machine intelligence

Introduction Automated production in electronics manufacturing can produce high-quality products, but it might lead to a particular failure without human interventions. With the rapid technology development, such as the Industrial Internet of Things (IIoT), big data analysis, cloud computing, artificial intelligence (AI), many manufacturing processes can be more intelligent, and Industry 4.0 can then be realized in the near future[1]. Smart manufacturing adopts real-time decision-making based on operational and inspectional data and integrates the entire manufacturing process as a unified framework. Then, the future manufacturing process transforms cyber-physical systems digitally and responds to any uncertain situations proactively while ensuring higher efficiency. In Surface Mount Assembly (SMA) lines, equipment status and quality data can be collected via IIoT technology. Data-driven solutions, such as AI and machine learning algorithms, can be applied to diagnose abnormal defects and adjust optimal machine parameters in response to unexpected changes/situations during production. Collaborating with various SMT industry partners, the research team at the State University of New York at Binghamton (aka Binghamton University) developed a novel framework based on AI-based closed-loop feedback control and parameter optimization to implement a smart manufacturing solution in the PCB assembly for yield and throughput improvement. This AI-based framework could provide a potential road map for data-driven process control in SMA. Machine Intelligence in SMA Each SMA process has a critical effect on the final PCB product quality and throughput. Notably, the solder printing process is a critical operation because over 60% of the PCB assembly soldering defects can be traced back to this stage. An inadequate volume of solder paste transferred to any PCB pad is a printing fault, which leads to board failure and substantial reworking and repair costs. The pick and place (P P) process is the highest cost procedure, including expensive machine investment and extended production time. In the soldering reflow process (SRP), the reflow oven temperature and other related settings determine the solder joints' quality and reliability. Hence, multiple inspection machines in the SMA processes have been introduced, including solder paste inspection (SPI) and automated optical inspection (AOI) machines. Particularly, two independent AOIs could be employed to detect the components' defects before and after SRP. Because many electronics components become small-scale (e.g., ), more assembly-related failures are often observed in recent SMA processes. The Smart Electronics Manufacturing Laboratory (SEML) at Binghamton University is fully equipped with two solder paste printers, two chip-mounters, and a reflow oven along with SPI and AOI machines. The research team tested more than 8,000 PCB at SEML. The results show that numerical methods based only on physical properties might have practical limitations in explaining small-scale components' behavioral patterns. It might be caused by unknown environmental factors (e.g., temperature and humidity), machine calibration, measurement accuracies, vibrations, etc., which could have influenced the quality of the SMA outcomes. However, recent research shows that AI-based methods can increase product quality up to 35%, reduce scrap rates, and optimize fab operations in semiconductor manufacturing, compared to traditional approaches[2]. It implies that a data-driven intelligent SMA process control has the potential to advance SMA processes. The goal of the smart SMA is to maintain optimized settings in both offline and online scenarios. The AI and data analytics solution can optimize all SMA process parameters before production (i.e., offline control) and during production (i.e., online control). The overall schematic of the AI-based closed-loop feedback control framework is illustrated in Figure 2. Intelligent SMA Modules In the solder printing process, four machine intelligence modules are considered: Printing advising module (PAM) Printing optimization module (POM) Printing diagnosis module (PDM) Dynamic stencil cleaning process control (CPC) Figure 2. A schematic diagram of the AI-based closed-loop feedback platform Figure 3. PAM effectiveness over customer’s best-known printing parameter setting PAM aims to recommend the ideal initial setting of the printer critical parameters, such as printing speed, printing pressure, and separation speed, using hybrid machine learning and heuristics optimization techniques[3]. As a case study, the research team validated the PAM's performance with an automotive PCB testbed and compared the printing results to the best-known printing parameters. The experimental results show that PAM can achieve over 50% higher Cpk (i.e., process capability index, as shown in Figure 3). POM optimizes printing parameters in real-time by monitoring printing quality and fine-tuning offset and process parameters for adapting to dynamic conditions[4]. The experimental results show that POM achieves more than 30% production quality improvement in terms of the Cpk by adjusting printing parameters compared to the offline control. PDM provides anomaly detection and diagnosis of the potential printing failure cases to improve process quality and reduce downtime[5]. The experimental results show that the PDM can achieve more than 87% accuracy in predicting different types of defects, improper printer hardware issues in the board support, squeegee, paste conditions, etc. The CPC uses the SPI information to estimate residue buildup level on the stencil undersurface and assess the stencil cleaning profile and cycle control, as illustrated in Figure 4[6]. Upon