Embedded Files
AI in
AI is rapidly transforming engineering design, introducing tools that accelerate innovation, enhance precision, and streamline workflows. This post explores key advancements, from machine learning enhanced automation over surrogate models to generative design systems, and their real-world applications.
Note that for this article, I am listing my references in the small "link" icons below each segment to make it easier for people to read up on these topics more in depth.
Also note that this article is best consumed on a Laptop or Tablet. For Mobile the figures might be a little squeezed.
Last updated: 17th February 2025
Executive Summary
Executive Summary
Engineering (design) is undergoing a transformation through the implementation of machine learning or artificial intelligence. Clear patterns emerge in both research and commercial applications. I reviewed 62 research publications and 20 commercial AI applications on the market to understand the emergence of these methods better.
After an explanation of basics of engineering processes, like the process steps themselves and the modalities of data in use, some core thematic streams emerged:
Where do Engineers use AI already:
Here, after analysing available literature and information commercial applications core areas of application can be identified.Research suggests that Design Automation leads adoption at 42% of applications, followed by Simulation Enhancement at 24% and Knowledge management (13%)
Commercial solutions mirror this distribution, with notable clusters in design automation and simulation tools.
What neural models are mainly used:
To better understand how certain applications can be enabled, the most prominent neural models need to be portrayed.Transformers dominate technical approaches (20%), with specialised architectures for engineering applications, like Text to CAD.
Graph Neural Networks (13%) and Physics-Informed Neural Networks (11%) show strong presence in simulation
Multi-modal AI systems, particularly Vision-Language Models and Large Language Models, are enabling new capabilities in engineering documentation and design interpretation
How is the market developing:
Market maturity remains fairly low, with only 1% of companies reaching AI maturity as of 2023
The Commercial landscape shows uneven adoption across the engineering process chain with clear technology clusters forming around design automation (e.g., nTop, Hyperganic, Synera) and simulation (e.g., Altair, NVIDIA, Neural Concept, Autodesk)
Looking ahead, it could be expected, that an acceleration of AI adoption would take place through 2025, with particular growth in knowledge management and process integration solutions. Furthermore, the emergence of geometry-aware neural networks and physics-informed approaches suggests a shift toward more sophisticated, engineering-specific AI applications.
Table of contents
Table of contents
Core Concepts and Terminology
Core Concepts and Terminology
In this section some core concepts and terminologies relevant to this post are explained. Let's start out with engineering design and the relevant processes.
Engineering Design
Engineering Design
Engineering design is a systematic process that combines scientific knowledge, engineering expertise, and creative thinking to conceptualise, develop, and optimise products. It involves several stages, from problem identification and research to concept generation, detailed design, and prototyping. Engineering development, on the other hand, focuses on turning design concepts into functional and manufacturable products. The key engineering design principles are:
Design for Functionality
Design for Safety
Design for Manufacturability
Design for Reliability
Design for Sustainability
Design for Adaptability
There are several methodologies that engineers follow and that help defining the structure of the engineering design process. Examples include: Design Thinking, the Waterfall Method, Agile Methods, Concurrent Methods like the V-Model and more. Each industry specifies clear Engineering Design standards and methodologies. As an example, engineering design within the European Cooperation for Space Standardisation (ECSS) framework is a structured, standards-driven process optimised for reliability, collaboration, and compliance in space systems development. For simplicity a schematic engineering design approach is shown in the following figure, that includes all relevant steps to get from an idea to a final product.
In the above figure, the engineering design process is split up into seven steps that are categorised in four distinct categories: "Conceive", "Design", "Fabricate", and "Service". These categories are also widely used in Product Lifecycle Management (PLM) systems, which streamline design workflows by centralising data and enabling cross-functional collaboration.
Read some more about CAD systems and their history here as well as the Geometric Modelling Landscape here.
