The field of computational design in mechanical engineering typically starts with identifying a problem or goal, followed by reviewing relevant literature, resources, and systems that could be employed to tackle the issue. However, at the Design Computation and Digital Engineering (DeCoDE) Lab at MIT, the focus is on pushing the boundaries of possibility. Under the leadership of ABS Career Development Assistant Professor Faez Ahmed, alongside graduate student Amin Heyrani Nobari, the lab collaborates closely with the MIT-IBM Watson AI Lab to merge machine learning and generative AI techniques with physical modeling and engineering principles. Their aim is to confront design challenges and enhance the development of mechanical systems.
One notable project, termed Linkages, examines innovative ways to connect planar bars and joints in order to trace curved paths. In a recent discussion, Ahmed and Nobari provided insights into their ongoing research and its implications for the field.
Q: How is your team approaching mechanical engineering problems from an observational standpoint?
Ahmed: We are exploring how generative AI can be utilized in engineering applications. A fundamental challenge in this area lies in embedding precision within generative AI models. Specifically, we use self-supervised contrastive learning approaches to learn representations of linkages and curves in design, essentially capturing what the design looks like and its functionality. This concept aligns with automated discovery: Can AI algorithms generate new products?
A critical aspect of our research pertains to linkages and generative AI models. Precision is a key concern across these models, and our learnings from this research could resonate with other engineering fields. We aim to demonstrate a proof of concept that could eventually extend to various applications like ship and aircraft design, and even image generation tasks.
When designing linkages, we create mechanisms that trace user-defined paths, significantly increasing efficiency. For example, our method achieves a 28-fold reduction in error and speeds up the design process by a factor of 20 compared to previous techniques.
Q: Can you elaborate on your linkages method and how it compares with others?
Nobari: Our approach utilizes contrastive learning between mechanisms, which we represent as graphs. Each joint corresponds to a node in these graphs, embedding various features such as position and joint type, whether fixed or free.
We employ a neural network architecture designed to capture the essential characteristics of a mechanism’s kinematics. In this structure, one model processes curves to create embeddings while another computes embeddings of mechanism graphs. By interconnecting these two modalities via contrastive learning, we can identify new mechanisms while maintaining precision.
Once potential mechanisms are identified, we apply an optimization step to align them closely with target curves. If we achieve a near-precise mechanism configuration, we can use gradient-based optimization to fine-tune joint positions for optimal performance.
Our work provides examples, like tracing letters of the alphabet, which traditional methods struggle to accomplish. Conventional machine learning systems often grapple with limited configurations—typically analyzing mechanisms with just four or six bars. In contrast, we demonstrate that with a small number of joints, we can closely match complex curves. For instance, writing the letter "M" has been historically challenging, but our method makes it feasible.
Investigating generative models for graphs revealed their inherent challenges. Often training these models proves ineffective, especially when integrating continuous variables sensitive to kinematic parameters. Approaching the design task through optimization leads to a complex mixed-integer nonlinear problem, rendering traditional methods inadequate as we expand beyond a handful of joints.
The leading edge of deep learning approaches tends to rely on reinforcement learning. They use a trial-and-error method to assemble mechanisms randomly in hopes of creating a suitable design—often taking far longer than our approach, which accomplishes results in 75 seconds compared to the over 45 minutes required using reinforcement learning.
Q: What broader implications does advancing techniques like linkages hold for future human-AI co-design collaborations?
Ahmed: The primary application lies in the design of machines and mechanical systems, as we’ve demonstrated. Critically, we’re addressing a dual challenge of learning in both discrete and continuous domains—navigating how linkages interact (discretely connected or not) while varying node positions in a continuous fashion. Most machine learning methods focus either on discrete datasets, like in language processing, or continuous spaces, as with computer vision.
Understanding and addressing this integration opens up potential applications across various engineering fields, from metamaterials to complex network design and structure formulations.
Looking ahead, we plan to explore more complex mechanical systems that incorporate various elastic properties, as well as different component types. We are also investigating how to instill precision in large language models and implementing generative models that autonomously create designs rather than retrieving from datasets.
Nobari: In mechanical engineering, inverse kinematic synthesis serves as a crucial application, such as in car suspension systems, where establishing a specific motion path is essential. As we build on our framework, introducing compliant mechanisms—where continuous changes occur instead of rigid linkages—could prove particularly transformative.
This approach combines both combinatorial and continuous design parameters. Our work endeavors to bridge gaps in mechanical engineering through frameworks that automate various intricate processes, laying the groundwork toward significantly improved future design capabilities.
This research advances with the support of the MIT-IBM Watson AI Lab, aiming to redefine the possibilities of mechanical system design and foster future collaborations between human ingenuity and AI development.