the implementation of CPC, it is expected that the robustness of the printing quality and the Cpk can be improved by 34% and 10%, respectively, compared to the best-known cleaning parameters using in the production line. Figure 4. Smart residue buildup prediction for stencil cleaning operation control During the P P procedure, the mounter optimization module (MOM) and the mounter diagnosis module (MDM) can be applied as a machine's automation process while optimizing the P P machine's parameters. By utilizing self-alignment effects appropriately, MOM identifies the optimal placement position by predicting the component's post-reflow positions based on the data collected by SPI, Pre-AOI, and Post-AOI machines, as shown in Figure 2. MOM also offers the placement positions adaptively during active production. In the MOM framework, multiple dynamic placement options are first generated based on the solder paste offset information. The components' final offsets in both x and y directions are predicted by a hybrid AI model that stacks on the k-nearest neighbor regression and the gradient boosting regression models. The optimal placement, which has the minimum predicted post-reflow misalignment, can be identified by MOM. The experimental results show that MOM can decrease 18% of the final misalignments compared to a conventional P P placement method (i.e., placing a component on the pad center). MDM is a prescriptive and predictive maintenance method that uses P P machine operational and AOI inspection data to trace back the root causes of P P defects and prevent future failure. MDM can achieve an accuracy of 84.50% in identifying the known root causes of certain defects, such as improper nozzle size, parts' contamination, and feeder problem. It shows that different mounting defects can be detected and classified automatically when the abnormality is detected through AI-based diagnosis algorithms. Figure 5. The illustration of the optimal placement position in the MOM One of the reflow oven issues to be addressed is to find the optimal reflow oven temperature settings, which would affect the final quality of the PCB products. Solder paste manufacturers usually provide a target profile based on the solder paste composition's physical properties, and solder joint temperature is required to meet the given profile. Hence, reflow engineers should fine-tune reflow oven temperature manually to ensure a thermal profile outcome from the reflow oven to correctly meet a target profile, requiring substantial cost and effort. The research team proposes an automated reflow recipe optimization model based on the PCB thermal profile and its recipe. Figure 6. Optimized thermal recipe and thermal profile First, the initial recipe collects the data for the prediction model and identifies the relationship between the thermal profile and the corresponding recipe. Then, an AI-based model is developed to predict the thermal profile based on the input recipe. Compared to traditional methods, the AI-based method generates an optimal reflow oven recipe to minimize the gap between the predicted temperature and the given profile. As a result, the AI-based prediction model allows us to achieve promising results, such as 97% of fitness in the given profile temperature curve within one hour of processing time. The proposed model has other significant advantages, such as saving time, labor, and materials. It enhances the degree of automation of the PCB reflow process. In the future, data from multiple inspection machines will be integrated so that the reflow optimization process is fully automated and generates more reliable results. Summary and Conclusion The small-scale electronics products make the SMA processes much more complicated to maintain high-quality PCB products, and theoretical interpretations of the SMA processes can be challenging due to many uncertain factors. With the help of AI and big data collected from various inspectional operations, SMA processes can be intelligent and flexible in response to dynamic environmental situations. While retaining the optimal control parameters throughout the SMA processes, the final PCB product quality can be enhanced while maintaining the designed throughput. Automated and smart systems bring about the opportunity to next level of electronics manufacturing, which utilizes the data and information from the end-users through edge/cloud computing and fastens the customized product manufacturing with increasing efficiency for high-mix/low-volume manufacturing. Also, it can increase verities of design and fasten the delivery time. About the Author Prof. Sang Won Yoon is a recipient of the SUNY Chancellor’s Award for Excellence in Scholarship and Creative Activities in 2019 and a highly successful researcher who leads many productive long-term industry collaborations. Prof. Yoon received his doctoral degree in School of Industrial Engineering at Purdue University, and he joined the faculty of the Watson School in the Department of Systems Science and Industrial Engineering at State University of New York at Binghamton in 2010. Prof. Yoon has been studying how to extract useful insights from expanding data sets to support intelligent decision-making processes. His research not only resides in better understanding large-scale data set by using statistical learning methodologies, but also leverages optimization, soft computing, simulation, and complex theories with conventional machine learning algorithms. As a result, Prof. Yoon has published in over 130 internationally renowned journals and conference proceedings. He was also a member of the Data Science Transdisciplinary Area of Excellence (TAE) initiative and is an active member of the Health Sciences TAE at his institution. The author recognizes