Engineering Data Types
Engineering Data Types
In engineering, data is the fundamental building block of understanding and innovation. The way we structure and process information spans a rich spectrum from rigidly organized grids to fluid, unstructured networks. This diversity reflects the complex challenges engineers face—whether analyzing precise sensor readings, mapping intricate design relationships, or interpreting nuanced system interactions.
From regular meshes and voxel grids to social network graphs and raw volumetric scans, each data structure offers unique capabilities for modelling, simulation, and analysis. Understanding these structural variations enables engineers to select the most appropriate approach for transforming raw information into actionable insights, bridging the gap between raw data and meaningful engineering solutions.
In the figure below you'll see how engineering data can be structured into buckets that are defined by three different data representations and by the nature of this data, i.e. whether it is structured, semi-structured or unstructured:
Grid: Regular or irregular Grids of data points, like tables or voxel data.
Graph: Regular or irregular Graphs of data points, like surface meshes, knowledge graphs or social networks
Series: Regular or irregular Series of data points, like sensor readings or text
The entries in the above table can apply to multiple buckets. Therefore for further illustration, find descriptions of practical data types you might encounter in the dropdowns below.
Documentation
Documentation
Text/ Logs
Engineers generate extensive documentation for project planning and execution. Technical specifications, project requirements, and design documents form the foundation of engineering planning, often stored in version-controlled repositories alongside code. Engineers also maintain detailed documentation of system architecture, API specifications, and deployment procedures, while creating progress reports and meeting notes to track project development.
Technical Drawings and Sketches
Engineering drawings serve as the universal language for technical communication, encompassing detailed 2D and 3D representations of components, assemblies, and systems. These drawings follow standardised conventions including dimensioning, tolerances, and symbols to precisely convey design intent and manufacturing requirements. Technical drawings range from initial sketches and assembly drawings to detailed part drawings and wiring diagrams, all typically created using Computer-Aided Design (CAD) software that enables version control and collaborative editing. Construction drawings specifically detail building plans, elevations, sections, and mechanical/electrical systems, while manufacturing drawings include material specifications, surface finishes, and quality control requirements.
Sensor Data
Engineers working with sensor and machine data manage both structured and unstructured formats. Structured data includes organised, predefined formats like numerical sensor readings (e.g., temperature, pressure, GPS coordinates), timestamps, and relational database entries from IoT devices or industrial machines. This data is stored in SQL databases or data warehouses and is easily queryable for tasks like performance monitoring or predictive maintenance. Unstructured data, such as raw video feeds, audio recordings, LiDAR scans, social media posts, and maintenance logs, lacks a fixed schema and requires advanced tools like Natural Language Processing (NLP) or computer vision for analysis. Semi-structured formats like JSON logs from IoT sensors or XML files bridge the gap, offering partial organisation without rigid schemas. The convergence of these data types—from structured telemetry to unstructured multimedia—demands scalable storage solutions (e.g., data lakes) and integration techniques to extract actionable insights for applications like autonomous systems or real-time anomaly detection. See an example of accelerometer time series sensor data in the figure below. Figure taken from here.
Geometry Representations
Geometry Representations
In this section the different 3d geometry representations are considered. The figure below shows different representations of the same sphere. These representations are explained in the following.
Point Clouds
Point clouds are collections of 3D data points representing the surfaces of physical objects or environments, captured using technologies like LiDAR, photogrammetry, or RGB-D cameras. Each point contains spatial coordinates (x, y, z) and attributes like color or intensity, creating a highly detailed digital replica of real-world structures. Engineers leverage this data across industries for precision modelling, analysis, and decision-making. Point clouds can be expressed in as structured graphs or as unstructured coordinate lists.
Engineers leverage point clouds across industries for precision, efficiency, and visualisation:
Construction & Architecture
Creating as-built models for renovations by capturing existing structures.
Monitoring construction progress through real-time comparisons of scans vs. design models.
Conducting topographic surveys for infrastructure planning.
Manufacturing & Quality Control
Comparing scanned parts to CAD models to detect deviations in geometry.
