Perspectives on Various Issues of Interest

Aiding Surgical Outcomes with AI and Machine Learning

At its 2022 annual meeting, the Surgical Outcomes Club—a leading consortium of surgeons and health services researchers dedicated to advancing surgical outcomes science—hosted a panel of four experts to discuss the growing role of predictive analytics and artificial intelligence (AI) in surgical research. The discussion centered on three core domains where AI is poised to make a significant impact: computer vision, digital transformation at the point of care, and the utilization of electronic health records (EHR) data. The panel addressed both the opportunities and inherent challenges associated with integrating AI into surgical practice.

Computer Vision: Giving Machines Eyes in the OR

The increasing capture of surgical video—routinely generated during minimally invasive and robotic procedures, and now expanding into open surgeries—offers a new frontier for AI. Real-time video annotation powered by computer vision can help evaluate surgical performance, identify complex anatomy, and provide intraoperative feedback to mitigate technical errors. Beyond performance assessment, this technology holds promise for surgical education, allowing skills training and behavior review through tool and hand tracking, and phase annotation.

With the advent of convolutional neural networks and other advanced models, video-based AI tools can now approach the visual complexity of surgery. These innovations can enhance surgeon training and potentially assist with decision-making during procedures. However, real-world implementation remains limited by several barriers, including the complexity of surgical environments, insufficient generalizability of current models, and a lack of large, annotated, and diverse datasets. Data sharing limitations and institutional barriers further complicate the creation of robust open-source datasets.

While current efforts rely on public video sources of inconsistent quality, a recent consensus suggests retrospective training tools may be feasible within two years, and real-time applications may emerge within the next decade. Recognizing the early-stage nature of this field is critical to fostering collaboration between surgeons and engineers, ensuring that AI tools ultimately support, rather than replace, surgical expertise.

Building Surgical Intelligence Through Video-Based Analytics

Video analytics offer the potential to assist surgeons during operations by identifying key anatomical landmarks, outlining tumor margins, or analyzing instrument usage patterns. Particularly in rare or unexpected events—such as intraoperative bleeding—AI could provide scenario-based recommendations to guide next steps. Though not yet deployed in operating rooms, emerging research outlines how real-time decision support systems could soon become a reality.

Leveraging Data for Surgical Innovation

Despite the influx of new surgical devices, there's limited insight into how they compare to existing techniques. Video analysis can quantify how new technologies influence surgical workflow and learning curves, particularly during the adoption of minimally invasive or robotic procedures. Side-by-side comparisons of similar cases can inform best practices and highlight improvements or setbacks introduced by novel tools.

Surgical video provides “ground truth” data, offering unparalleled insights into intraoperative behavior. Beyond enhancing individual performance, these data can serve broader purposes—from reducing OR inefficiencies to defending medical decisions in legal contexts, and informing medical device development. However, building the necessary infrastructure for data processing and analytics requires technical expertise far beyond traditional clinical training. Successful implementation hinges on interdisciplinary collaboration between clinicians and data scientists.

Addressing the Complexities of Surgical Video Analysis

While promising, the use of surgical video raises significant legal, ethical, and logistical questions. Ownership of the footage remains unclear, and concerns about patient privacy, staff exposure, and potential misuse can discourage open sharing—especially in complex or unfavorable cases. Nonetheless, video evidence can also serve as proof of adherence to standard care protocols.

Other concerns include potential conflicts of interest, data security, and a disconnect between those who generate the data (surgeons) and those capable of analyzing it (engineers). Despite these challenges, carefully selected use cases and clear goals can help bridge these divides.

Calls to Action: Accelerating AI Integration in Neurosurgery

The increasing public awareness of AI offers a strategic moment to integrate it meaningfully into neurosurgical practice. Four key actions are proposed:

1. Establish AI Task Forces: Professional societies such as the American Association of Neurological Surgeons (AANS) and the Congress of Neurological Surgeons (CNS) should form joint task forces to define best practices, set data standards, and facilitate clinician-scientist collaboration. Subspecialty task forces should address domain-specific use cases, backed by dedicated research funding and aligned with interdisciplinary partners across related surgical and technical fields.

2. Create Multi-Institutional Research Organizations: Single-institution efforts lack the scale and diversity needed to train robust AI models. Instead, we should foster independent, multi-institutional research entities—either nonprofit or for-profit—that can secure funding, manage cross-institutional data, and develop reusable tools for ML integration.

3. Launch Conferences and Challenge Frameworks: There is a need for clinician-led conferences and grand challenges to define and advance AI use cases in surgery. Inspired by the “Common Task Framework” (CTF) model, such initiatives can attract diverse collaborators and reward clinically meaningful innovation. Dedicated tracks within surgical and technical conferences can help bridge the divide between these communities.

4. Standardize Data Capture and Sharing: Video is just one of many valuable data streams in the modern OR. Integrating and standardizing these streams for AI use remains a challenge due to technical and regulatory hurdles. Collaborative efforts between surgical and technical communities, supported by new regulatory frameworks, can unlock the potential of OR data for clinical improvement.

By aligning clinical insight with technical innovation, the surgical community can unlock the transformative potential of AI. Through multidisciplinary efforts, structured collaboration, and a shared vision, we can bring next-generation surgical analytics from concept to clinical reality—benefiting patients, providers, and the entire healthcare ecosystem.

As an example, a study developed an AI-ready dataset for model training by programmatically querying open surgical procedures on YouTube, selecting and manually annotating a subset of videos. This dataset was used to train a multitask AI model, subsequently applied in two proof-of-concept studies: (1) to generate “surgical signatures” that characterize procedural patterns, and (2) to identify hand motion kinematics indicative of surgeon experience and skill level.

The resulting Annotated Videos of Open Surgery (AVOS) dataset comprises 1,997 videos spanning 23 procedure types, sourced from 50 countries over a 15-year period. To test real-world applicability, additional deidentified surgical videos were prospectively collected from a tertiary academic medical center (Beth Israel Deaconess Medical Center [BIDMC]), with IRB approval and patient consent.

Multitask Model Architecture and Training

A multitask neural network was trained on the AVOS dataset to perform spatiotemporal analysis of hands, tools, and actions in surgical video. The model captured procedural flow and fine motor behaviors in near real time, enabling simultaneous analysis across multiple tasks. To improve generalizability across varied operative conditions, data augmentation techniques—including flipping, scaling, rotation, and occlusion testing—were applied during training. An alternating task training strategy was used to optimize both spatial and temporal branches, with a dedicated training stream for hand-pose estimation.

Inference was performed by extracting batches of four frames at five-second intervals from each video. Background actions were filtered to ensure consistent comparisons across procedures.

Proof-of-Concept: Generating Surgical Signatures

The model was tested on previously unseen videos of appendectomies, pilonidal cystectomies, and thyroidectomies—procedures well-represented in the AVOS dataset. These videos were manually reviewed to confirm the presence of key operative steps, with durations ranging from 2 to 30 minutes. Using temporal averaging of model outputs, distinct surgical signatures were generated for each procedure, reflecting expected progressions in tool use and action (e.g., from cutting to suturing).

These signatures serve as procedural benchmarks, and significant deviations from them may reflect disruptions in surgical flow, variations in technique, or complexity in a given case. This functionality offers the potential for early detection of surgical anomalies or challenges requiring expert intervention.

Proof-of-Concept: Quantifying Surgical Skill

To assess skill, the model was retrospectively applied to 101 prospectively collected surgical videos at BIDMC, including live procedures and simulated wound closures. Participants included 14 operators categorized as either trainees (medical students, residents) or experienced surgeons (fellows, attendings). Hand movements were tracked using bounding boxes and nine anatomical key points (thumb, index finger, and palm).

