Artificial Intelligence for Power Electronics in Electric Vehicles: Challenges and Opportunities

Artificial intelligence (AI) is now playing a ubiquitous role in accelerating the technology development and improving the robustness of various systems in renewable energy and energy efficiency sectors. Efforts to successfully utilize AI in these fields have benefited from, among many reasons, the confluence of the ability to extract large volume of data and the affordability of massive high-performance computing resources. The prediction capability of AI techniques can assist in energy forecasting and maximizing energy generation from wind and solar power plants [1,2]. AI can be considered as an overarching field in computer science which studies different ways to impart humanlike intelligence to computer systems, whereas machine learning (ML) is a subset of AI that identifies patterns within datasets and uses that information to classify or explore the relationship between different variables within the datasets. In the transportation sector, ML algorithms are finding a myriad of applications such as predicting consumer charging behavior, suggesting routes to electric vehicles (EVs) for maximizing driving range, and operating EV fleets for ride-hailing services [3–5]. In addition to system-level benefits, AI is driving key innovations in vehicle electrification at the component level. ML techniques developed to predict a battery's performance early in its cycle life [6] and optimize fast charging of batteries [7] can potentially have a significant impact in making widespread adoption of EVs a reality. Furthermore, power electronics inverters and electric machines can benefit from the application of AI and ML algorithms in their design, operation, and maintenance and a detailed study on these topics is provided in Refs. [8] and [9]. In this perspective, we turn our attention toward the use cases of AI in the field of automotive power electronics, with the goal of identifying major challenges that need to be overcome to realize a successful impact in the EV sector. Specifically, we focus on the following topic areas related to automotive power electronics modules: semiconductor devices, power module design and prognostics, and thermal management (Fig. 1). The literature that we review is not limited to automotive applications of power electronics; the pervasive nature of power electronics allows a great deal of knowledge transfer among multiple domains, and this is especially true in the case of AI techniques for the aforementioned aspects of automotive power electronics. We conclude this perspective with a brief discussion on the opportunities and challenges in accelerating the development and adoption of AI techniques for power electronics in EVs. The objective here is to stimulate ideas and discussion within the power electronics community to address the typical challenges associated with the adoption of AI and ML techniques.

The EV industry is currently in a major transition from silicon to wide-bandgap devices such as silicon carbide (SiC) and gallium nitride (GaN) for power electronics applications in electric drivetrains [10]. Gallium oxide (Ga₂O₃), an ultrawide-bandgap device with low-cost manufacturing methods [11], has the potential to make a revolutionary impact but is still in the research and development phase. The importance of technological innovations in semiconductor materials and device architectures for automakers cannot be overstated, as system-level targets such as cost reduction, miniaturization, and power density improvement are closely tied to device performance and reliability. A recent report by McKinsey & Company [12] predicts that AI will make significant strides in the semiconductor industry over the next few years, but the current value generation through the application of these techniques is only a fraction of the realizable potential. The semiconductor device manufacturing sector stands to gain the most value from automation through AI due to the large inherent costs involved. A similar argument is made in the 2020 IEEE International Roadmap for Devices and Systems [13], which states that AI applications can result in a dramatic drop in the device production processing times.

Although expert systems were investigated in the early 1990s for semiconductor manufacturing [14], ML algorithms are only recently being deployed in semiconductor fabs on a larger scale for tasks such as wafer defect detection and classification, yield prediction and improvement, and predictive maintenance of the manufacturing equipment. Automated defect classification techniques are already used in fabs; however, their accuracies are only around 80% and need a final verification by a human operator. X-ray and scanning electron microscope images that reveal the minute features and defect patterns in wafers are ideal data sets for training by deep learning algorithms. Yuan-Fu [15] applied singular value decomposition to extract high-dimensional features, performed hyperparameter tuning on a wafer map dataset, and trained different supervised learning algorithms to classify the wafer defects. A convolutional neural network was also trained on the same dataset. The convolutional neural network model outperformed the supervised learning techniques and offered a classification accuracy of 99.2%. Nakazawa and Kulkarni [16] used a convolutional neural network for wafer map defect pattern classification, achieving an overall classification accuracy of 98.2%. Interestingly, they used simulated wafer maps as the training data, as real data on defect patterns from experimental wafer maps were found to be highly imbalanced. This approach also allowed the inclusion of well-defined data on rare defects, which would otherwise be harder to obtain from the actual wafer maps. Jiang et al. [17] introduced a generic, scalable ML framework to predict the final test yield of semiconductor wafers at the fabrication stage itself. This framework incorporated data associated with a variety of manufacturing parameters and can even be customized to focus on a single production step.