Reverse engineering components using high-accuracy 3D scans.
Robotics & Autonomous Systems
Enabling environment mapping and obstacle detection for autonomous vehicles.
Supporting robotic navigation in complex industrial settings.
Energy & Infrastructure
Inspecting pipelines, power plants, and offshore rigs to reduce manual inspections.
Modeling digital twins for predictive maintenance.
BRep (solid)
For the sake of simplicity BReps and solids are presented at the same time in this post. BRep (Boundary Representation) and solid are closely related but distinct concepts in 3D modeling. A BRep is a method for defining an object’s geometry by describing its surfaces, edges, and vertices, while a "solid" refers to a watertight 3D volume that unambiguously separates interior and exterior space.
BRep (Boundary Representation) is a method for defining 3D objects by mathematically describing their enclosing surfaces. It combines topological data (how surfaces, edges, and vertices connect) with geometric data (equations for curves/surfaces like planes, cylinders, or NURBS).
Engineers use BReps across industries for precision modelling and interoperability:
CAD & Manufacturing
Creating parametric models of mechanical parts with classical CAD tools like Fusion 360, SolidWorks, PTC Creo or Rhino.
Performing Boolean operations (union, subtraction) and feature-based modelling (extrusions, fillets).
Quality Control
Validating designs by comparing BRep models to scanned point clouds.
Detecting geometric flaws like gaps, overlaps, or non-manifold edges.
Simulation
Using BReps in finite element analysis (FEA) tools for stress testing.
Meshes
Meshes are discrete representations of geometric shapes used to approximate complex surfaces or volumes in computational models. They consist of interconnected vertices, edges, and faces (polygons like triangles or quadrilaterals) that define the topology and geometry of an object. Meshes enable numerical simulations, rendering, and analysis across engineering and computer graphics by breaking down continuous domains into manageable elements. Meshes are fundamentally connected to graph theory through their structural representation, where vertices, edges, and faces form interconnected nodes and links. This relationship enables engineers to apply graph-based algorithms for optimisation and analysis, while meshes themselves remain critical tools for simulating complex physical phenomena in engineering design.
The importance of meshes in engineering design can be broken down into the following points:
Simulation Accuracy
Meshes discretise complex geometries into elements where partial differential equations (PDEs) are numerically solved
Adaptive meshing refines critical regions (e.g., stress concentrations) to improve accuracy while minimising computational cost
Workflow Efficiency
Structured meshes (hexahedral elements) reduce solver time for regular geometries, while unstructured meshes (tetrahedrons) handle irregular shapes.
Tools like Ansys Mechanical and OpenFOAM rely on hybrid meshing (combining hex/tet elements) to balance speed and precision.
Manufacturing and Validation
Surface meshes (STL files) communicate CAD data for 3D printing, though limitations in geometric fidelity persist.
Mesh quality directly impacts FEA/CFD validation: skewed elements or poor aspect ratios cause divergence errors.
Voxels
A voxel (short for "volume element") is the 3D equivalent of a pixel, representing a discrete unit of volume in a spatial grid. Each voxel contains spatial coordinates (X, Y, Z) and may store attributes like color, density, or material properties. Unlike polygons or BReps, voxels model volumetric data directly, making them ideal for representing complex internal structures and heterogeneous materials.
Engineers leverage voxels across industries for precision and computational efficiency:
Additive Manufacturing
3D Printing: Voxel-level control enables multi-material printing with spatially varying properties (e.g., flexible hinges in rigid components).
Lightweight Design: Voxel-based generative design optimises material distribution for weight reduction while maintaining strength. (see above an image illustrating topology optimisation, curtesy)
Medical Imaging
CT/MRI Analysis: Voxel grids reconstruct 3D anatomical models for surgical planning and tumor detection.
Bio-printing: High-resolution voxel grids guide cell placement in tissue engineering.
Geospatial Modeling
Digital Twins: Voxel grids model underground geology for oil/gas exploration or urban infrastructure planning.