Kinematic metrics—including velocity, rotation, and translation—were extracted and summarized into a single compound skill score using principal component analysis. Logistic regression analysis showed this compound feature significantly predicted surgeon experience, with each unit increase associated with a 3.6-fold increase in odds of being an experienced surgeon (95% CI: 1.67–7.62; p = 0.001).

Implications for Surgical Education and AI-Augmented Assessment

The multitask model demonstrated procedure-agnostic capabilities, performing reliably across variable video conditions such as lighting and camera angles. The ability to analyze both procedural flow and individual surgeon behavior marks a major advance toward automated, objective surgical feedback.

By linking hand motion patterns to surgical expertise, the model offers actionable insights for training. For instance, AI-driven feedback on motion economy and steadiness could allow trainees to iteratively improve performance, aligning with best practices observed in expert surgeons. This scalable, unbiased approach to skill assessment may facilitate faster and more reliable surgical training, especially in simulation-based environments.

Sources:

https://doi.org/10.1001/jamasurg.2022.5444

https://doi.org/10.3171/2023.4.JNS23640

https://doi:10.1001/jamasurg.2023.6262

Unsupervised NLP Methods for Analyzing User-Provided Data in Smart Product Service Systems

Product-Service Systems (PSS) are business models that go beyond delivering physical products by also offering complementary intangible services. With the rise of smart technologies and connected devices, these models have evolved into Smart Product-Service Systems (Smart PSS), enabling service providers to deliver personalized and data-driven offerings. By leveraging user-generated data from smart products, providers can tailor services to meet individual customer needs more effectively.

To harness the potential of smart devices in enhancing existing services or developing new ones, a structured Smart PSS design methodology is essential. Such a method enables enterprises and service providers to build personalized Smart PSS solutions using user-generated data, or to enhance current offerings through the application of deep learning techniques.

As a first step, service providers should adopt the customer's perspective to identify the root causes of any inconvenience experienced during service use. To support this, the use of a customer journey map is recommended as a tool to analyze and understand the customer’s mental and emotional experience throughout their interaction with a product or service.

A customer journey map is a visual method that outlines the steps a customer takes, combined with their subjective feelings at each stage. By illustrating the entire service experience through graphics and process flows, it helps organizations see their services through the eyes of the customer. This perspective enables companies to pinpoint weaknesses in the service process and use these insights as a foundation for improvement.

Furthermore, as many customer interactions today are closely tied to data collection, traditional customer journey maps may need to be adapted to incorporate data-driven insights that better align with the evolving needs of modern service providers. A modified customer journey map is needed to assist service providers in analyzing the customer experience, as illustrated in the following figure.

Two key enhancements include: the inclusion of “needed data” at each stage of the journey, and the addition of emotional states alongside their root causes, forming the basis of an emotional journey.

This modified approach is explained step by step in the following sections:

1. Service providers should place themselves in the role of a customer and break down the service process into distinct stages, reflecting how the customer experiences the journey.

2. Identify the key activities and touchpoints customers engage with at each stage. These are critical interactions that shape the overall experience and must be included in the journey map.

3. Consider the emotions customers might experience during each activity. Providers should empathize with the user to understand these feelings. Based on this, an emotional journey can be drawn to represent the fluctuations in customer sentiment. In this version, both the emotional state and its underlying causes are mapped, providing deeper insight into customer behavior.

4. Recognize the data customers rely on to complete key activities. This “needed data” should be clearly identified to help service providers understand what information supports each step of the journey.

5. Finally, service providers should reflect on potential problems or pain points within the process and explore corresponding solutions to enhance the customer experience.

The modified customer journey map is also applied to uncover user requirements across various industries, recognizing that different sectors may face unique service-related pain points. However, certain factors—such as system usability, user interface (UI) design, and pricing—remain consistently influential in shaping customer satisfaction across industries.

Therefore, when designing new services, providers must consider both industry-specific challenges and universal customer expectations.

In summary, the modified customer journey map allows service providers to gain a comprehensive understanding of the user experience, including the emotions associated with each step of the service process. By capturing these emotional insights, providers can more easily identify potential issues and opportunities for improvement, enabling them to make targeted adjustments that better align with customer needs and expectations.

After identifying user requirements, service providers must determine which pain points they aim to address or how to adapt existing services to better fulfill those needs. Once the specific challenges are defined, the next step is to gather the necessary data to support service redesign or improvement.

Data can be sourced from a variety of channels, including social media posts and replies, instant customer feedback, and open datasets provided by governments or academic institutions. Thanks to the widespread availability of online data, much of this information is easily accessible, well-organized, and often available at low cost.

However, publicly available data may not always provide the depth or specificity required for targeted service enhancements. To overcome this limitation, service providers are encouraged to collect their own data tailored to their objectives. For instance, one might scrape restaurant reviews from platforms like Google Maps to gain insights into customer sentiment and expectations.

The benefit of collecting data independently is that it can be customized to meet specific demands. However, this approach can also be time-consuming and resource-intensive, requiring careful consideration of the trade-offs between data quality and operational cost.

To effectively meet diverse customer needs, service providers must develop appropriate models that align with the data they’ve collected. Before selecting a suitable model, it is essential to clearly define the target problem and identify the type of data to be used. However, traditional machine learning models often struggle with unstructured data—such as text, images, and audio—due to their lack of inherent structure.

Focusing on text data one could demonstrate the advantages of integrating smart Product-Service Systems (Smart PSS) with deep learning techniques. Specifically, we utilize the Doc2Vec (Document-to-Vector) model, a deep learning method that extends the Word2Vec approach, and implement it using the Gensim® library.

Doc2Vec is an unsupervised learning algorithm designed to convert entire texts into fixed-length vector representations. It learns these representations by predicting words based on their surrounding context, leveraging the idea that the contextual relationships in a sentence can reveal its semantic structure. During training, sections of sentences are masked (hollowed out), and the model learns to predict these missing words using the surrounding context—without requiring labeled data.

Additionally, Doc2Vec incorporates a paragraph matrix that serves as a unique identifier for each document, capturing global information and maintaining coherence across different parts of the same text. This enables the model to retain not just statistical, but also semantic and contextual meaning, even across documents of varying lengths.

Compared to traditional techniques like TF-IDF, which rely heavily on frequency counts and disregard word order or meaning, Doc2Vec provides a richer representation of text. Unlike Word2Vec, which computes document vectors as averages of individual word vectors, Doc2Vec takes into account the structure and flow of entire paragraphs, making it more effective in capturing the topic or theme of a document.

Once trained, the resulting document vectors can be used for various tasks such as document classification, clustering, and similarity comparison. A commonly used technique for comparing these vectors is Cosine Similarity, which measures the angle between two non-zero vectors, offering a metric for semantic similarity between texts.

The final step focuses on encouraging service providers to develop a complete and practical solution tailored to customer needs. These solutions can take various forms—web platforms, mobile applications (such as Android apps), and other digital interfaces are commonly used to deliver services effectively.

For instance, the figure below illustrates the workflow of a smart Product-Service System (Smart PSS) designed for tourist recommendations, implemented by a network of independent taxi operators.

The system employs an Android application as the front-end interface for the proposed Smart PSS, while the back-end is powered by a Python server integrating a Doc2Vec model. The interaction process unfolds as follows:

1. In the first step (arrow 1), users input three types of information via the app: their current location, preferred travel distance, and a text description of personal preferences related to attractions and restaurants.

2. This information is then transmitted to a PHP and Python server (arrows 2, 3, and 4), where it is used as parameters to compute similarities with surrounding attraction data (arrow 5).

3. The top 5 most relevant results, based on similarity calculations, are sent back to the app (via arrows 6 and 7), and finally, the results are displayed to the user (arrow 8).

To enhance user experience, the recommendation system is also integrated with the Google Directions API to offer navigation services. It generates a personalized full-day itinerary, including suggestions for morning and afternoon attractions, as well as lunch and dinner spots.