Beyond the manufacturing stage, AI is being increasingly explored in chip physical design and verification, and for field-use monitoring purposes. Major players in the tech industry are starting to use AI algorithms for accelerating the chip design life cycle. Researchers at Google recently formulated a deep reinforcement learning approach for chip floorplanning [18], developing an edge-based graph convolutional neural network that can perform chip placements in just 6 h compared to several weeks by a manual design process. In a review article, Khailany et al. [19] identified growth opportunities for ML algorithms in chip design flow and introduced their work to optimize the very large-scale integration design process using a combination of graphics processing unit acceleration, neural network predictors, and deep reinforcement learning techniques. Generative adversarial networks are also being investigated to explore the chip design space and advance design principles for manufacturability [20,21]. It is noteworthy that AI in chip design is heavily tailored toward application specific integrated circuits in the micro-electronics industry; however, adapting and building upon these AI best practices by automotive power electronics device designers and manufacturers will become crucial in the efforts to eventually improve EV sales. Efforts to reduce material losses and improve efficiency and productivity at different stages of chip manufacturing become even more critical in the context of wide-bandgap and ultrawide-bandgap devices. Furthermore, the ML algorithms can be extended beyond the chip phase to electronic design automation of power electronics circuits. Similar to chip floorplanning, the optimal placement of the different passive components and its performance prediction can be solved using ML techniques.

Leveraging AI in the manufacturing and design of power devices will not be as easy and straightforward as a black box tool; designers will need to carefully identify the synthesis and process steps where ML or deep reinforcement learning techniques can be more effective in replacing—or in most cases augmenting—the manual operation efforts. The prowess of AI is best utilized in semiconductor manufacturing by combining it with deep domain expertise. More data sets are helpful to accelerate the use of AI in this area; however, the complexity and lack of enough field data can be sidestepped by training the AI on synthetic data, as demonstrated by Nakazawa and Kulkarni [16]. An area of interest would be conducting high-fidelity crystal growth simulations to generate synthetic datasets on which AI can be applied with an aim to reduce the defect distribution. With the increased utilization of AI in semiconductor manufacturing processes such as crystal growth [22], lithography, metrology, defect detection/classification, and yield prediction and improvement [23], it is expected that chip companies will be better prepared to manage any future disruptions to the global supply chain, thereby alleviating the impact on the automotive industry. As more semiconductor companies are now exploring the potential of AI, it will likely be integrated in a few of these manufacturing steps (Fig. 2) over the next few years with the goal to accelerate product development and minimize the overall production time.

A power electronics module can be defined as the integrated system of components that accommodate the semiconductor devices while enabling the transfer of electrical signals to and from external connections in addition to providing thermal pathways and mechanical support. Designing a power module is inherently a multiphysics task, as electrical, thermal, and mechanical functionalities need to be considered. Although the devices are typically made of Si, SiC, or GaN, there are some variations among automakers in the design of other components within an automotive power module [24–26]. These include the overall circuit layout, type of interconnects for electrical routing, materials for interface attachment, insulation layers, and heat sink designs. In certain cases, discrete packages are preferred over integrated power modules.