Environmental Simulations: Volumetric data tracks fluid flow in porous media or pollutant dispersion.
Implicit
Implicit surfaces (often called "implicits") are mathematical representations of 3D shapes defined by scalar fields, where the surface is the zero-level set of a continuous function f(x,y,z)=0. Unlike explicit representations like meshes or B-reps, implicits encode geometry through equations rather than vertices/edges, enabling infinite resolution and robust handling of complex topologies.
Engineers leverage implicits for precision, automation, and complex geometry creation:
Additive Manufacturing
Multi-Material Printing: Voxel-level control enables spatially varying properties (e.g., rigid-flexible transitions).
Lattice Structures: Generate lightweight designs with implicit TPMS patterns (gyroids, diamonds) at scale (1M+ cells in milliseconds).
Generative Design
Field-Driven Optimization: Combine stress/strain fields from FEA with implicit geometry to create performance-driven structures.
Topology Optimization: Convert simulation results directly into manufacturable implicit models, bypassing error-prone CAD conversions.(see above an image illustrating topology optimisation, curtesy)
Medical Imaging & Bioprinting
Anatomical Modeling: Reconstruct organs from CT/MRI scans using signed distance fields (SDFs).
Tissue Engineering: Guide cell placement with high-resolution implicit grids for 3D bio-printing.
See in the following figure a SDF of a Gyroid created in Shadertoy.
Summary on geometric file types
Engineering Design Process
Engineering Design Process
Shape Grammar
A shape grammar is a set of shape rules that apply in a step-by-step way to generate a set, or language, of designs. Shape grammars are both descriptive and generative. The rules of a shape grammar generate or compute designs, and the rules themselves are descriptions of the forms of the generated designs.
Shape grammars have properties aimed at making them especially suitable for designing, without sacrificing formal rigour. First, shape grammars are spatial, rather than textual or symbolic, algorithms. Second, shape grammars treat shapes as nonatomic entities--they can be freely decomposed and recomposed at the discretion of the designer. This liberty allows for emergence, which is the ability to recognise and, more importantly, to operate on shapes that are not predefined in a grammar but emerge, or are formed, from any parts of shapes generated through rule applications. Third, shape grammars are nondeterministic. The user of a shape grammar may have many choices of rules, and ways to apply them, in each step of a computation. As a design is computed, there may be multiple futures for it that respond differently to emergent properties, or to other conditions or goals.
See below a schematic depiction of how shape grammar works in 2d.
In Wallas et al. shape grammar is used to overcome the problem of geometry scalability by assembling a geometry through parametric objects that define a scope and an instance, effectively describing the component fully in 3d. See an example of that in the image below.
Design Intent
Design intent is generally understood simply as a CAD model’s anticipated behaviour when altered. However, this representation provides a simplified view of the model’s construction and purpose, which may hinder its general understanding and future reusability. It encompasses the logic behind how features, dimensions, and constraints are structured to maintain functionality and manufacturability when changes occur. This concept is closely tied to design trees—the hierarchical feature lists in CAD software that document modelling steps and dependencies.
Design trees inherit properties of graphs, including nodes (features like sketches, extrusions) and edges (dependencies or parent-child relationships), however unlike general graphs, they enforce Directionality and Hierarchy.
PLM Systemgraphs
PLM (Product Lifecycle Management) system graphs are transforming how engineering teams manage complex product data by leveraging graph theory and database technologies to create interconnected, dynamic representations of product information. These graphs enable unprecedented visibility into relationships between design, manufacturing, supply chain, and quality data across the entire product lifecycle.
Deep Neural Models
Deep Neural Models
The landscape of artificial intelligence (AI) in engineering has evolved into a rich ecosystem of specialized neural networks and traditional rule-based systems. Early AI relied heavily on rule-based approaches like shape grammars and design intent to automate design tasks. While still relevant, these methods have been largely complemented by adaptive, data-driven neural networks.