The detailed workflow is as follows:

• Step 1: Users define a preferred travel radius and select types of attractions they are interested in.

• Step 2: They enter a brief text describing their preferences (e.g., “quiet places with scenic views”).

• Step 3: If users choose to skip planning for any part of the day (morning, lunch, afternoon, or dinner), they can do so.

• Step 4: The application uses the user’s current location to set the search radius.

• Step 5: For each attraction within that radius, the system retrieves the top five user reviews from Google Maps, then converts both the user’s input and the review texts into 50-dimensional vectors using the Doc2Vec model.

• Step 6: Using cosine similarity, the system compares the user's preferences to each attraction's reviews and ranks the nearby options accordingly.

Each recommended attraction includes details such as the name, user rating, estimated travel time, and distance.

From the ranked list, the system presents the top 10 recommended attractions, allowing users to make selections for each time period. Once selections are made, a custom day plan is generated and route-planned using the Google Maps API, which defaults to the shortest driving route. To support diverse travel preferences, the app also includes a transportation mode selector, enabling users to switch between walking, public transport, biking, or driving.

Source: https://doi.org/10.1016/j.compind.2021.103421

 

Machine Learning Based Industry 4.0 Framework for Composite Autoclave Manufacturing

Artificial intelligence (AI), machine learning (ML), and data-driven techniques can be used to support and optimize the manufacturing of composite materials. The full model development lifecycle encompasses analysis, optimization, inverse problem-solving, and experimental validation. The Industry 4.0-inspired modeling pipeline tailored to composite processing combines both data-centric and model-centric knowledge engineering practices, with an emphasis on embedding domain expertise directly into the dataset curation and model development phases.

A high-level Industry 4.0 (I4.0) framework is composed of three interrelated sub-systems:

(a) Data acquisition from cyber-physical systems, which may include a combination of real-time production data and simulated outputs;
(b) A data pipeline that manages key preprocessing tasks such as data cleaning, dimensionality reduction, storage, and other logistical operations necessary to ensure data usability;
(c) A model development pipeline, which involves filtering incoming data, constructing relevant knowledge datasets, building and validating predictive models of the manufacturing system, and ultimately deploying these models to support real-time decision-making on the shop floor.

Focusing on the sub-system (c), the proposed framework builds upon the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology, a widely adopted and open-standard model for data analytics. It provides a structured, iterative approach to data-driven problem-solving, making it especially suited for dynamic, evolving manufacturing environments.

A key element in data-driven modeling is the anomaly detection system, which ensures that incoming data aligns with the distribution of previously collected data. This validation step can be implemented using a variety of techniques, such as classification, clustering, deep learning, or statistical models. If the chosen method determines that a new data point fits within the existing distribution, it is accepted and incorporated into the dataset; otherwise, it is rejected. To enhance this process, domain expertise can be integrated through a dataset knowledgeability exercise, allowing expert judgment to refine the dataset and establish a feedback loop. This hybrid approach contrasts with fully automated, knowledge-agnostic systems and is seen as a vital component in advancing toward Industry 5.0. By enforcing consistency in data distribution, the assumption of independently and identically distributed (IID) data is upheld—an essential condition for making reliable comparisons between models trained on distinct datasets.

USE CASE:

Composite materials have seen broad adoption across industries such as aerospace, automotive, and construction, thanks to their exceptional structural properties—namely, high stiffness-to-weight and strength-to-weight ratios—along with reduced maintenance needs and lower lifecycle costs. Despite these benefits, the production of composite components remains challenged by significant levels of uncertainty, both aleatoric (inherent variability) and epistemic (lack of knowledge), which hinder the consistent manufacture of high-quality parts.

In the aerospace sector, where structural reliability is paramount due to the potentially catastrophic consequences of failure, stringent qualification frameworks exist to ensure the integrity of materials, processes, and designs. While necessary for safety, these frameworks place considerable demands on manufacturers and often limit flexibility in production—particularly when it comes to cost optimization and real-time decision-making on the factory floor (Crawford et al., 2021a).

One example of these challenges can be seen in the use of “bus stop” autoclave cure cycles, where multiple parts, tools, and materials are batched together and cured simultaneously. This method offers practical advantages—such as reduced floor space requirements, more efficient use of capital equipment, and lower overhead—but also increases the complexity of planning and process control. In such settings, shop floor engineers and operators are often required to make in-situ decisions, relying on their expertise and tacit knowledge to maintain process conformance and part quality.

However, these decisions are typically unstructured and lack systematic optimization, revealing an opportunity for the integration of intelligent, technology-enabled decision-support systems. The success of such systems depends not only on access to historical process data, but also on the incorporation of expert insights—bridging human experience with data-driven methods to improve repeatability and efficiency in composite manufacturing.

In bus-stop autoclave curing runs for manufacturing composite aerospace structures, multiple small components—each with similar physical characteristics and qualified to undergo the same cure cycle—are stacked together in a single autoclave to improve production efficiency. Despite the shared cure cycle, each part must maintain a thermal history that stays within an acceptable thermal envelope to ensure product quality.

To achieve a successful cure across all components, parts are carefully selected based on physical similarities, such as laminate thickness, construction type (e.g., monolithic vs. sandwich panels), tooling material, and other relevant attributes. Two critical features derived from each part’s thermal profile—the peak exotherm (the maximum temperature reached during the cure) and the steady-state lag (the highest temperature differential between the part and the surrounding autoclave gas)—are used as indicators of process quality.

Lower values of these thermal metrics generally signify reduced process variability, which correlates with higher-quality outcomes. If these thresholds are exceeded, parts may develop defects like voids, ultimately compromising mechanical properties such as flexural strength, flexural modulus, and interlaminar shear strength. Thus, peak exotherm and steady-state lag serve as essential acceptance criteria for screening cured parts.

In the context of bus-stop autoclave cure cycles, interpretable surrogate models such as Logistic Rule Regression are integrated with expert knowledge through a fuzzy scoring system, enabling an assessment of dataset “knowledgeability” prior to the deployment of black-box models. To further support model validation, two metrics are introduced: specificity as a global confidence indicator, and the novel Decision Boundary Crispness Score (DBSC) as a local, sensitivity-based metric.

The modeling task at the factory level is framed as a binary classification problem, where the objective is to predict whether a carbon fiber prepreg part passes or fails the quality standards following autoclave curing. Specifically, the classification is based on two critical thermal processing criteria:

1. Pass if the peak exotherm temperature is less than 5 °C (Criterion 1)

2. Pass if the maximum lag temperature is less than 20 °C (Criterion 2)

Parts that meet both criteria are labeled as "pass" (class 1), while those that exceed either threshold are labeled as "fail" (class 0).

The model architecture, illustrated in the following figure, consists of two independent predictive models—one for each thermal outcome. This design allows for separate evaluation and parameter tuning tailored to each specific target, enabling more precise learning for each thermal metric.

Both models share an identical architecture, comprising five hidden layers with seven neurons per layer. This configuration was chosen through a parametric study, where it demonstrated the lowest error rate on the test dataset. Each neuron uses a sigmoid activation function, and model training is performed using the Adam optimizer with a binary cross-entropy loss function.

The results demonstrate that DBSC offers a more nuanced and conservative evaluation of both dataset and model quality, especially useful when dealing with uncertain or variable manufacturing conditions. The enhanced explainability and localized insight provided by these methods are particularly valuable to production engineers, supporting trust and accountability when using black-box models in high-stakes, real-time decision-making scenarios.

Source: https://doi.org/10.1016/j.compind.2021.103510

Augmented Reality for Training Applications in Small Enterprises

The fragmented and technically demanding nature of augmented reality (AR) content development continues to limit its broader adoption in industrial settings. Current AR authoring methods often require specialized knowledge in areas such as 3D modeling, programming, computer vision, tracking, and rendering. While automated tools attempt to streamline content creation, they are typically rigid and unsuitable for dynamic or undefined processes.

