in the lack of micro-mechanism and multi-scale data processing and the low prediction accuracy and efficiency in the life prediction of electronic components. Specifically as follows:

[0089] Referring to Figure 1 and Figure 2 To specifically illustrate this embodiment, the multi-scale electronic component life prediction method based on physics-enhanced deep learning described in this embodiment includes:

[0090] I. Acquisition of device parameters

[0091] SiC (silicon carbide) MOSFETs are widely used in electric vehicles, satellites, power transmission, and high-frequency power conversion due to their high efficiency and high reliability. In the environment of satellite on-orbit operation, SiC MOSFETs face complex and special environments such as high-radiation environments, extreme temperature variations, and vacuum environments. In the long-term operation tasks of satellites, it is required that core components such as SiC MOSFETs maintain stable performance throughout the operation cycle.

[0092] Real-time monitoring and acquisition of the source-drain on-resistance, threshold voltage, drain current, gate leakage current, transconductance, capacitance parameters, switching time, device temperature, environmental humidity, radiation dose, and power consumption and thermal resistance of the SiC MOSFET device.

[0093] II. Physical modeling of multi-scale analysis

[0094] SiC MOSFET devices are affected by interface state trap charges and oxide trap charges simultaneously in multiple complex environments (temperature, humidity, and radiation). Therefore, a dominant factor Δ(T, H, D) is introduced to comprehensively reflect the influence of temperature T, humidity H, and radiation dose D on the device reliability degradation mechanism. Define the thresholds Δ th1 = 1 and Δ th2 = 1.5. If Δ(T, H, D) < Δ th1 , then the interface state trap charges dominate; if Δ th1 < Δ(T, H, D) < Δ th2 , then it is in the transition region, and both interface state trap charges and oxide trap charges dominate; if Δ(T, H, D) > Δ th2 , then the oxide trap charges dominate.

[0095] Thus, the dominant factors for judging the interface state trap charge and oxide trap charge under the multiple complex environments of the device are established:

[0096]

[0097] α + β + γ = 1 (2),

[0098] wherein, T represents the current working temperature (Kelvin), H represents the relative humidity (%), D represents the irradiation dose (rad), T0, H0, and D0 respectively represent the nominal values of T, H, and D, and α, β, and γ all represent empirical exponential parameters.

[0099] 2.1 Threshold voltage

[0100] 2.1.1 During the operation of the device, when Δ(T, H, D) < Δ th1 , electron-hole pairs will be generated in the silicon oxide layer. The holes caused by the environment are usually trapped inside the SiO2, and interface reactions occur at the SiO2 / SiC interface, generating a large number of interface state trap charges. For the interface state trap charge ΔC IS , it is solved by the following formula:

[0101] ΔC IS = qD IS (3),

[0102] wherein, q represents the elementary charge, and D IS represents the total number of interface state trap charge captures.

[0103] The voltage change amount ΔV IS caused by the interface state trap charge ΔC IS can be expressed by the oxide capacitance R CA as:

[0104]

[0105] The oxide capacitance R CA can be calculated and solved by the following formula:

[0106]

[0107] wherein, ε CA represents the dielectric constant of the oxide, C represents the length of the oxide, K represents the width of the oxide, and T CA is the thickness of the gate oxide layer of the measured electronic component.

[0108] The relational expression between the total number of interface state trap charge captures D IS and T, H, D is established:

[0109]

[0110] In the formula, A IS and B IS both represent the fitting parameters of the interface state trap charge under complex environmental factors.

[0111] Therefore, substituting into Equation (4) can further establish the formula for the threshold voltage ΔV TH1 affected by the interface state trap charge:

[0112]

[0113] 2.1.2 During the operation of the device, when Δ(T, H, D) > Δ th2 , as the device operates in the environment for a long time, the oxide trap charge simultaneously dominates and affects the performance of the device. For the oxide trap charge ΔC OS it can be calculated by the following formula:

[0114] ΔC OS = qD OS (8),

[0115] In the formula, D OS represents the total number of oxide trap charge captures.

