April 2024PDECC—Sakthivel (TATA Consulting Engineers)

HP Tagline--PDECCArtificial intelligence for engineering a better tomorrow—Part 1S. Sakthivel, TATA Consulting Engineers Ltd., Mumbai, IndiaArtificial intelligence (AI) in engineering a better tomorrow refers to the concept of utilizing AI technologies to enhance and innovate the field of engineering. This concept includes the integration of AI into various engineering disciplines to improve efficiency, accuracy and innovation. AI’s role in engineering could involve automating complex tasks, optimizing design and manufacturing processes, predicting system failures and enhancing decision-making. This approach targets a more advanced, efficient and sustainable future in engineering, ultimately contributing to greater societal and environmental outcomes. This article primarily focuses on the development of engineering drawings by using AI within engineering consultancy firms.A_HLFH_WEBCAST-FLIR-April-24Engineering diagrams play a crucial role throughout the entire engineering lifecycle, involving the early-stage process development, detailed engineering, construction, operation and decommissioning of a plant. In the chemical industry, key engineering diagrams include block flow diagrams (BFDs), process flow diagrams (PFDs), and piping and instrumentation diagrams (P&IDs). These diagrams encapsulate vital information about chemical processes, including process topology, major unit operations, control equipment and piping details. P&IDs serve as schematic representations, interpreting the interconnectedness of process equipment and instrumentation essential for process control. The successful development of P&IDs requires continuous collaboration among various engineering disciplines like process, instrumentation, piping and mechanical engineering. Today, the preparation of P&IDs remains a tedious, manual and time-consuming task. Moreover, several practical challenges are associated with storing and transmitting engineering diagram files in analog form, including:

The process of modifying and updating engineering diagrams is both costly and time-consuming. The need for modification often arises to keep the updated materials for safety or regulatory compliance reasons.

Analog engineering diagrams lack interoperability with other software applications, such as those related to the Internet of Things (IoT), digital twins, process automation, advanced processes, cloud computing and simulation software.

Extracting information from analog engineering diagrams is not readily determinable. As these diagrams are not machine-readable, they cannot be seamlessly integrated into knowledge bases. Consequently, information about processes becomes challenging to locate.¹ To address these issues, it is necessary to transition to the digitization of engineering diagrams in the chemical industry.

Today, AI is widely recognized as a disruptive technology, deeply influencing various sectors including information technology (IT), healthcare (medicine), manufacturing, business, retail, education and urban development. Its impact extends beyond transformed traditional computing approaches, as AI continues to penetrate and reshape numerous industries. In addition, engineering firms can utilize AI technology to automate routine tasks associated with drawing preparation, allowing engineers to allocate their valuable time to more critical project elements. AI assistants have the capability to analyze data, recognize patterns and produce precise drawings with minimal human intervention. This not only improves the accuracy of the drawings, but also encourages improved collaboration among engineers, architects and other stakeholders. In recent years, AI technologies have experienced significant advancements, introducing a range of capabilities that can substantially improve the engineering design process. AI can offer several advantages in the formulation of P&IDs, streamlining the process and improving overall efficiency. AI tools can bring about significant enhancements in the design and engineering sector. These improvements include automated symbol recognition, intelligent piping routing, intelligent annotation and labeling, support for design optimization, error detection and prevention, automatic tracking of revisions, integration of data and collaboration, compliance with standards and regulations, and improved data visualization. The objective of this article is to develop a drafting procedure for P&IDs using AI tools. Furthermore, the significance and relevance of AI techniques are examined in the context of next-generation process design and engineering applications. This highlights the enormous potential of incorporating AI into the creation of engineering drawings. The integration of AI assistants enables engineers to streamline workflows, mitigate manual errors and expedite project timelines.AI in engineering. Today, a variety of computer-aided design (CAD) tools and software platforms exists for drafting BFDs, PFDs and P&IDs—e.g., SmartPlant P&ID, CADWorx P&ID, SolidWorks P&ID, SmartDraw, SketchUp, CADMATIC, and Bentley OpenPlant P&ID. While these tools may not be explicitly categorized as AI tools, some of them integrate AI or advanced automation features. Furthermore, many proprietary process simulation software solutions lack support for interoperability and data exchange. The communication of flowsheet information through documents poses a barrier to adopting findable, accessible, interoperable and reusable (FAIR) practices. This constraint hampers the effective utilization of advanced data analysis and processing tools. Today, specific elements of chemical process design entail laborious and repetitive tasks, yet FAIR process data holds the promise of streamlining automated data processing. To tackle the challenges mentioned above, incorporating AI through trainable neutral networks can help. AI operates on the foundation of the neural network approach, a technique within the dominion of AI that instructs computers to analyze data in a manner inspired by the human brain. This methodology—classified as deep learning within the domain of machine-learning—employs interconnected nodes or neurons arranged in layered structures, mirroring the organization of the human brain. When it comes to using neural networks for developing engineering drawings, various types of neural networks can be employed. The choice of the neural network type depends on the specific requirements of the engineering drawing task, such as image recognition or generation, or understanding of sequential dependencies within the drawing. The following neural network types are commonly used in this context:

