Lifetime Prediction of GaN Power Devices Based on COMSOL Simulations and Long Short-Term Memory (LSTM) Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors presented a machine learning method to predict GaN lifetime. However, the manuscript lacks critical details in several areas, which significantly undermines the reproducibility and credibility of the work. The work in its current form is not justified to be considered a publication.
1.    “Long Short-Term Memory (LSTM) networks, as a class of recurrent neural networks 70 (RNNs), are particularly advantageous for processing time-dependent sequences, making 71 them well-suited for modeling degradation processes that evolve over time.” This has been stated in the previous paragraph and is repetitive. 
2.    What type of GaN power device was studied? There is no mention of the manufacturing method, device geometry, use case, etc.
3.    “These include careful sequence design to ensure temporal coherence” Explain “temporal coherence”
4.    It’s not clear what failure mode does figure 2 represent. Need more detailed captions.
5.    The decapsulated device shows wire bonding is a major failure point; but wire bonding is not inherently related to GaN material. Can the wire bond failure be prevented with more advanced bonding techniques?
6.    Where do chloride ions, copper, and Si contamination come from? Should the contamination still cause issues if a different oxide is used, such as SiO2 or SiNx?
7.    Elaborate “To predict component lifespan, simulations were conducted using Icepak and compared with experimental data. The simulated results deviated by no 168 more than 5% from actual measurements.”  Why using Icepack here instead of  COMSOL, and how were the experiments conducted? 
8.    What is thermal-thermal coupling in Figure 3?
9.    The COMSOL modeling requires significantly more details. What’s the input in the electro-thermal coupling modeling? Do you consider different Vds and Vgs? If so, what are the electrical parameters used in the model, such as electron mobility, 2DEG properties, contact resistance? What kind of devices did you simulate? What is the substrate, material stack, and layer thicknesses?
10.    A device schematic should be included.
11.    “We monitor the drift of the conduction resistance or the change in threshold voltage, combine infrared thermal imaging to locate hotspots, and analyze 204 failures such as solder layer peeling or gate metal electrical migration.” Are the authors testing a real device here? GaN is transparent to IR, how is the hotspot identified? 
12.    The authors mention Figure 4 depicts multiple different devices, but it’s not clear how these samples are different from each other. Are the devices having different structures, or are the only packaging different? It is also not clear what’s the difference is between figure 2 schematics in Figure 4a. 
13.    The differences in the packaging also lack description.
14.    The weights are described as being "empirically determined". The authors provide no justification or systematic process for selecting these specific weights, making this critical step seem arbitrary and unscientific.
15.    The LSTM network is not learning from raw experimental data but is instead being trained to mimic Weibull-Arrhenius model. The final validation is then performed against "actual measured real data," but the details of these experiments are also missing. This defeats the purpose of using machine learning to find convoluted patterns in experimental data. 
16.    Figure 7f has no units

Author Response

Response to Reviewers

We sincerely thank the reviewers for their careful reading of our manuscript and for providing valuable comments and suggestions. We have carefully revised the manuscript according to the feedback and believe that the revisions have significantly improved the clarity, quality, and scientific content of the work. Below, we provide a point-by-point response to each comment.

Reviewer 1 Comments and Responses

Comment 1: “Long Short-Term Memory (LSTM) networks, as a class of recurrent neural networks 70 (RNNs), are particularly advantageous for processing time-dependent sequences, making 71 them well-suited for modeling degradation processes that evolve over time.” This has been stated in the previous paragraph and is repetitive. 
Response:
       Thank you for pointing this out. Therefore, I have removed the repetitive parts. This part was originally located at the 68th line on the second page.

 

Comment 2: What type of GaN power device was studied? There is no mention of the manufacturing method, device geometry, use case, etc.
Response:
       We thank the reviewer for the comment. The GaN device used in this research is a GaN HEMT MMIC Power Amplifier chip. Therefore, I have added relevant information about the manufacturing process and applications of GaN devices in the introduction section of the article. This part is located at line 150 on page 4.