the following for their assistance with this article: Daehan Won, [email protected], Assistant professor Jingxi He, [email protected], Ph.D. candidate Shrouq M. Alelaumi, [email protected], Ph.D. candidate Yuanyuan Li, [email protected], Ph.D. candidate Yuqiao Cen, [email protected], Ph.D. candidate References ​​​​​​​[1] Qi, Q., and Tao, F., 2018. Digital twin and big data towards smart manufacturing and industry 4.0: 360 degree comparison. IEEE Access, 6, pp.3585-3593. [2] 10 Ways machine learning is revolutionizing manufacturing in 2019. https://www.forbes.com/sites/louiscolumbus/2019/08/11/10-ways-machine-learning-is-revolutionizing-manufacturing-in-2019/?sh=7cd2e9e22b40. [3] Khader, N. and Yoon, S.W., 2018. Stencil printing process optimization to control solder paste volume transfer efficiency. IEEE Transactions on Components, Packaging and Manufacturing Technology, 8(9), pp.1686-1694. [4] Lu, H., Wang, H., Yoon, S.W. and Won, D., 2019. Real-Time stencil printing optimization using a hybrid multi-layer online sequential extreme learning and evolutionary search approach. IEEE Transactions on Components, Packaging and Manufacturing Technology, 9(12), pp.2490-2498. [5] Alelaumi, S., Wang, H., Lu, H. and Yoon, S.W., 2020. A Predictive Abnormality Detection Model Using Ensemble Learning in Stencil Printing Process. IEEE Transactions on Components, Packaging and Manufacturing Technology, 10(9), pp.1560-1568. [6] Alelaumi, S., Khader, N., He, J., Lam, S. and Yoon, S.W., 2021. Residue buildup predictive modeling for stencil cleaning profile decision-making using recurrent neural network. Robotics and Computer-Integrated Manufacturing, 68, p.102041.

Constant coverage of an invigorating topic like machine intelligence in the media often urges us to consider its use in EDA technology. As is often the case, there are many myths and falsehoods that consume our time and effort when trying to apply machine intelligence to EDA. This article aims to uncover the myths and to provide helpful advice on applying machine intelligence to your EDA project or product.Value PropositionFirst, there needs to be a clear value proposition for adding machine intelligence to an EDA product. Using machine intelligence to create a me-too product adds no value. EDA customers are too busy to understand or care about an EDA tool’s underlying technology. They just want to use the tool and get results. If the tool delivers value, if it delivers tangible benefits, then they’ll use it. Otherwise, they won’t.Currently, EDA tool developers are already experimenting with AI and machine intelligence without considering this fundamental truth – without a higher-end objective. AI must deliver something better or new, whether a speed advantage, a performance advantage, new features, new insights, or perhaps even something pleasantly surprising. Before you write a single line of AI-enhanced code, you need to clearly understand how AI will enhance the product. What is the value proposition?Use ModelThere’s a major barrier to customer adoption of AI and machine intelligence technology for EDA tools: EDA users are averse to make decisions based on probabilistic results. Instead, half a century of EDA tool use has conditioned them to expect deterministic outcomes from their tools.Back in 2003, a prominent visionary and EDA investor was quoted in an interview, saying: “If I open my eyes five years from now, all static analysis in VLSI will be statistical.” Many EDA luminaries have been proven wrong over time for betting that EDA users will accept statistical results. As enthusiastic as I am about using machine intelligence to improve EDA tools, I must urge caution based on the history of EDA failures that employed a probabilistic use model. Decision-makers and EDA tool users want to see deterministic answers to questions about yield or slack, not probabilistic ones.Our experiences at Paripath in developing the PASER (Paripath Accelerated Simulation Environment) tool also bear this out. We discovered that delivering results 50x faster but with 92% accuracy was simply not good enough for end users. EDA users only started to use PASER when its answers became 98+% accurate. To be adopted in the production flow, the tool had to deliver 99% accuracy.Data EngineeringThere are specific ways to achieve these accuracy goals. The first is data engineering. Machine intelligence is a new approach to EDA tool development and it needs to be trained on a data set. If the data is poor or incomplete, training will create an inaccurate model. Fundamental software-development rules still apply. Garbage in, garbage out.Without good training data, there’s no way for you to build good neural-network models. If you train a model with garbage data, you’ll get a garbage model. You must cleanse the data before you use it for training. Otherwise, the model will draw inaccurate conclusions and customers will not use your tool. The model is not to blame here. The model’s not wrong. The problem lies in poor data engineering, poor data cleansing, and a lack of discipline to prepare input data.High DimensionalityNext, machine intelligence has a unique ability to quickly solve problems of high dimensionality. Pure EDA problems often have high dimensionality. Over the years, EDA developers have perfected the art of segmenting the problems into sequencing solutions with lower dimension. Machine intelligence technology can handle problems with thousands of dimensions, but you need to be careful when tackling problems that have high dimensionality. Too many dimensions can produce confused or inaccurate results with AI and deep-learning technology.It helps to visualize the problem and to analyze the data set before using the data to train an AI-enhanced EDA tool. Several visualization