Many AR development approaches depend on existing resources like CAD models or PDF manuals. However, in cases where these materials are outdated, incomplete, or unavailable—such as with legacy equipment or obsolete electronics—content creation becomes a significant bottleneck.

Although model and symbol libraries are emerging to address this, they require constant updates and rely on fiducial markers for registration. This introduces further challenges, including the need for infrastructure changes (e.g., marker placement and visibility), time-consuming setup, and ongoing maintenance. Natural feature recognition can eliminate the need for markers but instead demands complete, high-quality 3D models and clear visibility of all object features, which is often impractical in industrial environments. Furthermore, while depth sensors and image capture can help generate 3D models without CAD data, these methods may interrupt workflows, require invasive equipment, and struggle to capture small or occluded components.

To address these challenges, AR content creation methods must evolve to accommodate non-experts, operate without dependencies on pre-existing resources or infrastructure modifications, and adapt to changes in processes or tasks. The proposed method introduces a low-disruption, template-based approach that bridges traditional and AR interfaces by capturing expert knowledge—both explicit and tacit—through eye-tracking and structured information mapping.

A Multi-Modal Framework for AR Content Creation and Delivery

The framework supports two user roles:

• Content creators (e.g., trainers or experts), and

• Content consumers (e.g., trainees or task novices).

It comprises five main steps:

1. Business Need Identification – Pinpoint a requirement for efficient knowledge transfer between experienced and inexperienced staff, particularly for hands-busy tasks.

2. Task Recording via Eye Tracking – The expert performs the task while wearing eye-tracking glasses, generating visual and audio data that are then broken down into clear training steps. These are populated into a Unity 3D-based AR content template.

3. Iterative Refinement – Content is refined through usability testing and feedback loops.

4. User Feedback Collection – Feedback during real-world use is captured to guide further optimization.

5. Knowledge Preservation and Empowerment – Organizations can retain critical knowledge through archiving and moderation, while users gain autonomy through self-paced learning.

(a) Task Recording & Information Mapping: The initial stage involves capturing the training task using eye-tracking technology. The recorded content is then structured through information mapping to define clear instructional steps. (b) Unity Development Environment: Unity is used as the core development platform, providing a visual interface for organizing and integrating multimedia elements into the AR training experience. (c) Augmented Repair Training Application Template with Mixed Reality Toolkit: The template, built on Unity and enhanced with the Mixed Reality Toolkit, offers a pre-configured structure for rapidly building AR training applications with intuitive interaction components. (d) Content Import: Trainers upload task-related media—such as video clips and images—into the template to visually support each instructional step. (e) Step Customization: Trainers define and organize the number of training steps, and for each one, they add supporting images, descriptive text, and other guidance to enhance user comprehension.

The training content (video, audio, gaze data) is captured directly during task execution. Eye-tracking glasses combine a front-facing video feed with a gaze fixation marker and integrated audio. This method enables the transfer of explicit procedural steps and tacit expertise, such as visual attention and workflow patterns. The resulting media is edited using standard software (e.g., Apple iMovie) and mapped into structured instructional steps.

Design of the AR Interface: Augmented Repair Training Application 

The template organizes content using user-centric design principles. A numbered step menu allows users to track progress and navigate between instructions. Controls such as “home,” “back,” and “next” replicate traditional interface metaphors, easing the transition for new AR users. To enhance clarity, content is displayed with high-contrast text (white font with blue accents).

Instructional content is structured using information mapping, a method that includes:

• Chunking related steps and concepts

• Highlighting essential information

• Ensuring consistency in formatting and structure

• Integrating visuals effectively

• Presenting content hierarchically for intuitive learning

Verification of Learning and Task Completion

Verification of knowledge transfer can be achieved through interactive assessments or expert review. While quizzes may suit well-defined tasks with quantifiable outcomes, they lack flexibility for variable, real-world processes. For this application, expert review was chosen for its simplicity and alignment with the goal of maintaining learner autonomy while ensuring training effectiveness.

USE CASE: 

The AR training assistance method was implemented to create shop floor training content for a small enterprise. While participants expressed enthusiasm about the innovative approach, many were unfamiliar with the tools involved—specifically Unity3D, eye-tracking software, and the AR device itself. This unfamiliarity led to several usability challenges during the content editing and deployment stages. Technicians required support to navigate unexpected issues such as accidentally deleted scripts or misplaced files.

Despite these challenges, subjective feedback from stakeholders was highly positive. Senior management and the sales team viewed the AR training tool as a valuable and forward-thinking investment. When showcased at trade events and to existing clients, the tool generated strong interest and favorable responses. Shop floor staff were also enthusiastic about the interface, interpreting it as a sign that their company was embracing modern technologies and investing in employee development. However, a small number of staff expressed discomfort using certain features of the AR device, such as voice commands and air tap gestures.

In response to the challenges identified during deployment, it became evident that a more accessible content creation interface would improve usability and reduce technical barriers. To this end, the development of an "intermediary interface" is proposed for future work. This interface would offer a form-based editing environment, simplifying the interaction between non-expert users and the underlying AR content formatting system.

Key features of the proposed intermediary interface include:

• Form-style input fields for guided data entry (e.g., text, video, audio annotations)

• Built-in tools to edit and align eye-tracking outputs into AR-friendly formats

• Tooltips and help documentation embedded within each section to guide users step-by-step

• Automated conversion of form inputs into AR code via a secure back-end service

• Responsive layout using a bootstrap-style grid system to accommodate various AR form factors

The Integrated Development Environment (IDE) would be decoupled from the front end and managed in the back end, ensuring secure storage, script execution, and formatting integrity. Additional layers of security and content moderation can be introduced to prevent accidental corruption, unauthorized access, or unapproved deployments.

This redesigned workflow aims to empower non-technical users to contribute to AR content creation while preserving the integrity, flexibility, and scalability of the training system.

Source: https://doi.org/10.1016/j.compind.2020.103208

Demand Forecasting with Deep Reinforcement Learning

Case: Implementing deep reinforcement learning based demand forecasting for a leading global semiconductor components distributor who has to deal with a huge number of short life cycle products with very different demand patterns and potential perishability concerns. The organization serves as a franchise partner for hundreds of suppliers worldwide. The distributor equipped with a strong service team provides a range of value-added services such as product information, demand generation, turnkey solution, technical support, after-sales services, warehousing, logistics, e-commerce, and credit to different customers including OEM (Original Equipment Manufacturer), ODM (Original Design Manufacturer), EMS (Electronic Manufacturing Services), and SME (Small and Medium Enterprises). With orders from customers all over the world, the distributor must satisfy thousands of product demands with local flexibility.

Demand forecasting plays an important role for smart production in semiconductor supply chains due to long lead-time for capacity expansion, shortening product life cycles and pervasive uncertainty.

Some of the traditional forecasting techniques have limitations such as lack of data for estimating parameters for employing multivariate models (it is also very time-consuming when there are many products involved) and aggregating demand at monthly or quarterly levels reduces supply chain agility needed for smart production. Reinforcement learning (RL) enables agents to learn to select an action that maximizes the reward from the interaction with environment. High dimensional state of most real-world problems make it challenging for RL to derive efficient representation of the environment. Additionally the use of completely raw inputs for learning optimal control policies create additional limitations in terms of memory complexity and computational complexity. With increasing computing power being widely available, neural network and deep learning are providing effective solutions to overcome these obstacles. Deep Reinforcement Learning (DRL) is a combinational agent of deep learning and the convolutional reinforcement learning that can learn to optimize a sequential decision making under demand uncertainty through the environment state. 