 

Comment 3: “These include careful sequence design to ensure temporal coherence” Explain “temporal coherence”

Response:
       We appreciate the detailed suggestions. The temporal consistency here refers to the requirement of sampling according to the same time standard when collecting data. Therefore, I have provided an explanation for it in the original text. This part is located at line 75 of the second page.”

 

Comment 4: It’s not clear what failure mode does figure 2 represent. Need more detailed captions.

Response:
       We appreciate the reviewer’s insightful comments . The fault shown in the figure is caused by the fact that the voltage borne by the capacitor medium exceeds the designed withstand voltage value, resulting in excessive concentration of local electric field and ultimately causing the medium to break down. Therefore, I have provided relevant explanations in the article. This part is located at line 134 on page 3.

 

Comment 5: The decapsulated device shows wire bonding is a major failure point; but wire bonding is not inherently related to GaN material. Can the wire bond failure be prevented with more advanced bonding techniques?

Response:
       We thank the reviewer for this valuable suggestion. We obtained the data through thermal simulation, without involving semiconductor simulation. The main focus of our work was on the algorithms. In practical operation, the ultrasonic/thermal compression process can be improved to ensure a tight bond at the solder joints. Control the bonding parameters (pressure, ultrasonic power, temperature) to avoid voids and cracks. Enhance the cleanliness of the package to prevent contaminants from reducing the bonding strength of the solder joints. It is also possible to change the packaging technology and adopt flip-chip, lead-free packaging (LGA, BGA, WLCSP), and embedded die packaging to avoid wire bonding failures.

 

Comment 6: Where do chloride ions, copper, and Si contamination come from? Should the contamination still cause issues if a different oxide is used, such as SiO2 or SiNx?
Response:
       We appreciate the reviewer’s insightful comments. In the manufacturing process of GaN devices, chlorine-based gases (such as Cl₂, BCl₃, HCl) are commonly used for etching or cleaning. The residual chlorine ions are prone to cause contamination. The chloride ions (Cl⁻) can easily lead to metal corrosion, electrode failure, increase leakage current and reliability risks. Copper impurities usually come from electrode materials, while silicon usually originates from the glass fiber support material. Even if other different oxides are used, impurity contamination still has an impact, although the manifestation may be different. Therefore, I have provided supplementary explanations in the original text. This part is located at line 172 on page 4.

 

Comment 7: Elaborate “To predict component lifespan, simulations were conducted using Icepak and compared with experimental data. The simulated results deviated by no 168 more than 5% from actual measurements.” Why using Icepack here instead of  COMSOL, and how were the experiments conducted?

Response:
       We thank the reviewer for this valuable suggestion. This place was indeed a typing error. The simulation software we finally chose was COMSOL. Therefore, I have replaced "Icepak" in the original text with "COMSOL". This part is located at line 179 on page 5.

 

Comment 8: What is thermal-thermal coupling in Figure 3?

Response:
       Thank you for pointing this out. The expression here is not precise enough. It should actually be the thermal-mechanical coupling. Therefore, I have changed "thermal-thermal" in the figure to "thermal-mechanical". This part is located at Figure 3 on Page 5.

 

Comment 9: The COMSOL modeling requires significantly more details. What’s the input in the electro-thermal coupling modeling? Do you consider different Vds and Vgs? If so, what are the electrical parameters used in the model, such as electron mobility, 2DEG properties, contact resistance? What kind of devices did you simulate? What is the substrate, material stack, and layer thicknesses?

Response:
       We thank the reviewer for this insightful comment. Our research is not focused on the semiconductor of a specific device, nor does it involve the simulation of the semiconductor within the device. We are studying the entire power amplification chip. Therefore, we only consider the power and heat dissipation issues brought about by the chip.

 

Comment 10: A device schematic should be included.