Modern AI offers a diverse toolkit for engineers. Multi-Layer Perceptrons (MLPs) handle non-linear data patterns, while Convolutional Neural Networks (CNNs) process grid-structured data like images and 3D volumes. Recurrent Neural Networks (RNNs) are essential for sequential data, enabling time-series forecasting and predictive maintenance. Graph Neural Networks (GNNs) excel at learning from graph-like data, making them ideal for simulation surrogates, while Reinforcement Learning (RL) optimizes processes through trial and error, such as generating shape grammars.
Generative models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) enable innovative design exploration, from material synthesis to accelerating topology optimization. Physics-Informed Neural Networks (PINNs) and Deep Operator Networks (DeepONets) offer physics-based solutions for complex systems, embedding physical laws directly into their training.
The advent of Transformers revolutionized AI, leading to Large Language Models (LLMs) and Vision-Language Models (VLMs). These models enable advanced capabilities in text generation, multimodal analysis, and automation of engineering tasks, bridging language and vision in new ways.
The following graph provides a visual overview of these neural network architectures as well the hierarchy of development. For instance: After the initial description of neural networks, convolutional neural networks were developed to process image (2d grid data) data, which again opened up the possibilities for further developments of 3d generative models, like the U-Net.
In the graph shown the basic architectural features are illustrated:
Dense (fully connected) layers as shown by rectangles
Convolutional layers (3d cuboids)
Graph like connections as shown by the graph schematic
Data flow as illustrated by arrows
Abbreviations:
in Physics Informed Neural Networks: PDE: partial differential equation
in Variational Auto-encoders and Transformers: E: Encoder, D: Decoder
in Generative Adversarial Networks: G: Generator, D: Discriminator
in Deep Operator Networks: B: Branch Net, T: Trunk Net
In the following the individual network architectures and their respective functionality is explained in more detail including some links to educational videos by well respected institutions.
Fundamental Networks and Concepts
Fundamental Networks and Concepts
Multi Layer Perceptron
In deep learning, a multilayer perceptron (MLP) is a name for a modern feedforward neural network consisting of fully connected neurons with nonlinear activation functions, organised in layers, notable for being able to distinguish data that is not linearly separable
Convolutional Neural Networks
Convolutional Neural Networks (CNNs) first introduced in 1987, are deep learning architectures designed to process grid-structured data (e.g., images, videos, 3D volumes) by leveraging spatial hierarchies and translation-invariant features. CNNs utilise convolutional layers, essentially filter-based feature extraction, pooling layers to downsample for translational invariance, and fully connected layers for the final classification/ or regression to hierarchically learn patterns.
Recurrent Neural Networks
Recurrent Neural Networks (RNNs) are a class of artificial neural networks designed to process sequential data by maintaining an internal state (memory) that captures information from previous inputs. RNNs process input sequences one element at a time, updating their hidden state at each step. RNNs find application in time-series forecasting (e.g., energy demand prediction, equipment failure anticipation) and sequential data processing.
Graph Neural Networks
Graph Neural Networks (GNNs) are deep learning models designed to process graph-structured data (nodes connected by edges), enabling relational reasoning and pattern discovery in non-Euclidean domains. GNNs use message passing to propagate and aggregate features across neighboring nodes, capturing structural relationships. This allows nodes to learn embeddings that reflect both their attributes and their position within the graph. GNNs can typically be used to learn graph like data such as meshes and therefore lend themselves ideally for simulation surrogates.
Deep Reinforcement Learning
(Deep) Reinforcement Learning, first introduced in 1996, fundamentally differs from the other Machine Learning Methods, in that it learns without a dataset in an unsupervised fashion through a large set of trial and error interactions between an actor and an environment. This is typically done through some reward signal being sent to the actor after taking actions, based on the effects of said actions on the environment. In this scenario, the actor’s goal is to maximize the rewards it receives by making decisions (i.e., taking actions) such that the total reward is maximized. From this point of view, reinforcement learning can be thought of as an approach similar to optimization, where an objective (maximizing the reward) is being optimized. RL is applied in various fields in engineering, like generations of shape grammars.