It is important to design a forecasting selection mechanism that may integrate a number of forecasting approaches. The selection of forecasting model has limitations. In particular, rule-based approach may have the problem of overfitting or insufficient definition. Feature-based selection is highly dependent on the feature formulation and selection step that is time-consuming. Since the product quantities and types for a semiconductor distributor are quite large, it is not suitable to apply the rule-based method due to the difficulty to define all the required rules. In addition, as demand fluctuation is high in the semiconductor industry, the feature formulation will be complex or insufficient due to diverse features and thus affect the capability for further forecasting. For the ensemble model, since the weight assigned to each predictor will not be updated in each step, it may have limitations for forecasting the products with rapid changing time-series patterns such as high-tech products with shortening product life cycles. The DRL model provides the advantage that it can select the forecasting model automatically and update the chosen model dynamically through the interaction with the environment.

The implementation involves six phases - (1) defining the problem scope; (2) identifying gaps in decision quality; (3) establishing objectives, considering alternatives, and identifying uncertainty Influences; (4) assessing and outlining likely outcomes; (5) conducting holistic evaluations and value assessments; and (6) assessing trade-offs between attributes and making a choice.

Defining the Problem Scope

The framework begins with defining and structuring the problem at hand. Both the decision maker and the context are central to this process. For this case, the decision content revolves around estimating future demand, while the decision maker is the demand-planning manager. This individual responds to demand forecasting for a semiconductor distributor, relying on both past experience and customer forecasts.

To meet the global demand of customers, a semiconductor distributor must place timely and accurate orders with suppliers based on these demand forecasts. However, challenges arise from the semiconductor industry's relatively short product life cycle, long lead times, and the high volatility in supply and demand. These factors make precise forecasting particularly difficult. Overestimating demand often leads to excess inventory and high storage costs, while underestimating demand increases the risk of stockouts. Therefore, the decision maker requires a reliable, efficient, and accurate method to forecast future demand and mitigate unnecessary inventory costs.

Identifying Gaps in Decision Quality

Semiconductor components are known for having diverse demand patterns, some of which exhibit significant fluctuations due to the short product life cycles inherent in the industry. Many companies lack advanced data acquisition systems that can collect sufficient data for comprehensive, multifactorial analyses. As a result, the focus is on improving decision quality using univariate data.

Since future demand for various semiconductor products must be forecasted, and given that each forecasting method has its own strengths and limitations, it is essential to automatically select the appropriate forecasting model for each product. By combining analytical insights into demand with expert knowledge from domain specialists, companies can make more informed and intelligent decisions within a smart supply chain. Furthermore, due to the rapid changes in time series data linked to the short product life cycle, choosing the right forecasting model at different time intervals becomes a critical challenge.

Establishing Objectives, Considering Alternatives, and Identifying Uncertainty Influences

After defining the problem and identifying key areas for improvement, relevant demand data is prepared. Using the theoretical demand pattern classification rule, demand patterns are classified into four groups to enhance the forecasting model’s prediction accuracy. These categories are based on two key thresholds: inter-demand intervals (IDI) and the coefficient of demand variation. The four groups are smooth (low IDI and low CV), erratic (low IDI and high CV), intermittent (high IDI and low CV), and lumpy (high IDI and high CV).

To address demand uncertainty, this step explores the relationships between various sources of demand uncertainty, considering the unique characteristics of each forecasting model. A comparison of different models, each with its own properties, is performed and integrated into a comprehensive strategy. Specifically, traditional forecasting methods such as the Naïve forecast, simple moving average, single exponential smoothing, and the Syntetos-Boylan approximation (SBA) are evaluated. Additionally, three machine learning approaches—artificial neural network (ANN), recurrent neural network (RNN), and support vector regression (SVR)—are selected for comparison.

Assessing and Outlining Likely Outcomes

Deep Reinforcement Learning (DRL) is utilized to evaluate the potential outcomes of various actions, enabling effective learning by combining deep neural networks with reinforcement learning (RL). Given that demand forecasting is performed at a weekly granularity, Deep Q-Network (DQN), a DRL model well-suited for discrete states, is selected to establish a model selection mechanism.

Traditional Q-learning stores the discrete states in a table, but it faces limitations when it comes to extracting meaningful features from high-dimensional data, especially as the number of states grows. To overcome this issue, DQN employs stochastic gradient descent within a deep neural network to update the model, instead of storing large volumes of Q-values in a table. This approach makes DQN ideal for learning the complex and stochastic demand patterns of semiconductor products and selecting the most appropriate forecasting method. The agent’s objective is to maximize the cumulative future reward by optimizing the action-value function.

Conducting Holistic Evaluations and Value Assessments

To evaluate the overall performance of the selected methods, Mean Absolute Scaled Error (MASE) is used as the primary metric. Given the high fluctuations in the demand patterns, some periods in the time series may contain zero values. The advantage of using MASE is that it accommodates the presence of zeros in the historical demand data.

Assessing Trade-offs Between Attributes and Making a Choice

The decision maker will make the ultimate decision based on the performance evaluations from the previous stage. The proposed approach, with its self-learning function, will update periodically and select the method that minimizes forecast error. The quantified forecasting results, along with the measurement metrics, will serve as a reference for the demand-planning manager, who can then adjust the final demand forecast using their domain expertise and insights.

OVERALL FINDING

The DRL model is compared with the seven forecasting models, the ensemble model that combines the forecasting results from the seven forecasting model with the simple average, and with Support Vector Machine whose input is the same as the observed state in the proposed DRL model. The output is the selected forecasting model. The DRL model had the lowest MASE for the all demand scenarios - smooth, erratic, intermittent, and lumpy.

Source: https://doi.org/10.1080/00207543.2020.1733125

Reinforcement Learning in Operations & Supply Chain Management

Reinforcement learning (RL), a branch of machine learning, is increasingly being applied in operations and supply chain management to optimize decision-making, enhance efficiency, and improve performance. 

In Inventory Management applications, RL algorithms enable dynamic pricing, optimizing pricing strategies based on real-time demand and inventory levels, maximizing revenue while minimizing stockouts or overstock situations. Amazon employs RL algorithms to optimize pricing strategies dynamically based on factors such as demand, competitor prices, and inventory levels. This ensures that prices are adjusted in real-time to maximize revenue while maintaining competitiveness. RL models learn optimal inventory replenishment policies by considering demand forecasts, lead times, and cost constraints, ensuring optimal stock levels. Alibaba leverages RL algorithms for optimizing inventory management across its e-commerce platforms. By considering factors such as demand forecasts, lead times, and supply chain constraints, RL models help Alibaba maintain optimal inventory levels, reduce stockouts, and improve customer satisfaction.

For Warehouse Management, RL algorithms help with order fulfillment by optimizing picking routes and allocation of orders to minimize travel time and labor costs within warehouses. RL is used to optimize warehouse layout and storage configurations, reducing congestion, improving throughput, and enhancing operational efficiency. Amazon utilizes RL for optimizing warehouse operations, including inventory allocation, picking routes, and replenishment decisions. RL algorithms help minimize travel time, improve order fulfillment efficiency, and optimize resource utilization within warehouses. DHL utilizes RL for automating warehouse operations, including order picking, packing, and sorting. RL algorithms help DHL optimize warehouse layouts, design efficient workflows, and deploy robotic systems for tasks such as inventory management and order fulfillment.

In Transportation and Logistics, RL algorithms optimize delivery routes, considering factors such as traffic conditions, vehicle capacities, and delivery time windows, to minimize transportation costs and delivery times. UPS uses RL for optimizing delivery routes and schedules to minimize fuel consumption, reduce transportation costs, and improve delivery times. RL algorithms consider factors such as traffic conditions, package volumes, and delivery time windows to optimize route planning. RL models optimize fleet operations by dynamically allocating vehicles and resources based on demand fluctuations and operational constraints. UPS employs RL to optimize fleet operations, including vehicle routing, scheduling, and maintenance decisions. RL models help allocate vehicles dynamically based on demand fluctuations and operational constraints, ensuring efficient use of resources.