[0116] Therefore, the voltage change amount ΔV OS caused by the oxide trap charge ΔC OS can be calculated as follows:

[0117]

[0118] Establish the relational formula for the total number of oxide trap charge captures D OS associated with T, H, D:

[0119]

[0120] In the formula, A OS and B OS both represent the fitting parameters of the oxide trap charge under complex environmental factors.

[0121] Therefore, substituting the voltage change amount ΔV OS into Equation (9) can further establish the formula for the threshold voltage ΔV TH2 affected by the oxide trap charge:

[0122]

[0123] 2.1.3 During the operation of the device, when there is a transition region Δ th1 < Δ(T, H, D) < Δ th2When in this interval, interface states and oxide trap charges both play a dominant role in affecting the device performance. Therefore, the threshold voltage ΔV under the simultaneous influence of interface states and oxide trap charges TH3 is established by the formula:

[0124]

[0125] 2.2 On-resistance

[0126] Multiple complex environments (temperature, humidity, and irradiation) will simultaneously affect the source-drain on-resistance R of the device DS(ON) , and a multi-scale physical model is established for R DS(ON) :

[0127]

[0128] In the formula, R G represents the resistance of the channel region of the electronic component under test, R L represents the resistance of the structural accumulation region of the electronic component under test, R C represents the correction factor, V GS represents the gate voltage of the electronic component under test, K G and K L respectively represent the widths of the channel and structural accumulation regions of the electronic component under test, C G and C L represent the lengths of the channel and structural accumulation regions of the electronic component under test, Q G and Q L represent the electron mobilities of the channel and structural accumulation regions of the electronic component under test, V TH represents the initial threshold voltage of the electronic component under test.

[0129] By solving the partial derivative, a multi-scale physical model formula for the change in on-resistance ΔR of the electronic component under test is established DS(ON) :

[0130]

[0131] Among them, ΔV TH is the threshold voltage under unknown influences.

[0132] Thus, for R G and R L , the corresponding calculations are:

[0133]

[0134] In summary, considering the threshold voltage ΔV in Equation (12) TH3 , the model formula for the change in on-resistance can be modified as:

[0135]

[0136] 2.3 Establishment of Comprehensive Indexes for Other Parameters

[0137] For the drain current I D 、gate leakage current I G 、transconductance g m 、capacitance parameter C, switching time t sw and the power consumption P loss of the device and the thermal resistance parameter R θ etc., a comprehensive index (health degree) is established through the adaptive weighted method.

[0138] First, establish the mapping between the above physical parameters and mathematical parameters:

[0139]

[0140] Subsequently, perform parameter normalization to eliminate the influence of parameter dimensions. For transconductance and capacitance, use inverse mapping normalization, and for the others, use forward mapping normalization:

[0141]

[0142] In the formula, and are the minimum and maximum values of the parameter x i respectively; for inverse mapping normalization, x′ i (t) = 1 indicates that the parameter reaches the optimal level, and 0 indicates the worst; for forward mapping normalization, x′ i (t) = 0 indicates that the parameter reaches the optimal level, and 1 indicates the worst.

[0143] Set the reference weight i of the parameter x to satisfy:

[0144]

[0145] To make the weight dynamically adjusted with device aging and environmental conditions, add an adaptive correction term Δw i (t) to each weight:

[0146] Δw i (t) = γ i [x′ i (t) - x′ i,target [ + δ i Φ i [E(t)] (22),

[0147] In the formula, γ i and δ i are scalar coefficients, x′ i,targetDenote the target normalized value, Φ i [E(t)] represents that the environmental stress (temperature, humidity, and irradiation) factor is equivalent to Δ(T, H, D).

[0148] Combining the comprehensive reference weight and the correction term, the final adaptive weight w i (t):

[0149]

[0150] At time t, establish the drain current I D and the gate leakage current I G and the transconductance g m and the capacitance parameter C, the switching time t sw as well as the power consumption P of the device loss and the thermal resistance parameter R θ The comprehensive index model H(t) of the parameters:

[0151]

[0152] III. Deep learning model based on physical mechanism

[0153] By constructing a loss function, introduce the physical model of multi-scale analysis into the established LSTM-Transfomer deep learning model to establish a deep learning model based on physical mechanism.