Convolutional neural networks (CNNs): CNNs are particularly effective in image-related tasks, making them suitable for tasks involving engineering drawings. They use convolutional layers to detect patterns and features in images, making them well-suited for tasks such as object recognition and segmentation in engineering drawings.²

Graph neural networks (GNNs): GNNs are a class of neural networks designed to operate on graph-structured data. In this context, graphs consist of nodes and edges that represent relationships between the nodes. GNNs have gained prominence for their effectiveness in various tasks involving relational data, such as social network analysis, recommendation systems and molecular chemistry.³Various variants of GNNs—including the graph convolutional network (GCN), the graph attention network (GAT) and the graph recurrent network (GRN)—have showcased groundbreaking performances across numerous deep learning tasks.⁴

Generative adversarial networks (GANs): GANs consist of a generator and a discriminator, trained simultaneously through adversarial training. In the context of engineering drawings, GANs can be used to generate realistic and high-quality drawings, aiding in the augmentation of datasets for training purposes.⁵

Recurrent neural networks (RNNs): RNNs are designed for sequential data and are often used in tasks involving time series or sequential information. In the context of engineering drawings, RNNs can be applied to capture dependencies and relationships between different elements in a drawing, which is crucial in design processes.⁶

Long short-term memory (LSTM) networks: LSTM networks are a type of RNN designed to overcome the vanishing gradient problem, allowing them to capture long-range dependencies in sequential data. They can be applied to tasks involving the understanding of the context and relationships between different components in engineering drawings.