Production Planning and Scheduling is facilitated as RL algorithms optimize production sequences and schedules to minimize changeover times, idle capacity, and production costs while maximizing throughput and resource utilization. P&G utilizes RL for optimizing production planning and scheduling in its manufacturing facilities. RL algorithms help P&G minimize changeover times, optimize production sequences, and improve resource utilization, leading to increased throughput and operational efficiency. RL is used to optimize maintenance schedules for production equipment, considering equipment health data, production priorities, and resource availability. DHL employs RL for predictive maintenance of its logistics infrastructure, including vehicles, equipment, and facilities. RL models analyze sensor data, equipment usage patterns, and environmental conditions to predict maintenance needs, reduce downtime, and optimize asset performance.

For Supply Chain Optimization, RL models help in supplier selection since it can learn to select optimal suppliers based on criteria such as quality, reliability, and cost, optimizing supply chain performance and reducing supply chain risks. Alibaba uses RL to optimize various aspects of its supply chain, including supplier selection, order fulfillment, and logistics operations. RL algorithms help Alibaba identify optimal sourcing strategies, improve order processing efficiency, and enhance supply chain visibility and responsiveness. RL algorithms improve demand forecasting accuracy by learning from historical data, market trends, and external factors, enabling better inventory management and resource planning. P&G employs RL for improving demand forecasting accuracy by learning from historical sales data, market trends, and promotional activities. RL models help P&G better anticipate demand fluctuations, optimize inventory levels, and enhance supply chain planning. 

RL help with Resource Allocation decisions as it can be used to optimize labor allocation and scheduling in manufacturing facilities and distribution centers, ensuring adequate staffing levels to meet demand while minimizing labor costs. RL algorithms optimize energy usage in facilities by dynamically adjusting equipment settings, production schedules, and energy consumption patterns to minimize costs and environmental impact.

Risk Management applications are getting enhanced with RL since these models can learn adaptive strategies for responding to supply chain disruptions, such as supplier failures, natural disasters, or geopolitical events, minimizing the impact on operations and mitigating risks. RL algorithms assist in developing contingency plans and risk mitigation strategies by simulating various scenarios and recommending optimal responses in real-time. Ford uses RL for modeling and simulating supply chain risks and disruptions. RL algorithms help Ford identify vulnerabilities, evaluate mitigation strategies, and develop contingency plans to minimize the impact of supply chain disruptions on its operations.

Multi-Agent Systems can be modeled and simulated with RL. For instance, RL is used to model and simulate market dynamics in supply chains, where multiple agents (e.g., suppliers, manufacturers, distributors) interact and make decisions autonomously, optimizing overall system performance. Ford employs RL for optimizing supplier relationships and collaboration. RL models help Ford identify strategic suppliers, negotiate contracts, and manage supplier performance, ensuring a resilient and efficient supply chain.

Robotic Process Automation (RPA) in Sourcing

Robotic Process Automation (RPA) is revolutionizing sourcing operations by automating repetitive tasks, streamlining procurement processes, and improving efficiency. 

RPA bots are used to collect vendor information from various sources (such as emails, forms, and websites) are collected. The bots then validate the data, perform background checks, and populate the information into the sourcing system. This streamlines the vendor onboarding process, reduces manual effort, and ensures data accuracy. This provides a use case for automating vendor registration. IBM utilizes RPA to automate the supplier onboarding process. RPA bots collect vendor information from various sources, such as emails and forms, validate the data against predefined criteria, and populate it into the sourcing system. This streamlines the supplier registration process, reduces manual effort, and ensures data accuracy.

RPA aids in spend analysis by automating data extraction and analysis. RPA bots extract spending data from multiple sources, including invoices, purchase orders, and contracts. The bots categorize and analyze the data to identify spending patterns, supplier performance, and cost-saving opportunities. This allows sourcing teams to make data-driven decisions more efficiently and optimize procurement strategies. Accenture uses RPA for automating spend analysis activities. RPA bots extract spending data from multiple sources, such as invoices, purchase orders, and contracts, and categorize it based on predefined criteria. The bots then analyze the data to identify spending patterns, cost-saving opportunities, and supplier performance insights, helping Accenture optimize procurement strategies and improve sourcing decisions.

RFx Management is aided where RPA automates RFx preparation and evaluation process. RPA bots generate RFx documents (Request for Proposal, Request for Quote, etc.) by pulling information from templates and existing contracts. The bots distribute the RFx documents to selected vendors, track responses, and analyze proposals based on predefined criteria. This accelerates the RFx process, improves consistency, and facilitates vendor evaluation. Nike employs RPA for gathering market intelligence and competitor analysis in sourcing activities. RPA bots collect data from various sources, such as industry reports, news articles, and social media, and analyze it to identify trends, pricing strategies, and supplier performance benchmarks. This helps Nike make informed sourcing decisions, negotiate better deals, and stay competitive in the market.

RPA bots assist in contract management by extracting key terms and clauses from contracts, updating contract repositories, and sending alerts for renewal or expiration dates. The bots also help in contract negotiation by flagging discrepancies and ensuring compliance with regulatory requirements. This increases contract visibility, reduces risk, and enhances contract lifecycle management. Johnson & Johnson uses RPA for managing supplier data and documentation. RPA bots collect and organize supplier information, such as contracts, certifications, and contact details, from various sources. The bots validate the data, update supplier records in the sourcing system, and send alerts for expiring contracts or missing documentation. This ensures compliance with regulatory requirements, reduces risk, and enhances supplier data accuracy.

RPA helps with supplier performance monitoring where the bots collect and analyze supplier performance data from various sources, including feedback surveys, delivery reports, and quality assessments. The bots consolidate the data into performance dashboards, generate reports, and trigger alerts for underperforming suppliers. This enables proactive supplier management, fosters accountability, and drives continuous improvement. Amazon utilizes RPA for monitoring supplier performance and compliance. RPA bots track key performance indicators (KPIs), such as on-time delivery, quality metrics, and pricing accuracy, for each supplier. The bots analyze the data, generate performance reports, and flag underperforming suppliers for further review or action. This enables Amazon to maintain high standards of supplier performance and ensure customer satisfaction.

Purchasing order processing is facilitated since PO creation and approval can be automated. RPA bots automate the creation of purchase orders by extracting order details from requisition forms or ERP systems. The bots route the POs for approval based on predefined rules and escalate exceptions for manual review. Once approved, the bots transmit the POs to suppliers electronically, eliminating manual data entry and reducing processing time. Coca-Cola utilizes RPA for automating purchase order processing tasks. RPA bots extract order details from requisition forms or ERP systems, validate the information, and generate purchase orders automatically. The bots then route the orders for approval based on predefined rules and escalate exceptions for manual review. This accelerates the procurement process, reduces processing time, and improves order accuracy.

RPA bots manage the end-to-end process of onboarding and offboarding suppliers, from initial registration to termination of contracts. The bots collect necessary documentation, perform compliance checks, and update supplier records in the sourcing system. This ensures regulatory compliance, reduces administrative burden, and enhances supplier relationships. This helps in automating supplier lifecycle management. P&G employs RPA for streamlining supplier relationship management processes. RPA bots automate tasks such as supplier performance monitoring, contract management, and communication with vendors. By automating routine tasks, P&G frees up procurement professionals to focus on strategic activities, such as supplier development and negotiation, thereby enhancing supplier relationships and driving value for the organization.

RPA can help with price comparison and negotiation. RPA bots can automate price analysis and negotiation by collect pricing information from multiple sources, including catalogs, marketplaces, and historical data. The bots analyze price trends, identify cost-saving opportunities, and support negotiation strategies by providing real-time insights. This enables sourcing teams to negotiate better deals, optimize procurement costs, and achieve savings targets.