[0154] 3.1 LSTM-Transfomer deep learning model

[0155] The LSTM-Transfomer deep learning model combines time series modeling (LSTM) with the attention mechanism (Transformer MHA), and cooperates with the sliding window and MLP (multi-layer perceptron) for high-dimensional feature extraction, so as to realize efficient and reliable life prediction modeling.

[0156] The model mainly includes an input layer, an LSTM layer (which can return an entire sequence), a linear layer, a multi-head attention layer (the attention mechanism of Transformer), a dropout layer, a flattening layer, a multi-layer perceptron (MLP), and an output layer (linear activation for regression). The source-drain on-resistance, threshold voltage, drain current, gate leakage current, transconductance, capacitance parameter, switching time, device temperature, environmental humidity, irradiation dose, as well as the power consumption and thermal resistance of the device are used as the model input, and the model output is the predicted remaining life Y pred .

[0157] Extract additional local features from the scale physical parameter sequence through the sliding window technique to improve the prediction accuracy and practicability. Given the sliding window size is W and the step size is S, then the sliding window Window j is:[[]]

[0158]

[0159] Therefore, in the input layer, the feature matrix for the model inputs (source-drain on-resistance, threshold voltage, drain current, gate leakage current, transconductance, capacitance parameters, switching time, device temperature, environmental humidity, irradiation dose, and power consumption and thermal resistance of the device) is as follows:

[0160]

[0161] where N represents the number of sliding windows, d model represents the feature dimension of each time step, and the t-th row of X represents the feature vector x t .

[0162] Subsequently, the established enhanced LSTM layer is introduced to update the input x t at a single time step t, the hidden state h t-1 at the previous time step, and the memory cell c t-1 as follows:

[0163] h t = o t ⊙ tanh[γc t + (1 - γ)c t-1 (27).

[0164] where o t = σ(W o [h t-1 , x t + α o c t-1 ), i t = σ(W i [h t-1 , x t + α i c t-1 ), f t = σ(W f [h t-1 , x t + α f c t-1 ), σ represents the sigmoid function, tanh represents the hyperbolic tangent, W i , W f , W o , W c represent trainable weight matrices, α i , α f , α oIt represents a learnable scalar, which is used to enable the input gate, forget gate, and output gate to additionally refer to the size of the memory cell at the previous moment during decision-making. β represents a trainable coefficient for secondary reference to the memory cell. γ ∈ [0, 1] represents the mixing ratio of the old and new memories in the output hidden state.

[0165] Subsequently, for the multi-head self-attention layer of the Transformer, learnable factors q0, k0, v0 and a scaling factor τ are introduced to adapt to the requirements of multi-feature physical data and enhance the data feature capture ability, as follows:

[0166] Q = xW Q +α Q q0

[0167] K = xW K +α K k0 (28),

[0168] V = xW V +α V v0

[0169]

[0170] In the formula, Q, K, and V represent the query matrix, key matrix, and value matrix respectively; W Q 、W K and represent the weight matrix, α Q 、α K and α V represent trainable scalars, d head represents the dimension of a single attention head of the attention layer.

[0171] For the dropout layer, a dynamic dropout rate adjustment is established to make the model regularization adaptive:

[0172] p t =min(p base +γ′·e,1) (30),

[0173] In the formula, e is the number of training steps (epoch), p base represents the initial dropout rate, γ′ represents the increase rate, and p t represents the dynamic dropout rate.

[0174] 3.2 LSTM-Transfomer Deep Learning Model Incorporating Physical Mechanisms

[0175] By establishing a loss function introduce the physical model of multi-scale analysis into the LSTM-Transfomer deep learning model,

[0176]

[0177] w R + w V + w H + w D = 1 (32),

[0178] wherein, represents the loss error of the multi-scale physical model of the on-resistance, represents the loss error of the multi-scale physical model of the threshold voltage, represents the loss error of the comprehensive index, represents the loss error of deep learning, w R 、w V 、w H and w D represent the loss weights at the corresponding loss positions.