Leveraging GNNs for the interpretation of P&IDs presents a transformative approach that involves node classification, link prediction, graph clustering and generating, image and text classification, and graph and node classification. Real-time consistency checks during the engineering phase are facilitated, effectively preventing errors and minimizing workforce costs by employing GNNs. This application not only supports the engineering process, but also improves a comprehensive understanding of machine-readable P&IDs as a spacious information model, marking a significant advancement in the efficiency and accuracy of the design and operational phases. GNNs constitute a category of neural networks designed to handle data structured in the form of graphs and serve as a class of methodologies proficient in conducting inference on data characterized by graph structures. These networks showcase adaptability by supporting prediction tasks at the node, edge and graph levels. The primary categories of GNNs cover the recurrent graph neural network, the spatial convolutional network and the spectral convolutional network.Two neural networks are used for the development of process diagrams like GNNs and jumping knowledge networks (JKNs). In GNNs, the features of nodes in one layer are utilized for the calculation of features in the next layer. This sequential utilization of node features helps the network capture hierarchical representations of the graph. The information flow is typically limited to nodes and their immediate neighbors in each layer. In contrast, JKNs employ a different approach by saving the features of all layers and reusing them at the end of the network. The features from all hidden layers are concatenated and utilized as the input for the final layer. This allows for a holistic consideration of features from multiple layers. Nodes in JKNs receive information solely from their neighbors, emphasizing local connections within the graph structure. GNNs rely on the sequential propagation of information through layers, while JKNs adopt a unique approach by aggregating features from all layers to provide a more comprehensive understanding of the graph’s structure and relationships. The challenges associated with GNNs include issues related to the network architecture, particularly in terms of the number of layers used. In the case of two-layer GNNs, the aggregation of information is limited to nodes that are only two steps away. This can result in inadequate modeling of larger contextual information, limiting the network’s understanding of broader patterns and relationships within the graph. Conversely, when employing multiple layers in GNNs, the aggregation process can accumulate an excessive amount of information. This may lead to a phenomenon known as over-smoothing, where the features of different nodes become overly similar. Over-smoothing can hinder the network’s ability to distinguish between nodes and capture meticulous variations in the graph, impacting the overall effectiveness of the GNNs in learning and making accurate predictions. Balancing the number of layers in GNNs is crucial to address these challenges and optimize their performance in handling graph-structured data.P&IDs as graphs. Transforming a P&ID into a graph structure by using AI tools involves leveraging technologies like GNNs or other graph-based approaches. The description of how P&IDs can be conceptualized as graphs, using AI tools, is depicted in FIG. 1.

Node representation: The representation of equipment and instruments as nodes forms the foundational structure of a graph in a P&ID. Each piece of equipment (such as pumps, reactors or vessels, and instrumentation elements like sensors or valves) can be represented as nodes in the graph. These nodes store information about the type, properties and connections of the corresponding equipment or instruments.

Edge representation: The interconnections and relationships between equipment and instruments find expression in the graph through edges. The lines and connections between equipment and instruments on the P&ID are translated into edges in the graph. These edges represent the relationships and flow of materials, energy or information between the different nodes.

GNNs: The application of AI tools—particularly GNNs—takes the analysis of P&ID graphs to a sophisticated level. GNNs excel in learning and comprehending the elaborate relationships embedded within the graph. Capable of capturing hierarchical features and dependencies between nodes, GNNs provide a robust modeling framework for the complex structure of process systems depicted in P&IDs. This advanced learning capability enhances the system’s understanding and contributes to more informed decision-making in industrial processes.

Semantic understanding: Semantic understanding of P&IDs is achieved through symbolic interpretation facilitated by AI tools, which can be trained to recognize and interpret the symbolic representations of equipment and instrumentation on the P&ID. This involves associating each symbol with a specific node type, enabling the AI model to comprehend the semantics of the diagram.

Graph embeddings: AI tools generate embeddings or vector representations for each node in the P&ID graph. These embeddings encapsulate the essential features of the equipment and instruments, facilitating tasks such as similarity analysis or anomaly detection. The ability to distill essential information into vector representations enhances the efficiency of various analyses conducted on the P&ID graph.

Topology analysis: AI tools can analyze the topology of the P&ID graph to understand the overall structure of the process system. This includes identifying critical paths, potential bottlenecks or areas where modifications may optimize the system. Such structural analysis contributes to strategic decision-making for process system optimization.

Anomaly detection: Anomaly detection is a vital application of AI tools in the context of P&ID graphs. By incorporating anomaly detection algorithms, these tools can meticulously scrutinize the graph to identify irregularities or deviations from standard process configurations. This capability is instrumental in quality control and system monitoring, ensuring the integrity and reliability of industrial processes.

Integration with CAD systems: The integration of AI tools with CAD systems further streamlines the process of P&ID analysis. AI tools can integrate with CAD systems to facilitate automatic graph generation from digital P&ID designs, streamlining the process of converting engineering drawings into graph structures. This integration enhances the efficiency of converting intricate engineering drawings into comprehensible and analyzable graph structures, bridging the gap between design and operational phases in industrial processes.