A survey conducted by Zhao et al. [8] revealed that the majority of AI applications for power electronics systems are historically geared toward control techniques, although rising trends are recently noted in design and maintenance. The design of a power module can be conceptualized as a multi-objective optimization task with the goal of minimizing parasitics and overall thermal resistance or junction temperature while improving power density and reliability. Cost reduction is a primary target in the automotive industry, which can be factored into the design process through metrics such as power density (capital cost) and reliability (lifetime cost). Ji et al. [27] conducted an optimization study of a standard insulated-gate bipolar transistor (IGBT) power module to improve its lifetime under both power cycling and thermal cycling loads. A surrogate model was initially created to capture the correlation between geometric design parameters as inputs and thermal resistance and strain energy density as outputs. A nondominated sorting genetic algorithm (GA) was then applied to explore the design space and obtain Pareto-optimal solutions. Dragičević et al. [28] incorporated design-for-reliability principles to develop an automated tool for optimizing the design of a power converter. Their design methodology exploited the computational speed of artificial neural networks in discovering the relationship between design and operating parameters on one end and junction temperature and lifetime consumption on the other. This approach is unique in the sense that lifetime consumption was calculated under a mission-profile loading condition as opposed to accelerated profiles such as thermal or power cycling. Although this design optimization work was conducted on a photovoltaics inverter, the technique is generic and can be used as a building block to design similar tools for automotive power electronics systems. A study by Nwanoro et al. [29] found superior thermomechanical performance by Al ribbon bonds compared to wire bonds under power cycling and used a multi-objective GA to optimize the ribbon loop height and thickness. The objectives were to minimize the plastic strain range within the ribbon bond and its temperature. In another study, Pang et al. [30] adopted a sequential approach to optimizing a power module design to improve its electrical and thermal performance. Parasitic inductance and capacitance were first extracted through numerical modeling, followed by an electromagnetic interference analysis to determine the device electrical stress and common mode current. The best performing electrical layouts were then selected for the thermal simulations, where the component thicknesses constituted the design parameters and the objective was to minimize the junction temperature. The sizing of the interface attachment layers was noted to have a higher impact on the junction temperature than the ceramic layers. A common aspect among these studies is the use of AI techniques for optimizing the design and not necessarily in creating the design itself (i.e., number of devices and component layers, electrical interconnect and material types, and electrical isolation techniques), which depends on the overall electric drive architecture, cost, and scalability, and is mostly based on the design expertise of the manufacturer. In Ref. [31], Noor provides the example of an airplane partition structure designed by a generative design tool to predict the capabilities of AI in transforming the engineering design process. However, it is important to note that the data that is used to train an AI-based design tool consisted of human inputs. As such, the outputs of a generative design tool can be considered only as an advanced version of the engineer's design albeit with intricate patterns and features.

Overall, the progress made in the power module design process through implementing advanced optimization algorithms and targeting design for reliability principles is encouraging. These approaches can even be extended to printed circuit board-based inverter designs to predict the copper spread and number of thermal vias to minimize the overall thermal resistance. Nevertheless, the sequential design approach (electrical, thermal, and thermomechanical) is still the widely accepted practice today. It is imperative that the multiphysics functionalities of a power module be properly accounted for early in the design stage to achieve a balance between conflicting objectives. To this end, multi-objective optimization tools such as PowerSynth [32] are a step in the right direction. Furthermore, the design team should have a detailed practical knowledge of the combined realm of electrothermal and reliability principles. Not only will these practices lead to shrinking the power module developmental cycle but also significant cost benefits by limiting the number of design iterations. The designers must also ensure that the optimized designs can be manufactured without incurring high costs. Future power module design efforts will need to consider design for reuse, recycling, and remanufacturing principles, but it is still not clear how AI will play a role in this area. One idea is to impose a modularity constraint within the design optimization tool that will help in easily separating out the failed components from the salvageable healthy components.

Power module prognostics are another key area where AI can play a role in tasks such as condition monitoring, fault prognostics, failure detection and classification, and remaining useful lifetime (RUL) prediction. Compared to manufacturing and design, research efforts on AI techniques for RUL estimation and condition monitoring of power electronics systems are relatively advanced in terms of ideation and prototyping. This trend can be attributed to the overall progress achieved in the field of data-driven prognostics and health management of electronics, as well as sensor-based data collection tools. Also, failure rates in power electronics can be quite high, and reliability improvements of individual power electronics components can significantly reduce system-level costs.

Several studies [33–37] have used lab data collected for performance characterization and degradation monitoring of a power module to make predictions about its future health. The data-driven techniques in this field include but are not limited to particle filters, Kalman filters, Gaussian process regression, neural networks, and support-vector machines. Ali et al. [38] studied variations in the collector–emitter voltage drop of different discrete IGBTs subjected to accelerated power cycling experiments and developed a real-time RUL model. The proposed lifetime model employed a Gaussian process regression incorporating Bayesian inference. Alghassi et al. [39] developed a time-delay neural network to estimate the health status of IGBT modules under power cycling and used Monte Carlo simulations to minimize the uncertainty associated with the RUL prediction. The collector–emitter voltage drop in the device was again used as the characterizing feature in this study. Olivares et al. [40] compared the performance of different ML algorithms such as logistic regression, random forests, and gradient boosting in their attempts to develop a generalized approach to estimate the reliability of power metal-oxide-semiconductor field-effect transistors under thermal and radiological stress. Peyghami et al. [41] conducted stress-strength analysis to determine the constant lifetime curves of a power electronics converter under a limited set of operating conditions. An artificial neural network algorithm was then trained on the data obtained from the preliminary analysis to expand the scope of lifetime estimation to a wide range of operating conditions.