[0179] For the loss error of the multi-scale physical model of the on-resistance the loss error of the multi-scale physical model of the threshold voltage the loss error of the comprehensive index and the loss error of deep learning the model formula is established as follows:

[0180]

[0181]

[0182] wherein, n = 1, 2,..., N k , N k represents the total number of samples in a batch, β D 、β R 、β V and β H represent scaling factors for adjusting the error sensitivity of the loss function, γ D 、γ R 、γ V and γ H represent translation factors for controlling the asymmetry of the loss function, cosh represents the hyperbolic cosine function, x kn represents the nth input feature of the kth device, ΔR pred represents the on-resistance of the device predicted by the model, ΔR phy represents the on-resistance obtained from the physical model, ΔV pred represents the threshold voltage of the device predicted by the model, ΔV phy represents the threshold voltage obtained from the physical model, ΔH pred represents the comprehensive index predicted by the model, ΔH phy represents the comprehensive index obtained from the physical model, Ypred Denote the remaining useful life value of the device predicted by the deep learning model, Y actual Denote the actual remaining useful life value of the device, and θ denote the training parameters of the deep learning model.

[0183] In summary, this embodiment provides a multi-scale electronic component life prediction method based on physical deep learning. Considering the device under stress conditions such as electrical stress, humidity, temperature, and irradiation, an MPIDL (multi-scale physics-guided deep learning) model is established. The model takes the failure mechanism and charge accumulation relationship as constraints, and gradually establishes a multi-scale life prediction model for electronic components from the microscopic charge scale to the macroscopic device scale. In the part of modeling from the charge scale to the device scale, a dominant factor is introduced for judgment, and the influence of interface state trap charges and oxide trap charges is analyzed in three stages. An innovative multi-scale physical model of source-drain on-resistance, a multi-scale physical model of threshold voltage, and a health index of other multi-parameters are established. An innovative LSTM-Transformer model is established, and the fusion of multi-modal physics and deep learning is achieved by establishing a modified loss function, and finally, an efficient and accurate prediction of the life of electronic components is realized.

[0184] Although the present invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed, as long as they do not depart from the spirit and scope of the present invention as defined by the appended claims. It should be understood that different dependent claims and the features described herein can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. A method for predicting the lifespan of multi-scale electronic components based on physics-informed deep learning, characterized in that, Including: During the operation of the electronic component under test, a dominant factor that synthesizes reaction temperature, humidity, and radiation dose and can affect the reliability degradation mechanism of the electronic component under test is introduced. Based on this dominant factor, a threshold voltage model affected by interface states and oxide trap charges simultaneously is constructed; Considering the threshold voltage model affected by interface states and oxide trap charges simultaneously, a change amount model of the on-resistance is constructed; A comprehensive index model including drain current, gate leakage current, transconductance, capacitance, switching time, and power consumption and thermal resistance of the electronic component under test is established; Taking the output results of the threshold voltage model, the change amount model of the on-resistance, and the comprehensive index model as the input of the deep learning model, the lifetime of the electronic component under test is predicted through the deep learning model.

2. The method for predicting the lifespan of multi-scale electronic components based on physics deep learning according to claim 1, wherein The expression of the dominant factor that synthesizes reaction temperature, humidity, and radiation dose and can affect the reliability degradation mechanism of the electronic component under test is as follows: In the formula, T is the temperature, H is the humidity, D is the radiation dose, T0, H0, and D0 respectively represent the nominal values of T, H, and D, and α, β, and γ all represent empirical exponential parameters; Define the threshold Δ th1 = 1 and Δ th2 = 1.5, When Δ(T, H, D) < Δ th1 the electronic component under test is affected by the interface state trap charges; When Δ th1 <Δ(T, H, D)<Δ th2 the electronic component under test is affected by both interface state trap charges and oxide trap charges at the same time; When Δ(T, H, D) > Δ th2 the electronic component under test is affected by oxide trap charges.