Sakthivel-Fig-01Employing AI tools to represent P&IDs as graphs enhances the understanding of complex process systems, allowing for intelligent analysis, decision-making and automation in the field of industrial engineering. The graph-based representation provides a powerful framework for leveraging advanced AI techniques for system optimization and management.Simplified flowsheet input-line entry system (SFILES). An SFILES is a text-based representation designed for chemical process flowsheets. The concept behind SFILESs draws inspiration from the simplified molecule input-line entry system (SMILES). This tool integrates functionalities facilitating the uniform conversion between PFDs/P&IDs and SFILES 2.0 strings. Opting for an SFILES as the storage and exchange format for flowsheets offers numerous advantages over the use of images and graphs. SFILES 2.0 introduces a concept for describing essential control structures vital for the functioning of chemical plants. Sakthivel-Fig-02In a study by Vogel,⁷ efforts were made to establish the groundwork for developing a database on SFILES 2.0 specifically designed for PFDs and P&IDs. Chemical plant engineering diagrams can be expressed as graph-based and text-based representations of process diagrams or directed graphs through the utilization of SFILES 2.0—e.g., an illustrative process (FIG. 2) featuring a reactor with level control and a recycle loop with flow control.⁸ In this representation, nodes correspond to unit operations and control units, while directed edges represent material streams and signals connecting these nodes. PFDs can be effectively represented as directed graphs. FIG. 3 represents a graph representation of FIG. 2. The topological details of a process graph can be effectively stored by utilizing SFILES 2.0, as illustrated in FIG. 3, in a graphical/textual format. Sakthivel-Fig-03The construct of P&IDs as graphs involves representing the diagram elements, using directed graphs.

Nodes: Represent various equipment and components within the process. Each node corresponds to a specific element such as pumps, tanks, reactors, vessels or instruments.

Edges: Represent connections between equipment and components, typically indicating piping or signal lines. Edges define the relationships and flow paths between different elements in the system.

Direction: The direction of the edges signifies the direction of the process streams. This directional information is crucial for understanding the flow of materials, energy or signals within the system.

For adopting a graph-based conceptualization, P&IDs can be analyzed using graph theory principles, allowing for a structured and systematic approach to understanding the relationships and interactions between different elements in a process. This conceptualization provides a foundation for computational methods, facilitating various analyses and optimizations in the field of process engineering.Data Exchange in the Process Industry (DEXPI) is supporting an initiative to develop a P&ID by using AI tools. This effort was launched in 2011 by a consortium comprising owner/operators, software vendors, research organizations and engineering companies. The DEXPI initiative within the industry has pioneered a data format that facilitates the shift from a document-centric to a data-centric approach, aligning with the ongoing industry transition. To facilitate this evolution, P&ID data is made amenable for AI applications through the utilization of the DEXPI format. This format serves as a machine-readable, manufacturer-independent exchange standard explicitly tailored for P&IDs. A recent study explored the application of AI in drafting P&IDs, using the DEXPI standard.⁹ The research showcased the efficacy of deep-learning models trained with DEXPI P&ID data in aiding both the engineering and drawing processes of P&IDs. The utilization of RNNs for assisted equipment prediction in P&IDs, along with consistency checks based on GNNs, was demonstrated to significantly reduce workforce time and costs associated with P&ID development. The study identified two use cases for AI-supported P&ID synthesis, along with their corresponding modeling methodologies.⁹ In the initial use case, an RNN was employed for node prediction, providing recommendations for subsequent components. These suggestions aim to assist users and streamline the drawing process, ultimately reducing the time required. The second approach involved the use of GNNs, specialized neural networks designed for graph-based modeling. GNNs can learn the topologies of process plants represented as graphs, allowing for a consistency check during drawing by comparing the models with the drawn P&IDs. Leveraging a neural network provides the advantage of learning patterns and rules from existing plant topologies for the generation of P&IDs. As depicted in FIG. 4, the process involves converting the corresponding DEXPI file of the P&ID into a graph, using Python to make the P&ID structure accessible for subsequent processing.Sakthivel-Fig-04Part 2. Part 2 of this article—to be published in the May issue—will focus on using an AI tool for autocorrecting engineering drawings. When creating P&IDs, some errors may occur, such as missing or inaccurate instrumentation details, incorrectly sized lines, incomplete or inaccurate piping information, deviations from design standards, and inadequate review processes. These mistakes can cause major safety risks, along with delays in development, leading to operational inefficiencies and extra costs. These errors can be reduced, enabling engineers to streamline workflows, reduce manual mistakes and realize faster project completion times by integrating AI. HPLITERATURE CITED