Advanced Networks and Concepts
Advanced Networks and Concepts
Generative Adversarial Networks
Generative adversarial networks (GANs) provide a way to learn deep representations without extensively annotated training data. They achieve this through deriving back-propagation signals through a competitive process involving a pair of networks. The representations that can be learned by GANs was originally used in a variety of applications, including image synthesis, semantic image editing, style transfer, image super-resolution and classification. GANs overwhelmingly find application in the synthesis of (meta-)materials and in acceleration of Topology Optimisation.
Variational Autoencoders
Variational Autoencoders (VAE) were first introduced in 2013 by Kingma et al. Autoencoders are unsupervised embedding algorithms consisting of an encoder that maps an input design into a (typically) lower-dimensional latent space and a decoder that reconstructs the design as accurately as possible from the latent space. The encoder and decoder are conventionally implemented using deep neural networks. To generate new samples, latent vectors are sampled from the latent space and fed through the decoder. Typically, the distribution of the real data mapped to the latent space of an autoencoder is sparse, meaning that sampling a realistic latent vector is difficult. This limitation is addressed with the introduction of the Variational Autoencoder (VAE). VAEs mainly find application in the generation of novel (meta-)materials.
Large Language Models
Large Language Models (LLMs) recently demonstrated extraordinary capability in various natural language processing (NLP) tasks including language translation, text generation, question answering, sentiment analysis, etc. They primarily employ the Transformer architecture, which utilises self-attention mechanisms to weigh relationships between words in a sequence, enabling the capture of long-range dependencies and contextual understanding. LLMs undergo a two-stage training process: pre-training on massive datasets to learn general language patterns, followed by fine-tuning on specific tasks to adapt their capabilities. Key architectural components include self-attention layers, positional encodings, feed-forward networks, and layer normalization, which collectively enable LLMs to process and generate human-like text. While LLMs have found applications in chatbots, text generation, code synthesis, and knowledge retrieval, they face challenges such as hallucinations, bias amplification, and high computational costs, necessitating ongoing research to address these limitations.
Vision Language Models
Vision-Language Models (VLMs) are multimodal AI systems designed to process and understand both visual and textual information simultaneously, enabling tasks like image captioning, visual question answering, and document analysis. They typically employ a combination of image encoders (such as CNNs or Vision Transformers) to process visual data, text encoders (often transformer-based) to handle language, and fusion mechanisms to align these modalities, allowing for coherent multimodal reasoning. VLMs undergo a two-stage training process: pre-training on large datasets of image-text pairs to learn general associations, followed by fine-tuning on specific tasks to adapt their capabilities to particular domains or applications. Key architectural components include visual and textual encoding layers, cross-modal attention mechanisms, and task-specific decoders, which collectively enable VLMs to interpret and generate content that integrates both visual and linguistic understanding. While VLMs have shown remarkable progress in bridging vision and language, they face challenges such as handling text-rich images, resolution limitations, and mitigating biases inherent in training data, driving ongoing research to enhance their performance and applicability across diverse scenarios.
Physics Informed Neural Networks
Physics-Informed Neural Networks (PINNs) first introduced by Karniadakis et al., are not inherently classified as deep generative networks, but they share conceptual overlaps and can be integrated with generative approaches in certain applications. Designed to solve partial differential equations (PDEs) by embedding physical laws (e.g., conservation principles, boundary conditions) into neural network training. They act as universal function approximators for physics-driven systems.
Deep Operator Networks
DeepONet (Deep Operator Network) is a neural network architecture designed to approximate nonlinear operators that map input functions to output functions, enabling real-time predictions for complex systems like differential equations. Developed by researchers at Brown University, it builds on the universal approximation theorem for operators, allowing it to learn continuous mappings between infinite-dimensional function spaces. DeepONets are part of the family of PINNs.