Response:
       Thank you for pointing this out. We are not conducting the simulation based on a single component, but rather conducting research on the entire chip. It does not involve research on individual components; instead, it is a simulation of the entire chip. The internal structure of the chip is shown in Figure 2(a).

 

Comment 11: “We monitor the drift of the conduction resistance or the change in threshold voltage, combine infrared thermal imaging to locate hotspots, and analyze 204 failures such as solder layer peeling or gate metal electrical migration.” Are the authors testing a real device here? GaN is transparent to IR, how is the hotspot identified?

Response:
       We thank the reviewer for this insightful comment. The devices we tested were the real ones. During the test, we deposited a layer of black paint on the surface of the GaN device as an infrared absorption/emission material. As a result, the surface would radiate infrared based on temperature, and the thermal imager could indirectly "see" the distribution of hotspots.

 

Comment 12: The authors mention Figure 4 depicts multiple different devices, but it’s not clear how these samples are different from each other. Are the devices having different structures, or are the only packaging different? It is also not clear what’s the difference is between figure 2 schematics in Figure 4a.

Response:
       Thank you for pointing this out. There are differences in the packaging of the tested devices. In the first two figures of Figure 4(a), the heights of the packaging shells are different. One is 5mm and the other is 4.4mm. Therefore, I have added this description to the original text. This section is located at line 296 on page 7.

 

Comment 13: The differences in the packaging also lack description.

Response:
       We appreciate the detailed suggestions. I have added detailed packaging information in the title of Figure 4.

 

Comment 14: The weights are described as being "empirically determined". The authors provide no justification or systematic process for selecting these specific weights, making this critical step seem arbitrary and unscientific.

Response:
       We thank the reviewer for the comment. This weight is the ratio when both the APE (Absolute Percentage Error) and MAPE (Mean Absolute Percentage Error) values are at their minimum in the experimental results. It indicates that, based on the comprehensive data obtained, this weight is the ratio that, on the whole, makes the predicted results closest to the actual situation. Therefore, I have included this basis in the article. This section is located at line 451 on page 12.”

 

Comment 15: The LSTM network is not learning from raw experimental data but is instead being trained to mimic Weibull-Arrhenius model. The final validation is then performed against "actual measured real data," but the details of these experiments are also missing. This defeats the purpose of using machine learning to find convoluted patterns in experimental data.

Response:
       We thank the reviewer for the comment. In the research, we adjusted the temperature data obtained from the COMSOL simulation based on the data from the accelerated decline test. We have added the data table for the accelerated decline experiment. This part is added at the 303rd line on page 7 of the article.

Comment 16: Figure 7f has no units.

Response:
       Thank you for pointing this out. In some of the charts, there were indeed cases of information omission. Therefore, I have made modifications to these charts and added relevant information. The modified images include Figure 4 on Page 7, Figure 7 on Page 11, and Figure 8 on Page 13.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript proposes a hybrid approach that combines COMSOL simulations with LSTM networks to predict the lifetime of GaN power devices. While the topic is relevant and the methodology has potential, the manuscript requires further improvements to strengthen its persuasiveness.

1. Some sections are overly lengthy, particularly the introduction of basic LSTM principles. The methods section should focus more on technical details directly related to the research problem, while simplifying the general background of machine learning.

2. The choice of weighting factors (0.6:0.4) in the combined model (Weibull–Arrhenius + LSTM + COMSOL fatigue) lacks systematic optimization or theoretical justification.

3. Although the manuscript demonstrates the superiority of LSTM in prediction accuracy, the physical interpretability of the results is insufficient. A deeper discussion linking the prediction outcomes with underlying physical mechanisms is recommended.

4. Figures and tables are not prepared in a fully professional manner: some figures are of low resolution with blurred details and unclear annotations. The formatting is inconsistent, and some figures lack complete legends, axis units, or labels. All figures should be redrawn or reformatted to meet academic standards.