Theisen, M., et al., “Digitization of chemical process flow diagrams using deep convolutional neural networks,” Digital Chemical Engineering, December 2022.

Haar, C., K. Hangbeom and L. Koberg, “AI-based engineering and production drawing information extraction,” FAIM 2022, Detroit, Michigan, June 19–23, 2022.

Liuyue, X., et al., “Graph neural network-enabled manufacturing method classification from engineering drawings,” Computers in Industry, November 2022.

Zhou, J., et al., “Graph neural networks: A review of methods and applications,” AI Open, 2020.

Fu, B., Y. Gao and W. Wang, “Dual generative adversarial networks for automated component layout design of steel frame-brace structures,” Automation in Construction, February 2023.

Sherstinsky, A., “Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network,” Physica D: Nonlinear Phenomena, March 2020.

Vogel, G., E. Hirtreiter, L. S. Balhorn and A. M. Schweidtmann, “SFILES 2.0: An extended text-based flowsheet representation,” Optimization and Engineering, 2023.

Hirtreiter, E., L. S. Balhorn and A. M. Schweidtmann, “Towards automatic generation of piping and instrumentation diagrams (P&IDs) with artificial intelligence,” AIChE Journal, 2022.

Oeing, J., et al., “Using artificial intelligence to support the drawing of piping and instrumentation diagrams using DEXPI standard,” Digital Chemical Engineering, September 2022.

First Author Rule LineAuthor-pic-SakthivelS. Sakthivel works as the Deputy General Manager in the Technology Team at TATA Consulting Engineers Ltd., Mumbai, India. His expertise focuses on green chemicals, green fuels, energy transition and decarbonization, with a key focus on evaluating emerging technologies for commercialization. Dr. Sakthivel has experience in process engineering, technology analysis, screening and selection, as well as techno-economic assessments, pilot project development and process hazard analysis. His background also extends to basic, applied and market research, as well as powder and science technology. He has authored several articles in both national and internationally peer-reviewed journals.CoverAd—Shell revampContentsAd—MerichemMastheadAd—Honeywell UOPEditorial CommentAd—Curtiss Wright HVGlobal Project DataAd—E2G HVInnovationsAd—IRPCPodcastsAd—CCS Strategy (Early Bird)Business Trends—Yankovitz (Deloitte)Ad—GEI AwardsSF OpeningSPECIAL FOCUS—Lee (SK Energy)SPECIAL FOCUS— Rakers (PPG Protective & Marine Coatings)Ad—GastechSPECIAL FOCUS—Zhang (CastNewX Technology)SPECIAL FOCUS—Khanna (Engineering Consultant)Ad—WGLCProcess Optimization—Chua (University of Nottingham Malaysia)Process Optimization—Patron (Contributing Author)Ad—InstruCalcPetrochemical Technologies—Xu (Sinopec)Ad—HPI Market Data 2024PDECC—Sakthivel (TATA Consulting Engineers)Digital Technologies—Bowling (Seeq)Ad—PE Outlook 2024Environment and Safety—Khan (Contributing Editor)Ad—BoxscoreHydrogen—Chaudhuri (Fluor)Ad—H2T CircPower and Utilities—Srivastava (Fluor India)Sales OfficeAd—HP Circ (Back Cover)