A common goal among these studies on power device reliability using ML techniques is the attempt to create a real-time device RUL monitoring technique; however, these concepts are yet to transition from the research space to a commercial application. The challenge for EV companies is to minimize the capital cost of implementing a device monitoring circuit within the inverter. The associated long-term cost savings, although important, might not be sufficient to sell this as a feature to potential EV customers. A feasible solution would be to activate the device monitoring circuit at select intervals and compare the performance against a predefined threshold [42]. As a result, the amount of data generated can be minimized, but not at the expense of any information loss related to the degradation of devices or modules.

It should be noted that although device characteristics such as collector–emitter voltage or on-state resistance are usually selected as the temperature-sensitive precursor for condition monitoring and RUL prediction, the actual failure mechanisms are likely to occur within the bonded attachment layers or the wire bonds. Hence, it is important to leverage data-driven techniques with physics-of-failure models [43], which are necessary to develop the cause-effect understanding of the failure mechanisms. A drawback of these fusion models is that the physics-of-failure approach is based on a macroscale analysis of the thermomechanical behavior of the critical components within a power module, but the true causes of failure occur at the microstructural level. To incorporate microstructural data in the development of lifetime prediction models, multiscale modeling algorithms can be employed, with ML techniques facilitating information transfer between different length and time scales [44].

Operating conditions of power electronics devices in the field are much more complex than those performed in the laboratory. While laboratory tests are typically conducted under controlled temperatures and in a periodic fashion, operation in the field involves widely fluctuating temperatures and vibrations. RUL predictions in the laboratory are relatively simple compared to that of power electronics inverters deployed in electric vehicles. However, laboratory tests typically involve much smaller sample numbers than the number of inverters deployed. For example, ML techniques can be used to predict the degradation of solder layers or thermal interface materials during its lifetime. Under accelerated cycling or aging conditions of these materials, data such as thermal resistance or scanning acoustic microscope images can be obtained periodically to monitor its degradation behavior. With sufficient data, a time-series analysis can be performed using either statistical models or ML techniques to develop a defect/crack propagation prediction model. The scope of lifetime prediction models developed under laboratory conditions can be expanded by incorporating data from the field conditions. An ideal interaction between data from deployed assets and laboratory data as shown in Fig. 3 can be envisioned. Data in the laboratory typically involve multimodal and operando characterization techniques with physics-based models, but are limited in supply, whereas data in the field can be relatively simple (electrical and thermal) but plentiful. The plentiful data in the field may be sufficient to overcome the challenge of RUL prediction of such variable conditions, while laboratory tests can focus on conditions where vulnerabilities are exposed. As outlined in Fig. 3, a data collection strategy for field-deployed assets can be informed from lab-based testing where vulnerabilities to specific operating conditions are identified, e.g., time-resolved temperature, vibration, and electrical exposure. These operating conditions are then prioritized for measurement in deployed assets. As is common in today's deployed assets such as electric vehicles and portable electronics, communication between the operating system and the company's data and research facilities is common. Performance data of the deployed system along with data on the prioritized operating conditions can be gathered by the company to help build ML predictions on future system performance. However, while predictions based on data from deployed assets may be accurate and highlight vulnerabilities to specific operating conditions, they are of limited value for understanding failure modes. The operating conditions that are shown to correlate with reduced performance most significantly can be prioritized for lab-based research where more advanced characterization tools can be utilized. Information from the laboratory can then be used to inform measurement strategies in the field as well as help prioritize directions for further research and development. This synergistic approach of leveraging lab and field data is broadly applicable and is expected to help accelerate product development and accurate prediction of the performance of power electronics units in the field. Management of uncertainty before using the ML predictions will also need attention. A large amount of robust and high-quality data will likely be needed before predictions reach an acceptable level of uncertainty to inform decisions. There is a considerable challenge in creating a framework to collect and process data in this way. Significant progress on this type of interaction as well as on training an experienced workforce could be accelerated by open-source data from field and laboratory sources that emerging researchers can use in research institutions around the world.