3. The method for predicting the lifespan of multi-scale electronic components based on physics-based deep learning according to claim 2, wherein The constructing the threshold voltage model affected by interface states and oxide trap charges simultaneously based on the dominant factor includes: Constructing a threshold voltage model affected by interface state trap charges; Constructing a threshold voltage model affected by oxide trap charges; Combining the threshold voltage model affected by interface state trap charges with the threshold voltage model affected by oxide trap charges to obtain a threshold voltage model affected by interface states and oxide trap charges simultaneously.

4. The method for predicting the lifespan of multi-scale electronic components based on physics-based deep learning according to claim 3, wherein, The constructing the threshold voltage model affected by interface state trap charges includes: Calculate the interface state trap charge ΔC when the electronic component under test is affected by the interface state trap charge IS : ΔC IS = qD IS , where q is the elementary charge, D IS is the total number of interface state trap charge captures; Through oxide capacitor R CA Indicates the voltage change ΔV IS caused by the interface state trap charge ΔC IS : The oxide capacitor ε CA is the dielectric constant of the oxide, C is the length of the oxide, K is the width of the oxide, and T CA is the thickness of the gate oxide layer of the electronic component under test; Establish the total number D of interface state trap charge capture IS The relational expression associated with T, H, and D: Where A IS and B IS both represent the fitting parameters of the interface state trap charge under complex environmental factors; The oxide capacitor R CA , the total number of interface state trap charges captured D IS , the relational expressions associated with T, H, D, and the interface state trap charge ΔC IS are all substituted into the expression of the voltage change amount ΔV IS to obtain a threshold voltage model under the influence of interface state trap charges: where ΔV TH1 is the threshold voltage affected by the interface state trap charge; The constructing the threshold voltage model affected by oxide trap charges includes: Calculate the oxide trap charge ΔC when the electronic component under test is affected by oxide trap charges OS : ΔC OS = qD OS , where D OS is the total number of oxide trap charge captures; Via the oxide capacitance R CA Indicates the voltage change amount ΔV OS Caused by the oxide trap charge ΔC OS : Establish the total number D of trapped oxide charges OS The relational expressions associated with T, H, and D: Where, A OS and B OS both represent the fitting parameters of oxide trap charges under complex environmental factors; The oxide capacitor R CA , the total number of oxide trap charges trapped D OS , the relational expressions related to T, H, D, and the oxide trap charge ΔC OS are all substituted into the expression of the voltage change amount ΔV OS to obtain a threshold voltage model affected by oxide trap charges: where ΔV TH2 is the threshold voltage affected by oxide trap charges; The expression of the threshold voltage model affected by interface states and oxide trap charges simultaneously is: Among them, ΔV TH3 is the threshold voltage affected by the simultaneous influence of interface states and oxide trap charges.

5. The method for predicting the lifespan of multi-scale electronic components based on physical deep learning according to claim 1 or 4, characterized in that, The expression of the change amount model of the on-resistance is: Among them, ΔR DS(ON) represents the change in on-resistance of the electronic component under test, K G and K L respectively represent the widths of the channel and the structure accumulation region of the electronic component under test, C G and C L represent the lengths of the channel and the structure accumulation region of the electronic component under test, Q G and Q L represent the electron mobilities of the channel and the structure accumulation region of the electronic component under test, R CA is the oxide capacitance, ΔV TH3 is the threshold voltage affected by the simultaneous presence of interface states and oxide trap charges, V GS represents the gate voltage of the electronic component under test, R C represents the correction factor, V TH represents the initial threshold voltage of the electronic component under test.