Categorisation of Neural Models
Categorisation of Neural Models
Now to conclude this chapter let's couple the learning paradigm to the data structure and see which network falls into which bucket. In the table below you'll see how deep neural models can be categorised according to the data type as well as the overarching paradigms they are best applied for. The respective references to the entries can be found below or in the links in the relevant text segment.
AI in Engineering
AI in Engineering
In this section we try to build an understanding on where AI is used in engineering. For simplicity reason some high impact examples are picked and elaborated in more detail. Let's dive into this chapter with a high level overview about where AI models find application in engineering.
Literature Review
Literature Review
To understand the impact of AI in engineering better I reviewed 63 openly available sources published between 2004 and 2025, with roughly 80% (79.5%) of the reviewed papers published between after 2018. 62 of the reviewed literature are academic publications and one normal web-link. See a chart with the distribution of papers by publication date as well as a complete list of all references at the end of this section.
Categorisation of Literature
Categorisation of Literature
To effectively categorise the contents labels are introduced. The share of these categories with respect to the total number of papers reviewed is displayed (refer to the first chart on the right). The approach to select and assign labels can be seen in the data table at the end of this chapter.
Design Automation (42%): Accelerating or Augmenting the engineering design process. See a breakdown of the Design Automation Category in the second chart on the right.
Generative Design: Accelerating the constraint driven design for example by means of topology optimisation or other fully data driven generative methods.
Text 2 CAD: Improving the handling of CAD tools by automating CAD model generation with text prompts
Reverse Engineering: Data driven reverse engineering of CAD models for example from drawings
Meshing: Data driven optimisation of adaptive meshing
Simulation (24%): Accelerating or replacing physics simulation
Knowledge Management (13%): Data handling and knowledge driven engineering handling of documentation and knowledge graphs
Process Automation (12%): Automation engineering processes like data generation, or classical equipment automation
and 6. Data Management and Integration (11%): The smallest two categories revolve about handling and integrating date into existing engineering processes
The literature reveals a diverse landscape of AI applications, with particular emphasis on simulation, generative design, and data management workflows, indicating a clear trend toward computational enhancement of core engineering processes.
The technical approaches employed in these studies showcase a rich variety of neural methodologies, with Transformers* leading at 20% of implementations, followed by Graph Neural Networks (GNN) at 13%, and Physics-Informed Neural Networks (PINN), including DeepONets, as well as Convolutional Neural Networks (CNN) at 11% .
This distribution of neural network architectures reflects the engineering community's growing recognition of specialised neural architectures that can effectively handle the complex, structured data typical in engineering applications. See the third pie chart on the right.
The Transformer Category can be split up further into:
Vision Language Models (VLM) (30% of Transformers)
Large Language Models (35% of Transformers)
Other Transformer Architectures, that make use of the Attention mechanisms (35% of Transformers)
Correlation of Applications and Models
Correlation of Applications and Models
Using the labeled literature, we can establish the connection between individual models and their respective engineering applications. Each number correlates to a reference in the literature table later in this chapter.
Simulation: Physics Informed Neural Networks (including DeepONets) and Graph Neural Networks dominate simulation literature. Key considerations include available data and underlying physical laws.
Knowledge Management: Large Language Models and Knowledge Graphs are the primary models. Data modality remains a critical factor. Knowledge Graphs form the foundation for many NoSQL data structures.
Data Management, Integration, Process Automation: This field shows more model diversity, potentially due to varied data modalities and interfaces.
Design Automation: Most literature focuses on this area. Graph Neural Networks are prevalent in general design automation. Generative design relies more on Convolutional Neural Networks and Generative Adversarial Networks, likely due to their effectiveness with grid-like data. Transformers excel at bridging natural language and automated CAD model generation.
Legend: * Generative Design; ** Reverse Engineering; ` Text to CAD; `` Meshing
Identified Key Trends
Identified Key Trends
Lets look at the key trends identified:
The rise of geometry-aware neural networks, particularly in CAD and design automation, with multiple papers exploring novel architectures for processing 3D structures and mesh-based simulations. Recent works like "Geometric Deep Learning for Computer-Aided Design" and "DeepCAD" demonstrate significant advances in automated design synthesis and optimisation.