How are Manufacturing Firms Leveraging Digital Twins?

Digital twins, virtual replicas of physical assets, processes, or systems, are transforming the manufacturing industry by providing real-time insights, optimizing operations, and enabling predictive maintenance.

For predictive maintenance, manufacturing companies create digital twins of machinery and equipment, integrating sensors to collect real-time data on performance metrics such as temperature, vibration, and energy consumption. Machine learning algorithms analyze this data to predict potential equipment failures before they occur, allowing proactive maintenance to be scheduled, reducing downtime, and optimizing asset performance. For instance, Siemens utilizes digital twins to create virtual replicas of industrial equipment, such as turbines and motors. By integrating sensor data from physical assets with the digital twin models, Siemens can predict equipment failures, optimize maintenance schedules, and reduce downtime in manufacturing facilities.

Digital twins help in process optimization by simulating manufacturing processes and allowing companies to test and optimize production strategies in a virtual environment. By analyzing data from the digital twin, manufacturers can identify bottlenecks, streamline workflows, and improve efficiency. This leads to reduced cycle times, increased throughput, and higher quality output. GE employs digital twins to optimize production processes in manufacturing facilities. By creating virtual replicas of production lines and equipment, GE can simulate different scenarios, identify bottlenecks, and optimize workflows to improve productivity and efficiency. Boeing utilizes digital twins to optimize manufacturing processes for aircraft assembly. Digital twins enable Boeing to simulate assembly sequences, identify potential issues, and optimize workflow layouts to improve productivity and reduce assembly time. This enhances manufacturing efficiency and quality while reducing costs.

Digital twins also help in improving supply chain management practices by modeling the entire supply chain, from raw material sourcing to distribution channels. By simulating different scenarios, such as changes in demand, transportation delays, or supplier disruptions, companies can identify vulnerabilities and develop contingency plans. This improves supply chain resilience, reduces lead times, and enhances overall performance. Airbus employs digital twins to collaborate with suppliers and partners across its global supply chain. Digital twins help Airbus share design data, coordinate production schedules, and ensure compliance with regulatory requirements, leading to improved transparency, efficiency, and collaboration in the supply chain. Toyota employs digital twins to collaborate with suppliers and partners in its supply chain. Digital twins enable Toyota to share production plans, coordinate deliveries, and manage inventory levels in real time, leading to improved visibility, efficiency, and collaboration across the supply chain. Siemens employs digital twins to model and simulate supply chain processes, including production planning, inventory management, and logistics operations. Digital twins help Siemens optimize material flows, reduce lead times, and improve overall supply chain efficiency.

Digital twins are being used for quality control as they aid in monitoring and analyzing product quality in real time. By comparing data from physical products with their digital counterparts, manufacturers can identify defects or deviations from specifications early in the production process. This allows corrective actions to be taken promptly, reducing scrap, rework, and warranty costs. For instance, Boeing employs digital twins for quality control and inspection of aircraft components. Digital twins enable Boeing to compare digital models with physical parts, detect defects or deviations from specifications, and ensure compliance with safety and quality standards throughout the manufacturing process.

Digital twins help in product lifecycle management as it helps with tracking a product from design and development to manufacturing and end-of-life disposal. By capturing data on product usage, performance, and maintenance history, manufacturers can optimize product design, anticipate customer needs, and extend product lifespan. This leads to improved customer satisfaction and increased profitability.

Digital twins enable remote monitoring and control of manufacturing facilities, allowing operators to oversee operations from anywhere in the world. By accessing real-time data and simulation models, operators can make informed decisions, troubleshoot issues, and adjust parameters as needed to optimize performance and ensure safety.

Digital twins monitor energy consumption across manufacturing facilities, identifying opportunities to reduce waste and improve efficiency. This help in energy management. By analyzing data on equipment utilization, production schedules, and environmental conditions, companies can implement energy-saving measures such as load balancing, equipment upgrades, and process optimization, leading to cost savings and sustainability benefits.

Digital twins enable mass customization by simulating the production of custom-designed products. By capturing customer preferences and design specifications, manufacturers can create virtual replicas of individual products and simulate their production process. This allows for efficient batch size of one manufacturing, reducing lead times and costs while meeting customer demands for personalized products.

Digital twins are also helpful in asset performance management. GE utilizes digital twins for asset performance management in various industries, including aviation, healthcare, and power generation. Digital twins enable GE to monitor the health and performance of equipment in real time, identify potential issues before they occur, and optimize maintenance schedules to maximize uptime and reliability.

Digital twins can be used for product design. For instance, Airbus utilizes digital twins throughout the aircraft design and manufacturing process. Digital twins enable Airbus to simulate the performance of aircraft components, optimize designs for fuel efficiency and aerodynamics, and validate manufacturing processes before physical prototypes are built. This reduces development time, costs, and risks associated with aircraft production. 

Digital twins can also be used for facility design and layout. Toyota utilizes digital twins to design and optimize factory layouts for its automotive production facilities. Digital twins enable Toyota to simulate different layout configurations, analyze material flows, and optimize production processes to maximize efficiency and minimize waste. This helps Toyota achieve lean manufacturing principles and improve overall operational performance.

Capacity Planning Using Emerging Technologies

Capacity planning involves ensuring that a business has the resources it needs to meet current and future demand. Emerging technologies are transforming capacity planning by providing more accurate forecasts, optimizing resource allocation, and enhancing operational efficiency.

Artificial Intelligence (AI) and Machine Learning (ML) algorithms help with capacity planning by aiding more accurate demand forecasting. Organizations can analyze historical data, market trends, and external factors to predict future demand accurately. This allows companies to plan capacity based on realistic projections. Using optimization algorithms, machine Learning models optimize resource allocation, production schedules, and staffing levels by identifying patterns and predicting bottlenecks. Amazon uses AI-driven demand forecasting to optimize inventory levels and warehouse capacity, ensuring efficient order fulfillment.

Internet of Things (IoT) provide insights into current capacity and performance by enabling real-time monitoring whereby real-time data from equipment and production lines are collected. Siemens utilizes IoT sensors and predictive analytics for proactive maintenance of industrial equipment, minimizing downtime and optimizing production capacity.

Blockchain enables transparency and traceability by providing a secure, transparent ledger for tracking the movement of goods and resources across the supply chain, improving capacity planning accuracy. Smart contracts automate and enforce capacity-related agreements between suppliers and manufacturers, ensuring timely delivery and resource availability. Walmart implements blockchain for tracking produce from farm to shelf, enhancing transparency and efficiency in capacity planning.

Cloud Computing platforms offer scalable infrastructure that can be adjusted based on demand, providing flexible capacity planning. Cloud-based tools facilitate collaboration across different departments and geographies, enhancing coordination in capacity planning. Netflix uses cloud computing to dynamically scale its IT infrastructure based on user demand, ensuring seamless streaming experiences.

Advanced Analytics and Big Data tools process large datasets to uncover trends and insights that inform capacity planning decisions. Big data enables companies to simulate various scenarios and their impact on capacity, helping in strategic decision-making. Procter & Gamble (P&G) leverages advanced analytics to optimize production schedules and inventory management, ensuring efficient use of capacity.

Digital Twins create virtual replicas of physical assets and processes, allowing companies to simulate different scenarios and optimize capacity planning. This enables real-time monitoring where digital twins continuously updates the digital model with real-time data, providing a dynamic view of capacity and performance. General Electric (GE) uses digital twins to optimize the performance and capacity of their wind turbines, enhancing energy production efficiency.

Robotic Process Automation (RPA) automates repetitive tasks in capacity planning, such as data entry and report generation, freeing up human resources for more strategic activities. This helps with reducing the risk of human error in capacity planning processes, improving accuracy and efficiency. DHL employs RPA to automate warehouse operations and logistics planning, optimizing capacity and improving service delivery.

