Skip to main content

Do you still think Machine Learning and Artificial Intelligence are for Computer Science related fields? If so, you’re overlooking the incredible advancements reshaping the engineering landscape. In the worst case, failing to embrace these innovations could leave you trailing behind in a rapidly evolving world. This blog posts will help you introduce as well as explain the importance of generative designs in Mechanical Engineering. This is a beginner’s Introduction to generative design in engineering hence you don’t need to have prior knowledge of the concepts.


As engineers, keeping up with technology is important to leverage its advantages. The days of spending days brainstorming on better and more diverse designs are gone. In recent years, conventional design generation and manufacturing processes have been revolutionized. With the advancement of automation using Artificial Intelligence (AI) and Machine Learning (ML), generative designs are helping engineers generate innovative and optimized solutions. This has pushed the boundaries of what was previously possible.


In this blog, we will embark on a journey to demystify degenerative design and its application in mechanical engineering. Whether you’re a student, a seasoned professional, or simply curious about this intersection of design generation and AI, this blog will equip you with the knowledge and tools to dive into this fascinating domain.

We will cover following topics in this post:

  • Introduction to Generative Designs
  • Brief history of Generative Designs
  • Soft introduction to latest Deep Generative Models
  • Understanding why generative design is important in mechanical engineering with example papers.


What is Generative Design?

The generative design leverages computational power to explore vast design spaces and generate novel solutions. Traditional design methods used  rely on intuition and manual iterations. However, generative design empowers engineers with algorithms that can intelligently explore numerous design options, considering various parameters, constraints, and objectives.

Many generative designs, called Deep Generative Models (DGM), are available to automate design automation. Some prominent DGMs include Generative Adversarial Networks (GANs)Variational Autoencoders (VAEs), and feedforward Neural Networks (NNs), which have yielded astonishing results. Over the past decade, extraordinary results have been achieved in Solid Mechanics, Fluid Dynamics, and Thermodynamics.


Brief History of Generative Models:

Generative AI has a rich history and dates back to 1950s and 1960s, which shows it is not something discovered over the last decade. Models like Markov Chain Monte Carlo (MCMC) method by Metropolis et al. in early 1950s to Probabilistic Graphical Models in 1980s onwards to VAEs in early 2000s have lead the way to the major breakthrough made in 2014 by Ian Goodfellow with his novel approach of Generative Adversarial Networks.

Over the last decade, GANs have been rapidly modified and advanced into various domains. Some of the prominent variations of GANs include Conditional GANs (CGANs), Deep Convolutional GANs (DCGANs), and Wasserstein GANs (WGANs).

Generative models have been applied to diverse applications, including image synthesis, style transfer, text generation, music composition, drug discovery, and more. Researchers continue to explore novel architectures and techniques to enhance the capabilities of generative models.

It is not mandatory to be familiar with all the previous models. However it is important to note that the history of generative models is an ongoing and dynamic field with much potential. New advancements and breakthroughs continue to push the boundaries of this amazing field enabling exciting possibilities in AI, creativity, and innovation.

Generative Adversarial Networks (GANs)

Generative Adversarial Networks (GANs) were first introduced by Ian Goodfellow in 2014 in his paper Generative Adversarial Nets. They proposed a network with two models, a generator and a discriminator model, which were trained simultaneously. The generator model (G) works by mapping an arbitrary noise distribution to the design data distribution and using that to generate new data. However, the discriminator model (D) is trained to distinguish between real and fake data. Both models consist of neural networks, usually CNN, these days.

As the discriminator learns to differentiate between real and fake images, Generator takes the loss and improves by generating images that can ‘fool’ the discriminator.

There are many difficulties in training because of the instability of the GANs. However, many issues have been resolved by developing novel algorithmic techniques by modifying the GAN structures and including various loss functions.

Architecture of a GAN Model 



Variational Autoencoders (VAEs):

Variational Autoencoders were first introduced in 2013, and quickly gained success in many ML applications. The model consists of an encoder and a decoder which utilize unsupervised embedding algorithm to generate new images.

The encoder and decoder components in generative models are typically implemented using deep neural networks. The encoder takes input data and maps it to a lower-dimensional latent space, capturing important features and representations. The decoder, on the other hand, takes latent vectors sampled from the latent space and transforms them back into the original data domain, generating new samples.

This process allows for the generation of novel samples by sampling from the latent space and passing it through the decoder network. Deep neural networks enable the modeling and learning of complex mappings between the input data and the latent space, facilitating the generation of diverse and realistic samples.


Variational Autoencoders Structure

Why is Generative Design Important in Engineering?

The manual and iterative process of human designing is an important process which play an important role in the engineering design. However, accelerating the process will not reduce the computational time associated with the iterative process but also the cost whilst increasing productivity and efficiency of the design generation process.

Through AI-powered generative design techniques, the design process becomes more efficient and streamlined. By leveraging the computational capabilities of AI algorithms, a vast array of design possibilities can be explored and evaluated rapidly. This allows for the creation of optimal products that meet the unique requirements and preferences of individual users.

Here, we’ll look at few applications of generative design models in engineering, with example pictures where you can understand how these models are being implemented.

Learn more on Why Mechanical Engineers should learn Generative Models


1. Structural Optimization and Aesthetics:

Generative design enables engineers to optimize and improve mechanical designs for various parameters, such as weight, strength, energy efficiency, and performance. By exploring a vast design space and generating multiple design alternatives, generative design helps identify optimal solutions that meet specific requirements and objectives. This leads to more efficient and optimized mechanical systems.