Effective thermal management of power modules is essential to keep the device junction temperature under safe operating limits and prevent thermal runaway. In automotive applications, single-sided, single-phase liquid-cooling is the widespread choice due to its cost-effectiveness and moderate heat transfer coefficients, designed in the form of mainly channel flow and pin fin configurations [45]. However, double-sided liquid-cooling and air-cooling solutions are gaining traction in the context of wide-bandgap devices and miniaturization. Also, jet impingement flow configuration is now seen as a promising solution in electric vehicles due to its high heat transfer rate and improved thermal performance [46].

Although design and optimization of heat sinks is, in general, mature, review articles of thermal management for power electronics within electric vehicles suggest that the use of any AI techniques is practically nonexistent in the commercial space. This could possibly be due to the sufficient level of thermal performance offered by the conventional heat sink designs in dissipating heat loads typically generated in automotive silicon-based IGBT power modules. Nevertheless, studies that employ metaheuristic optimization algorithms to design heat sinks specifically to address the challenge of high heat fluxes from wide-bandgap devices are now picking up pace, as seen in the literature. Wu et al. [47] conducted design optimization of an air-cooled heat sink attached to a phase leg module of a 50-kW three-phase SiC inverter. This study utilized a combination of GA for the optimization process and finite element simulations to evaluate the thermal performance of each iterative candidate design. In comparison with a customized design from a heat sink manufacturer, the final optimized product was 27% less volume and achieved an ∼6 °C drop in junction temperature. Michalak et al. [48] also used a GA to optimize the jet impingement pattern in a heat sink with a multijet array configuration. The coolant was water-ethylene glycol mixture for a target application of a printed circuit board-embedded SiC onboard charger in EVs. The optimized heat sink geometry resulted in an ∼10 °C drop in average device junction temperature compared to the initial design but at the expense of a 12-kPa increase in pressure drop. Another application of GA for liquid-cooled heat sink design can be found in the optimization study by Gurpinar et al. [49], who used Fourier series to represent the fin height and aimed to improve the power density of the module and minimize the thermal imbalance among the SiC devices. The harmonic terms in the Fourier series were selected as the design variables. Finite element analyses showed that the optimized design reduced the thermal imbalance by 75% while achieving a twofold improvement in power density compared to the initial design, although a 7 °C rise in junction temperature was observed. In addition to GA, density-based topology optimization approaches have gained popularity for cooling power electronics in EVs. Dede et al. [50] employed a topology optimization routine to design a three-dimensional heat sink with pin fins for jet impingement air cooling and compared against high-performance benchmark solutions. The optimized heat sink, which was additively manufactured using AlSi12, surpassed the conventional heat sinks in coefficient of performance by a range of 7–44%, depending on the different designs considered in the study.

A major drawback of the optimization approaches to heat sink design is the associated manufacturing complexity associated with conventional machining methods. In almost all cases, three-dimensional printing is the only viable solution that can accommodate the geometric design features obtained as an output of the optimization process. Although optimization algorithms can create novel, unconventional heat sink designs, the level of adoption of these approaches by the EV industry remains to be seen. Rather than continuing to search for performance improvement, there is a greater possibility of one or two cost-effective cooling approaches being widely adopted once they are proven to reliably manage the heat flux levels from wide-bandgap devices. Also, the automotive industry is typically locked into a design cycle for several years, which leaves opportunities for transformational improvements few and far between.

The application of AI in automotive power electronics is certainly a growing area and will continue to make significant technological innovations in the electric vehicle drivetrain. In this section, we identify a few opportunities and challenges to fully utilize the potential of AI for power electronics in EVs on a commercial scale.

• A power electronics module comprises multiple component layers with different functionalities, and the material selection for each of these components is an important design decision. Designing novel material compositions with enhanced functionalities using ML techniques is an area worth exploring. Interface attachment materials with a graded coefficient of thermal expansion and elastic modulus could mitigate the failure mechanisms commonly observed in solder or sinter layers.