5. The manuscript contains minor issues in English grammar and word choice. Careful proofreading and language polishing are required.

Author Response

Response to Reviewers

We sincerely thank the reviewers for their careful reading of our manuscript and for providing valuable comments and suggestions. We have carefully revised the manuscript according to the feedback and believe that the revisions have significantly improved the clarity, quality, and scientific content of the work. Below, we provide a point-by-point response to each comment.

Reviewer 2 Comments and Responses

Comment 1: Some sections are overly lengthy, particularly the introduction of basic LSTM principles. The methods section should focus more on technical details directly related to the research problem, while simplifying the general background of machine learning.

Response:
       Thank you for pointing this out. There were some repetitions in the article when introducing LSTM. Therefore, I have removed the repetitive content. This part was originally located at the 68th line on the second page.

 

Comment 2: The choice of weighting factors (0.6:0.4) in the combined model (Weibull–Arrhenius + LSTM + COMSOL fatigue) lacks systematic optimization or theoretical justification.

Response:
       We thank the reviewer for the comment. This weight is the ratio when both the APE (Absolute Percentage Error) and MAPE (Mean Absolute Percentage Error) values are at their minimum in the experimental results. It indicates that, based on the comprehensive data obtained, this weight is the ratio that, on the whole, makes the predicted results closest to the actual situation. Therefore, I have included this basis in the article. This section is located at line 441 on page 11.

 

Comment 3: Although the manuscript demonstrates the superiority of LSTM in prediction accuracy, the physical interpretability of the results is insufficient. A deeper discussion linking the prediction outcomes with underlying physical mechanisms is recommended.

Response:
       Thank you for pointing this out. On page 11, lines 472 to 487, we conducted a rationality analysis of the hybrid model and explained the physical meaning of the model construction.

 

Comment 4: Figures and tables are not prepared in a fully professional manner: some figures are of low resolution with blurred details and unclear annotations. The formatting is inconsistent, and some figures lack complete legends, axis units, or labels. All figures should be redrawn or reformatted to meet academic standards.

Response:
       We thank the reviewer for this valuable suggestion. Some of the pictures do have the problem of insufficient clarity and incomplete information. Therefore, I have restructured these diagrams and added relevant information. The modified images include Figure 4 on Page 7, Figure 7 on Page 11, and Figure 8 on Page 13.

 

Comment 5: The manuscript contains minor issues in English grammar and word choice. Careful proofreading and language polishing are required.

Response:
       We appreciate the reviewer’s suggestion. I have made some improvements to certain words used in the article and some grammatical errors have been corrected.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have revised the manuscript and made some edits in response to my earlier review. They addressed my comments individually and I appreciate the effort. While these changes improve readability in places, they do not address the core problems I raised in the first round. I cannot recommend this for publication until the major issues are addressed. 
1.    My first concern is still with the COMSOL modeling. In the first review I asked for details of the electro-thermal coupling inputs, including bias conditions, semiconductor parameters, and the device stack. In the response, the authors state they are not modeling a specific device but only considering chip-level power and heat dissipation. Even if that were the case, the manuscript still does not provide the basic information needed for a device-level thermal model: there is no description of the heating area, the location where heat is applied, the density of the heat flux, or the boundary conditions at the interfaces. Without these, the modeling cannot be reproduced. In addition, the paper continues to describe the work as an electro-thermal coupling simulation, yet no electrical inputs, biasing conditions, or semiconductor physics are defined. It is therefore unclear what is being coupled, and the physical validity of the results remains questionable.
2.    My second concern is that the novelty of the work is very limited. The LSTM model is not trained on raw experimental data but on synthetic degradation curves that were generated from COMSOL simulations and then fitted with the Weibull–Arrhenius model. The small set of experimental results added in the revision is only used to make a minor correction to the simulated output and not as the primary input to the training. In practice, the machine learning is being trained to reproduce another model rather than to uncover degradation patterns in real data, which defeats the stated purpose of the approach. In addition, the COMSOL model itself is not a true device-level electro-thermal model because no electrical biasing, semiconductor physics, or explicit heating definitions are provided. This means that the claimed ability to generalize the method to other GaN devices is not convincing, since the underlying modeling is not connected to the actual physics of GaN operation.
Overall, the authors have only superficially addressed my earlier comments. As in the first round, the manuscript still lacks the necessary detail and scientific rigor to support its claims. The issues outlined above remain the main concerns, and without satisfactory justification or revision on these points I do not see value in spending time on minor details elsewhere in the manuscript.