Example Paper:

Deep Generative Design: Integration of Topology Optimization and Generative Models

S. Oh, Y. Jung et al. (2018)

This study emphasizes the importance of incorporating deep learning into generative design research. The proposed AI-based deep generative design framework combines topology optimization and generative models, such as GANs, to generate aesthetically pleasing and engineering-performance-optimized design options.

By iteratively exploring new designs, the framework produces a large number of options based on limited previous design data. Anomaly detection helps assess the novelty of generated designs, facilitating informed decision-making.

The case study on 2D wheel design demonstrates the framework’s superiority in terms of aesthetics, diversity, and robustness compared to previous generative design approaches.


Generated Images the Proposed framework


2. Complex Geometries and 3D Models:

Generative design techniques excel in generating complex geometries and organic shapes that are difficult to conceive using traditional design approaches. By leveraging algorithms and computational power, generative design can create intricate and optimized designs that maximize functionality and performance. This is particularly valuable in industries such as aerospace, automotive, and product design, where unique and complex shapes can lead to improved performance and efficiency.

Example Paper:


Amin Heyrani Nobari, Wei Chen, Faez Ahmed (2021)

In this research paper, the authors test their model on a task that involves creating new 3D shapes without the need for manual design or expert knowledge. They use deep generative models, such as GANs or VAEs, which are types of artificial neural networks capable of learning complex patterns in data and generating realistic samples. These models are well-suited for the task because they can understand and reproduce the intricate details of 3D shapes, allowing for the creation of new and useful designs.

Example of Range-GAN outputs with different volume ratio and aspect ratio conditions


3. Weight Reduction and Material Optimization:

Paper: A generalized framework for microstructural optimization using neural networks

Saketh Sridhara, Aaditaya Chandrasekhar, Krishan Suresh (2022)

Weight reduction is a crucial consideration in mechanical engineering, as it directly impacts factors such as fuel efficiency, load capacity, and overall performance. Generative design can help achieve weight reduction by optimizing designs and identifying material-efficient solutions. By removing unnecessary material and optimizing structures based on load requirements, generative design can create lightweight designs without compromising strength and durability.

This paper introduces a framework for optimizing microstructures using neural networks. The framework allows for the optimization of various microstructural quantities, such as bulk, shear etc by choosing one as the objective while the others are constrained. By automating the calculation of sensitivities, the framework eliminates the need for manual derivations, reducing the potential for errors.

Minimizing Poisson ratio with mass constraint using different NNs.

4. Additive Manufacturing and 3D Printing:

Generative design and additive manufacturing, such as 3D printing, go hand in hand. This techniques allow for the creation of complex geometries and internal structures that are challenging to produce using traditional manufacturing methods. Generative design techniques can generate designs specifically tailored for additive manufacturing, maximizing the benefits of this manufacturing approach, such as reduced assembly, increased design freedom, and material efficiency.


Examples of GAN predictions



This paper introduces a deep learning-based framework for topology optimization (TO) in additive manufacturing (AM) design. The framework leverages a conditional generative adversarial network (cGAN) to learn similarities between TO runs and predict near-optimal designs. It takes into account both loading conditions and manufacturing process configurations. The cGAN predictions align well with true optima, and additional TO iterations can improve accuracy further. Manifold learning techniques provide insights, revealing that cGAN error increases with the distance between load and boundary conditions. Notably, the cGAN generated structures sometimes outperform TO-calculated ones, suggesting potential applications in generative design.



5. 2D shape Synthesis

Paper: Airfoil GAN: encoding and synthesizing airfoils for aerodynamic shape optimization

Yuyang Wang, Kenji Shimada, Amir Barati Farimani (2023)

2D shape synthesis using Generative Adversarial Networks (GANs) has gained significant attention for its ability to generate diverse and visually appealing shapes. By training a GAN on a dataset, the network can learn underlying patterns and variations to generate new shapes that resemble the training data. GAN-based 2D shape synthesis offers applications in concept generation, design exploration, and artistic creation. Challenges include ensuring control over shape generation and evaluating the quality of generated shapes. This blog series aims to explore GAN architectures, training methods, and techniques for controlling the shape generation process.

Airfoils synthesized by sampling on the feature domain of (a) PCA, (b) VAE, and (c) VAEGAN.


This work focuses on utilizing deep generative models in mechanical engineering to address challenges in shape design. In this case, particularly in the context of airfoil design. The authors adapt the VAEGAN framework, combining the strengths of Variational Autoencoder (VAE) and Generative Adversarial Network (GAN) architectures. The VAE component enables encoding and reconstruction of existing airfoil shapes. Whereas the GAN component generates high-quality novel airfoil designs through adversarial training. Additionally, the authors introduce a genetic algorithm (GA) to optimize the aerodynamic properties of the generated airfoils, such as lift and drag coefficients. The GA is applied to both the initial VAEGAN-generated airfoils and further optimized designs to achieve desired aerodynamic properties.




In this blog, you were introduced to the concept of generative models in Mechanical Engineering on a beginner level. Explaining revolutionary concept as we provide you with the latest research papers. This allows you to grasp the significance and implications of these models in the field. We highly recommend you to uncover how these advancements are shaping the future of engineering. This will help you gain a deeper understanding of their importance through comprehensive analysis and exploration.



* indicates required

Intuit Mailchimp

Sania Shujaat

A Mechanical Engineer with a keen interest in applying AI to revolutionize Mechanical Engineering.


Leave a Reply