methods can help. For example, t-SNE (t-Distributed Stochastic Neighbor Embedding) lets you reduce a data set’s dimensionality from a very large number to a much lower number. Figure 1 shows a high-dimension dataset with a dimensionality of 2000, which has been reduced to a low dimensionality of 3. Figure 1: Visualizing the Data Set with Lower Dimensionality Reducing the dimensionality of a data set to 3 using t-SNE and visualization allows you to quickly see whether the data set defines an easy or a difficult problem. If the problem is difficult, you’ll likely need to lower the problem’s and the data set’s dimensionality before using the data to train a neural network.Technology SelectionOne factor that determines whether it will be easy or difficult to incorporate machine intelligence into your EDA tool is your choice of AI development tools. AI researchers have developed a long list of frameworks, libraries, and languages that they use to develop AI and machine-learning software. Frameworks and libraries such as TensorFlow, Caffe and MXNet are most popular for developing deep-learning models.However, these tools are not yet popular with the EDA development community. The languages of choice in the EDA community are traditionally C and C++ for development and Tcl for prototyping and creating user interfaces. The rest of the software world has moved on to newer development languages such as Python, Java, R, and such. Moreover, machine-learning development segments into two distinct processes: training (i.e. generating the model) and inference (i.e. using the model).Another question to consider is where to generate the model – at the vendor site or the customer site?Consequently, fitting AI and deep-learning development into EDA development environments can feel like fitting a square peg into a round hole. You may need to create corners in your hole.EDA is a very small player in the overall software market. Relatively few software developers are familiar with writing EDA tools. It’s best to select AI and deep-learning development tools that can provide some sort of interface that’s compatible with EDA’s development tools of choice. Some AI frameworks have lower-level C and C++ interface layers that provide a familiar entry point for experienced EDA developers.At Paripath, we chose TensorFlow for exactly this reason. TensorFlow has a lower-level C/C++ interface. Although the resulting development path becomes a longer one using this approach, it’s a more familiar path for EDA developers and therefore it’s a path that can ultimately lead your EDA development team to success. An elaborate study of comparing these frameworks has been published in the book Machine Intelligence in Design Automation.Integration into Legacy SystemsWhen you understand the value that you expect machine intelligence to add to your new EDA tool, when you’ve cleansed and then analyzed the data set, and when you have selected an appropriate set of development tools, you’re finally ready to add machine intelligence to your EDA development. There are two use models for AI-enhanced EDA tools. The first uses a trained model to guide the EDA tool’s decision-making. In this use case, the trained neural network doesn’t change. The software’s accuracy doesn’t improve with use unless the company that developed the EDA tool retrains the underlying neural network. This use case follows the familiar, existing use case associated with EDA tools developed using deterministic algorithms.For the second use case, the end user is able to retrain the underlying neural network, which allows the EDA tool to produce better, more accurate results over time. This use case produces a win/win situation because end users are able to hone their tools and improve them over time, without help from the EDA tool vendor’s application engineers. If the retrained models are also sent back to the EDA developer for incorporation into newer versions of the tool, all users benefit from other users’ training data.It’s not clear how you’d support this second use case in the current EDA business environment where most data sets are proprietary and are carefully guarded. Most large EDA tool customers want to keep their data in house under tight control. Even with this somewhat restrictive situation, however, EDA tools benefit from the incorporation of machine intelligence because each EDA tool customer can customize the tool and improve its results.Machine intelligence has much to add to EDA tools’ capabilities. Only time will tell if the customers want and will accept these new capabilities. Rohit Sharma, founder and CEO of Paripath Inc., is an engineer, author and entrepreneur. He has published many papers in international conferences and journals. He has contributed to electronic design automation domain for over 20 years learning, improvising and designing solutions. He is passionate about many technical topics including machine learning, analysis, characterization, and modeling. It led him to architect guna - an advanced characterization software for modern nodes. Sharma has written a book titled “Machine Intelligence for Design Automation.” You can download code examples and other information here.Note from SEMI-ESD Alliance: ESD Alliance’s Interoperability Committee brings together the industry to discuss interoperability. By focusing the efforts of the electronic system design community onto key compute operating systems, the Interoperability Committee seeks to define a stable, interoperable environment for tools and streamline the resources required to support these environments. The EDA Industry OS Roadmap presents guidelines to EDA vendors and customers for compute platforms to target for design starts. Learn more and view the OS Roadmap overview at our website.