Author Response

Dear reviewer,

We are extremely grateful for your valuable feedback once again. We have carefully revised the manuscript according to the feedback and believe that the revisions have significantly improved the clarity, quality, and scientific content of the work. Below, we provide a point-by-point response to each comment.

Comment 1: My first concern is still with the COMSOL modeling. In the first review I asked for details of the electro-thermal coupling inputs, including bias conditions, semiconductor parameters, and the device stack. In the response, the authors state they are not modeling a specific device but only considering chip-level power and heat dissipation. Even if that were the case, the manuscript still does not provide the basic information needed for a device-level thermal model: there is no description of the heating area, the location where heat is applied, the density of the heat flux, or the boundary conditions at the interfaces. Without these, the modeling cannot be reproduced. In addition, the paper continues to describe the work as an electro-thermal coupling simulation, yet no electrical inputs, biasing conditions, or semiconductor physics are defined. It is therefore unclear what is being coupled, and the physical validity of the results remains questionable.

Response:
We appreciate the detailed suggestions. Regarding your concern about the insufficient details in the COMSOL modeling, we have made the following revisions:

We have revised the expressions in the paper that might cause ambiguity in the device-level simulation.

In the methods section of the paper, we have supplemented the relevant parameters of the chip, including the operating conditions of the chip, the internal dimensions of the chip, the internal materials of the chip, and the simulation boundary conditions. These details ensure the repeatability of the simulation process. Please refer to Tables 1-4 in the article for details.

Regarding the expression of "electro-thermal coupling", we clarified it in the revised version: Our research objective is to model the thermal distribution based on the power dissipation at the chip level, rather than the electro-physical modeling at the specific device level. The paper has been modified to "thermal simulation based on dissipated power inputs", and the coupling relationship is explained as being achieved through the correspondence between power consumption and thermal distribution in the methods section.

 

Comment 2:  My second concern is that the novelty of the work is very limited. The LSTM model is not trained on raw experimental data but on synthetic degradation curves that were generated from COMSOL simulations and then fitted with the Weibull–Arrhenius model. The small set of experimental results added in the revision is only used to make a minor correction to the simulated output and not as the primary input to the training. In practice, the machine learning is being trained to reproduce another model rather than to uncover degradation patterns in real data, which defeats the stated purpose of the approach. In addition, the COMSOL model itself is not a true device-level electro-thermal model because no electrical biasing, semiconductor physics, or explicit heating definitions are provided. This means that the claimed ability to generalize the method to other GaN devices is not convincing, since the underlying modeling is not connected to the actual physics of GaN operation.

Response:
Thank you for pointing this out. Our data mainly comes from simulations, because real test experiments require a large amount of time and resources, so using simulation - degradation models is a feasible preliminary verification method. I have provided the necessary explanations in the article.

We have also made modifications to the application of LSTM. LSTM is no longer used to learn the Weibull - Arrhenius degradation model, but directly learns the data obtained from the COMSOL simulation. The simulation of COMSOL reflects the change of heat over time, which causes changes in thermal stress, affecting the lifespan of the device. LSTM directly learns the temperature data and links the easily observable information of temperature changes over time with the lifespan of the device, skipping the difficult-to-observe stress changes, and achieving a direct mapping of time-varying temperature data to the lifespan of the chip. Therefore, I have updated the relevant content and results in the article.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed the concerns raised in my initial review with reasonable and appropriate revisions. I recommend acceptance of the manuscript.