A strong emphasis on physics-informed approaches, where traditional simulation techniques are being augmented with neural networks to achieve faster, yet physically accurate results. This is evidenced by numerous publications on surrogate modelling and reduced-order modelling applications.
The emergence of multi-modal AI systems has enabled the interpretation and generation of engineering content across different representations – from natural language and sketches to 3D models and technical documentation. The development of systems like Text2CAD and Img2CAD exemplifies this trend. Furthermore, engineers are increasingly leveraging Transformer-based models, particularly Vision-Language Models (VLMs) and Large Language Models (LLMs), to process extensive documentation efficiently. A notable emerging technique in this context is Retrieval Augmented Generation (RAG), which enhances these models' capabilities by grounding their responses in relevant technical documentation and engineering knowledge bases.
A growing focus on integrating AI into engineering workflows through automation and integration solutions (14% of applications), suggesting a shift from purely theoretical research to practical implementation.
Relevance of Agentic AI
Relevance of Agentic AI
The reviewed literature further suggests a shift towards more autonomous and interactive AI systems. Especially the widespread use of GNNs, Transformers and most importantly PINNs shows that AI systems are built to automatically detect and prioritize relevant features (Attention in Transformers), learn connections between features (Graphs) and understand explicit physical constraints (PINNs) - mirroring how human engineers approach complex design and analysis tasks.
Commercial Applications
Commercial Applications
Machine Learning already today enhances various bespoke software products on the market. In this section, we'll cover some of these tools and structure them according to where and how they add value in the engineering design process. In the table below, companies are categorised based on their position in the design chain. While not exhaustive, this overview illustrates where companies are actively supporting engineers with AI-powered solutions.
What can be particularly noted is that distinct clusters have emerged in the fields of design automation and simulation, though the boundaries between these categories often overlap and remain fluid in practice. The design automation space shows a rich ecosystem of tools, including solutions like nTop, Hyperganic, and ToffeeX, which focus on generative design and optimisation. In the simulation domain, companies like Altair, NVIDIA, Neural Concept, and SimScale are leveraging AI to enhance traditional simulation workflows.
In knowledge management, players like Glean are developing sophisticated solutions for (engineering) knowledge capture and retrieval. Meanwhile, data management and integration tools like Toolkit3D and Synera are positioning themselves in the detailed design phase, though their applications may span multiple stages of the design process.
This distribution suggests that AI adoption in engineering is not uniform across all stages of the design process, with certain areas seeing more concentrated development and innovation than others. It's worth noting that this landscape is rapidly evolving, with companies continuously expanding their capabilities and often bridging multiple categories as the technology matures.
Find all links to the logos used in this table in the link section below.
A more exhaustive table of modern and legacy CAx tools has been done by Blake Courter, founder of Gradient Control Labs. For interested readers find his work embedded below.
State of Adoption
State of Adoption
While AI adoption in engineering is progressing rapidly, it's important to note that according to McKinsey Digital as of 2023, only 1% of companies believed they had reached AI maturity. However, according to LinkedIn sources by 2025, we can expect a more significant increase in AI deployment across enterprises, with an increase in the deployment of AI agents.
Some final Remarks
Some final Remarks
I wrote this article mainly due to my interest in AI in engineering, a field that is rapidly evolving and extremely versatile and fascinating. However, as with all new tech, there is of course also the danger of deception and overinflated narratives. Throughout the past years I was always missing one consolidated (educational) document, that summarises the current status of development and implementation. So now here it is. This document hopefully helps to educate and set the stage for ML adoption in engineering.
This was meant to be free to read for all and a great learning opportunity for me.
Feel free to reach me on LinkedIn if you have feedback or questions!
Page updated
Google Sites
Report abuse