• Similar to a battery management system, a monitoring system can be developed to manage the health of the power electronics inverter and can even be extended to include the electric motor. A cloud-based digital twin of the electric drivetrain for condition monitoring and RUL prediction can add value to the automaker in improving the overall system reliability and reducing the maintenance costs. Based on the trends observed in the precondition variables within the digital twin, feedback control signals can be sent to the inverter to adjust the operating conditions such as derating the devices or controlling the coolant flowrate and path to address the occurrence of any hot spots. Figure 4 demonstrates this idea applied to the cooling system. Electrical perturbations typically arise before thermal responses and can be used to forecast cooling needs that will be required locally within a power electronics module. ML methods can predict the time-resolved cooling required for certain local electrical perturbations, presenting opportunities to reduce the size of cooling systems, saving mass and cost in EVs. Similar active thermal management strategies are discussed in Refs. [51] and [52].

• Even though autonomous vehicles on a large scale might not transpire in the near future, EVs are the preferred candidates of the different companies in this area. As such, it is critical to prevent any cyber-physical attacks [53] on an autonomous electric drivetrain, and ML-based solutions can help detect any malicious disruptions to the power electronics systems and electric motor.

• As in most other fields, the four V's of data—volume, variety, veracity, and velocity—deserve a great deal of attention in power electronics. Data collection in certain domains, such as semiconductor manufacturing, can be very challenging. The notion that ML is only as good as the data content holds true universally, although there is a recent emphasis on working with minimal data. Following NASA's prognostics data repository model [54], an open-source database purely for power electronics component and system failures under the leadership of a consortium can go a long way in catalyzing ML applications in this field. Furthermore, emphasis should be given to understanding the role of different data types and their combinations to improve the accuracy of the ML algorithms.

• Although ML algorithms can simplify the task of estimating the lifetime of a power electronics system, it is critical to quantify the uncertainty associated with these predictions. The scope of an ML algorithm in what it can and cannot do must be delineated. For example, ML cannot reveal the fundamental causes of failure mechanisms within a power electronics module. Data interpretation must be combined with physics-based mechanistic models to really understand and improve the accuracy of lifetime predictions.

• An often understated cause of the growth of a new technology or its application in a certain field is the level of funding dedicated to research and development. Innovative AI techniques or ideas that involve the application of ML algorithms to minimize the cost and improve the reliability of power electronics systems must be strongly supported. Government agencies and industry consortiums can take a lead role in creating a roadmap for AI in power electronics, developing specific research programs, and training the workforce to advance and sustain progress in this area.

Power electronics systems are at the heart of the electrification and decarbonization of the global economy. AI has witnessed an unprecedented growth in the past decade and made significant contributions in various life cycle stages of power electronics. In this article, we briefly review the progress achieved in the application of AI to design and manufacture the critical components of a power electronics module and convey our perspective on opportunities and challenges. We note that AI techniques are starting to be investigated for automotive power electronics, especially in the design stage, and transitioning these concepts to a commercial scale can occur in the next few years, given the emphasis of major automakers in shifting their focus to EVs. There is ample opportunity for further adoption of AI techniques across a multitude of power electronics research areas, from semiconductor design to more effective control and cooling strategies. However, the research community is clearly faced with a challenge of obtaining robust, plentiful, and relevant data to train and advance AI techniques for power electronics. To enable accelerated growth of AI techniques for power electronics, more high-quality, open-source data sets are needed. Research institutions such as universities and national laboratories can release (and have already released) open-source databases for power electronics prognostics, but the data are often limited in scope by having highly specific and controlled design and operating conditions. Industry institutions have access to a plethora of extremely valuable field data that would also help further accelerate AI techniques and technology development.

The authors acknowledge support and guidance for the work provided by John Farrell, Laboratory Program Manager, Vehicle Technologies and Johney Green, Associate Lab Director, Mechanical and Thermal Engineering Directorate, National Renewable Energy Laboratory, Golden, CO.

This work was authored by the National Renewable Energy Laboratory (NREL), operated by Alliance for Sustainable Energy, LLC, for the U.S. Department of Energy (DOE) under Contract No. DE-AC36-08GO28308. The work was supported by the Planning and Assessment program at NREL. The views expressed in the article do not necessarily represent the views of the DOE or the U.S. Government.

• National Renewable Energy Laboratory (Funder ID: 10.13039/100006233).

No data, models, or code were generated or used for this paper.