Author Response

Dear reviewer:,
We greatly appreciate your valuable feedback once again. Thank you for your recognition of this article. We have improved the English expression of the article.

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

I appreciate the authors effort to provide a detailed revision. I think the manuscript need to be restructured. The authors need reconsider their novelty statement, and write the introduction accordingly. 

First, the main advantage I see of using LSTM is it captures time dependence. In the introduction, the authors should focus more on why time-dependence is important for predicting the failure of GaN power amplifiers, instead of saying traditional methods “require extensive experimental data that is costly and time-consuming to acquire”. Like the authors mentioned in the conclusion, using COMSOL to generate data is a limitation, not a feature. The authors can mention that this study uses COMSOL data to develop the method – which can be trained progressively with experimental data in the future.

Second, “empirical models are typically device-specific and offer poor generalizability across different geometries, materials, or usage scenarios.” is a weak argument. GaN devices vary drastically, and their failure modes differ based on their operation conditions. As the authors mentioned, they only did chip-level thermal simulation and did not consider semiconductor physics or varying heat flux during Vgs swing. Real devices indeed should be difficult to generalize. The authors need to focus specifically on their observed failure mode – wire bonding failure for example. Since the failure did not happen at the device level (thermal or electrical breakdown), this gives authors justification to not model device. However, doing so is not a “generalization” as the authors claim.

Author Response

Dear reviewer,

Thank you very much for your meticulous review of our manuscript and your valuable suggestions. The suggestions you made not only helped us better recognize the structural and logical deficiencies of the article, but also provided us with highly constructive guidance for our subsequent revisions. We highly value your feedback and have responded and improved accordingly in the revised version.

Comment 1: First, the main advantage I see of using LSTM is it captures time dependence. In the introduction, the authors should focus more on why time-dependence is important for predicting the failure of GaN power amplifiers, instead of saying traditional methods “require extensive experimental data that is costly and time-consuming to acquire”. Like the authors mentioned in the conclusion, using COMSOL to generate data is a limitation, not a feature. The authors can mention that this study uses COMSOL data to develop the method – which can be trained progressively with experimental data in the future.

Response:
Thank you very much for your valuable suggestions. In the introduction section of the article, I have elaborated in detail on the significance of time-related characteristics for predicting life expectancy. This part is located from line 51 to line 55 of the article. In the abstract section, I emphasized that this method was chosen for the initial prediction because it was not possible to obtain a large amount of data in the short term. Additionally, in the "Conclusion" section of the article, I also explained that if there are sufficient experimental samples in the future, the simulation data can be replaced. This part is located from line 20 to line 28 of the article, as well as from line 571 to line 576.

Comment 2: Second, “empirical models are typically device-specific and offer poor generalizability across different geometries, materials, or usage scenarios.” is a weak argument. GaN devices vary drastically, and their failure modes differ based on their operation conditions. As the authors mentioned, they only did chip-level thermal simulation and did not consider semiconductor physics or varying heat flux during Vgs swing. Real devices indeed should be difficult to generalize. The authors need to focus specifically on their observed failure mode – wire bonding failure for example. Since the failure did not happen at the device level (thermal or electrical breakdown), this gives authors justification to not model device. However, doing so is not a “generalization” as the authors claim.

We are extremely grateful for your detailed suggestions. In the introduction section, I removed the unreasonable descriptions and revised the statements regarding generalization. I also pointed out that the simulations in this paper are based on some common faults of gallium nitride chips (such as thermal breakdown and lead bonding failure), and these faults have temporal correlation in terms of temperature. This part is located from line 56 to line 60 of the article.

Author Response File: Author Response.pdf

Round 4

Reviewer 1 Report

Comments and Suggestions for Authors

My comments have been addressed.