6. The method for predicting the lifespan of multi-scale electronic components based on physical deep learning according to claim 1, wherein, The establishing a comprehensive index model including drain current, gate leakage current, transconductance, capacitance, switching time, and power consumption and thermal resistance of the electronic component under test includes: Establishing mathematical parameter expressions {x1, x2, x3, x4, x5, x6, x7} for the drain current, gate leakage current, transconductance, capacitance, switching time, and power consumption and thermal resistance parameters of the device; Normalizing each type of mathematical parameter respectively, and establishing a comprehensive index model H(t): where \(i = 1,\ldots,7\), \(x i '(t)\) represents the result after normalization of the \(i\)-th type of mathematical parameter, \(w i (t)\) represents the adaptive weight of the parameter \(x i .

7. The method for predicting the life of multi-scale electronic components based on physical deep learning according to claim 6, wherein The respectively normalizing each type of mathematical parameter includes: Performing inverse mapping normalization processing on the transconductance and capacitance, Performing forward mapping normalization processing on the drain current, gate leakage current, switching time, and power consumption and thermal resistance of the device.

8. The method for predicting the lifespan of multi-scale electronic components based on physical deep learning according to claim 1, wherein The deep learning model includes an input layer, an LSTM layer, a linear layer, a multi-head attention layer, a dropout layer, a flattening layer, a multi-layer perceptron, and an output layer; The update expression of the LSTM layer is: where h t represents the hidden state at time t, σ represents the sigmoid function, and x t represents the input at a single time step t, tanh represents the hyperbolic tangent, and W i , W f , W o and W c all represent trainable LSTM layer weight matrices, α i , α f and α o all represent learnable scalars of the LSTM layer, β represents the trainable coefficient of the LSTM layer, and γ ∈ [0, 1] represents the mixing ratio of the old and new memories in the output hidden state; The expression of the multi-head self-attention layer is as follows: Among them, Attention(Q, K, V) represents the output of the multi-head self-attention layer, x represents the input of the multi-head self-attention layer, Q, K, and V represent the query matrix, the key matrix, and the value matrix respectively, and W Q , W K and W V all represent the weight matrices of the multi-head self-attention layer, α Q , α K and α V all represent the trainable scalars of the multi-head self-attention layer, q0, k0, and v0 are all learnable factors, τ is the scaling factor, and d head represents the dimension of a single attention head of the attention layer.

9. The method for predicting the lifespan of multi-scale electronic components based on physical deep learning according to claim 8, wherein The loss function of the deep learning model The expression is: In the formula, represents the change amount model loss error of the on-resistance, represents the threshold voltage model loss error, represents the comprehensive index model loss error, represents the deep learning model loss error, w R 、w V 、w H and w D represent the loss weights corresponding to the loss errors, and there is w R +w V +w H +w D = 1.

10. The method for predicting the lifespan of multi-scale electronic components based on physical deep learning according to claim 9, wherein The loss error of the variation model of the on-resistance The expression is as follows: The loss error of the threshold voltage model The expression is as follows: The loss error of the comprehensive index model The expression is: The loss error of the deep learning model The expression is as follows: where \(n = 1, 2, \cdots, N\) k , \(N\) k represents the total number of samples in a batch, \(\beta\) R , \(\beta\) V , \(\beta\) H and \(\beta\) D all represent scaling factors, \(\gamma\) D , \(\gamma\) R , \(\gamma\) V and \(\gamma\) H all represent translation factors, \(\cosh\) represents the hyperbolic cosine function, \(x\) kn represents the \(n\)th input feature of the \(k\)th device, \(\Delta R\) pred represents the on - resistance of the device predicted by the model, \(\Delta R\) phy represents the on - resistance obtained from the physical model, \(\Delta V\) pred represents the threshold voltage of the device predicted by the model, \(\Delta V\) phy represents the threshold voltage obtained from the physical model, \(\Delta H\) pred represents the comprehensive index predicted by the model, \(\Delta H\) phy represents the comprehensive index obtained from the physical model, \(Y\) pred represents the remaining useful life value of the device predicted by the deep - learning model, \(Y\) actual represents the actual remaining useful life value of the device, and \(\theta\) represents the training parameter of the deep - learning model.

Images