Augmented Reality (AR) and Virtual Reality (VR) provide immersive training experiences for employees, improving skills and preparedness for managing capacity-related challenges. These tools enhance the visualization of capacity data and planning scenarios, making it easier for decision-makers to understand and act on insights. Boeing uses AR for assembly line workers to improve accuracy and efficiency, optimizing production capacity.

Long Short-Term Memory (LSTM) Networks for Forecasting

Long Short-Term Memory (LSTM) networks are a type of recurrent neural network (RNN) designed to effectively learn and remember long-term dependencies in sequential data. LSTMs address the limitations of traditional RNNs, which struggle with long-term dependencies due to issues like vanishing and exploding gradients. 

Architecture of LSTM Networks

LSTMs have a more complex architecture than standard RNNs. The key component of an LSTM is its memory cell, which maintains information over long periods. Each LSTM cell contains several gates that regulate the flow of information:

1. Forget Gate (𝑓_𝑡): Decides what information to discard from the cell state.
2. Input Gate (𝑖_𝑡): Determines which new information to add to the cell state.
3. Cell State (𝐶_𝑡): Represents the memory of the network.
4. Output Gate (𝑜_𝑡): Decides what information from the cell state to output.

These gates are controlled by sigmoid and tanh activation functions, which output values between 0 and 1 to modulate the information flow.

LSTM Equations:

LSTM Cell Structure

Here's a diagram representing the structure of an LSTM cell:

 

1. Forget Gate: This gate uses the sigmoid function to decide which parts of the cell state 𝐶_𝑡−1 to forget.
2. Input Gate: It controls which values to update with new information. The sigmoid layer decides which values to update, and the tanh layer creates new candidate values ​ to be added to the cell state.
3. Cell State Update: The cell state 𝐶_𝑡 is updated by combining the old cell state (scaled by the forget gate) and the new candidate values (scaled by the input gate).
4. Output Gate: This gate determines the output of the LSTM cell. It first uses a sigmoid function to decide which parts of the cell state to output, then applies the tanh function to scale the cell state before multiplying by the output gate's result.

Applications of LSTMs

LSTMs are widely used in various fields due to their ability to capture long-term dependencies:

• Natural Language Processing (NLP): For tasks like machine translation, text generation, and sentiment analysis.
• Time Series Prediction: Used for forecasting stock prices, weather prediction, and anomaly detection.
• Speech Recognition: LSTMs help in converting speech to text by maintaining the context over long audio sequences.

In the context of operations and supply chain management, LSTM networks are particularly well-suited for time series forecasting due to their ability to learn and retain dependencies over long sequences. Several companies across various industries have successfully implemented LSTM networks for forecasting purposes. Here are some real-world examples:

1. Amazon

Use Case: Demand Forecasting and Inventory Management

Amazon employs LSTM models to predict product demand based on historical sales data, seasonal trends, and promotional activities. These predictions help optimize inventory levels across its vast network of warehouses, reducing stockouts and overstock situations. The accuracy of LSTM in capturing temporal patterns improves Amazon's ability to meet customer demand efficiently.

2. Google

Use Case: Cloud Usage Forecasting

Google Cloud uses LSTM networks to forecast resource usage and manage cloud infrastructure more efficiently. By predicting future demand for computing resources, Google can optimize server allocation and reduce operational costs. The LSTM models consider past usage patterns, workload seasonality, and other relevant factors.

3. Uber

Use Case: Ride Demand Forecasting

Uber utilizes LSTM models to forecast ride demand in different geographical areas and at various times of the day. These predictions help in dynamically adjusting pricing (surge pricing) and ensuring an adequate supply of drivers. By accurately forecasting demand, Uber improves rider experience and driver satisfaction.

4. Netflix

Use Case: Content Demand Forecasting

Netflix uses LSTM networks to predict the demand for different types of content. By analyzing viewing patterns, user preferences, and seasonal trends, Netflix can make informed decisions about content acquisition and production. This helps optimize their content library and personalize recommendations for users.

5. Siemens

Use Case: Predictive Maintenance

Siemens employs LSTM networks for predictive maintenance of industrial equipment. The models analyze sensor data from machinery to predict potential failures and maintenance needs. By forecasting equipment health and performance, Siemens can schedule maintenance proactively, reducing downtime and extending the lifespan of equipment.

6. JD.com

Use Case: Sales Forecasting

JD.com, a major Chinese e-commerce company, uses LSTM models to forecast sales for various product categories. These forecasts inform inventory management and procurement strategies, ensuring that popular items are adequately stocked while minimizing excess inventory. The models consider factors like historical sales data, promotional events, and market trends.

7. Facebook

Use Case: Infrastructure Capacity Planning

Facebook uses LSTM networks to predict server loads and optimize the allocation of computational resources across its data centers. By forecasting demand for different services, Facebook can efficiently manage its infrastructure, ensuring high availability and performance while controlling costs.

8. Walmart

Use Case: Supply Chain Optimization

Walmart employs LSTM models for various aspects of supply chain optimization, including demand forecasting, inventory replenishment, and logistics planning. The models analyze vast amounts of data, including sales history, weather patterns, and economic indicators, to improve the accuracy of forecasts and enhance supply chain efficiency.

Key Benefits of Using LSTM for Forecasting

• Capturing Long-Term Dependencies: LSTM networks excel at understanding long-term dependencies in time series data, making them ideal for forecasting tasks that require consideration of historical patterns.
• Handling Non-Linear Relationships: LSTMs can model complex, non-linear relationships in data, providing more accurate predictions compared to traditional statistical methods.
• Adaptability: LSTM models can be trained and fine-tuned to adapt to various forecasting needs across different industries and use cases.

These examples illustrate the versatility and effectiveness of LSTM networks in forecasting applications across diverse industries, enabling companies to make data-driven decisions and optimize their operations.

Handling computational complexity, data requirements, and interpretability 

To handle computational complexity, it would be necessary to optimize the model architecture. Depending on the context, simple versions of LSTMs (e.g., Vanilla LSTMs, Gated Recurrent Units) might suffice and they are less computationally intensive. One could start with few layers and then add complexity as needed. Dropout regularization can help in preventing overfitting as well as computational load during training. The models used for LSTM can be compressed using techniques such as quantization (i.e., reducing the model’s precision from floating points to fixed point representation) and pruning (i.e., removing connections in the network that are not significant).

While reduction in complexity by optimizing model architecture and model compression will be helpful, it is still going to be important to deploy efficient hardware and software (e.g., high performance computing that utilize Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs) and facilitate parallel processing. Firms can also take advantage of distributed computing frameworks (e.g., Apache Spark, TensorFlow’s distributed computing) that allows for models to be trained across multiple nodes.

For managing data requirements, steps such as normalization of the data, using synthetic data to augment training set, integrating data from external sources that might be relevant for forecasting, using batch processing and mini-batch gradient descent (specifically, to effectively deal with large datasets), and dimensionality reduction (e.g., using Principal Component Analysis to reduce the number of features) can go a long way.

To improve interpretability, one could implement attention mechanisms within the LSTM architecture to highlight which parts of the input sequence the model is focusing on during prediction or could use saliency maps to visualize the importance of different input features and time steps in the prediction process. There are also some methods such as help in understanding of each feature to LSTMs prediction (e.g., SHAP – Shapley Additive Explanation) and to explain individual predictions by approximating the LSTM model with a simpler interpretable model locally (e.g., LIME - Local Interpretable Model-agnostic Explanations). Sometimes it is also helpful to integrate multiple models, e.g., using LSTM + Decision Trees or LSTM + Linear Models so as to have clarity on the predictions.

As is always the case, proper documentation and reporting protocols can ensure that the knowledge gained by implementing these tools are preserved and enhanced as a company continues the journey of using these tools.