Automotive component design at Nifco using generative AI and diffusion models

This post was contributed by Collin Pittro (Nifco), Chris Bulger (Nifco), Vidyasagar Ananthan, Ryan Qi, Prateek Jain, Julia Lazdins from AWS.

In previous posts, we explored using generative AI models can help with conceptual external design of cars and ways that machine learning (ML) speeds up design iterations in external aerodynamics.

In this post, we want to share a joint journey with Nifco USA – a leading manufacturer of automotive plastic components – to explore the cutting-edge application of generative AI diffusion models for structural design exploration. By combining generative AI with services like Amazon SageMaker and AWS Batch, Nifco is aiming to disrupt the boundaries of traditional design methodologies, to enable the creation of lightweight, high-performance automotive components.

Background

Vehicles subject their components to a lot of structural loads over their life. Among these components, ribbing structures play a crucial role in providing stiffness and strength while minimizing weight and material usage. Traditionally, ribbing structures are designed with repeating 2D patterns, like trusses (diamond, honeycomb, square, etc.) featuring a uniform thickness to facilitate injection molding manufacturing processes.

Topology optimization is a well-established field in structural engineering that aims to optimize the structural layout within a given design space for specific loading scenarios, while adhering to prescribed performance constraints. However, classical topology optimization techniques, such as Solid Isotropic Material with Penalization (SIMP) algorithms or level-set methods, are computationally expensive involving many physics-based simulations and requiring heuristics and manual tuning of parameters to work effectively.

Recent advancements in generative AI – particularly in the realm of diffusion models – have demonstrated remarkable potential in text-to-vision and image-generation tasks.

Diffusion models involve the iterative refinement of a random noise signal towards a target output, based on latent space bias through natural language (text) or image embeddings, exhibiting an exceptional ability to capture and reproduce intricate patterns and structures.

This post explores the applicability of diffusion models trained on classic SIMP results with volume fractions, applied loads, and boundary conditions as inputs. The joint AWS and Nifco team studied these to see if they could produce structures which are both manufacturable and of sufficient strength to withstand complex loading conditions encountered during their operational lifetime.

Automotive design requirements

The automotive industry is consistently seeking better fuel efficiency for internal combustion engine (ICE) vehicles and greater range in the rapidly growing electric vehicle market. This means that the opportunity to reduce weight is a constant consideration when designing automotive parts. One effective method for reducing weight is to convert metal bracketing and support components into lighter weight plastic designs.

While plastic bracket designs are typically lighter weight than metal brackets, metals like steel naturally have greater stiffness which is desirable for withstanding the dynamic conditions automotive components experience.

When designing vehicle components, engineers must account for various dynamic conditions that can impact performance and durability. Some key examples include:

Resonant Frequency (like in Figure 1) – the maximum accelerations and equivalent forces experienced at resonant frequencies within a specified frequency range can put peak loading stresses on vehicle components. These resonant conditions must be carefully analyzed to ensure the structural integrity of the design.

Power spectral density (PSD) as depicted in Figure 2 – these random vibrations, captured through power spectral density vibration curves from road testing, represent the real-world loading conditions the vehicle will experience during normal operation. Designing for these cumulative vibrations is crucial for long-term reliability.

Figure 1 – Acceleration (in units of g-force) against frequencies encountered in automotive components

Figure 2 – Power Spectrum Density for vibrations encountered during road tests [3]

Filler or ribbing structures implemented into plastic bracket designs are typically repeating geometric patterns of simple shapes, like rectangular, triangular, or honeycomb structures – shapes familiar to human minds and nature. While these types of standard shapes can be effective, the questions encouraging our exploration of generative AI was: what if the best ribbing structures are novel enough beyond the basic building-block shapes that humans intuitively prefer? Could generative AI develop unique structures that further optimize a component’s design for strength, stiffness, and weight?

Nifco’s motivation for optimizing for strength and weight in plastic designs ultimately comes down to material usage. If we can use less plastic material but still achieve our strength requirements, several benefits follow. Reduced vehicle weight aids the energy efficiency of the vehicle. And less plastic material demand means lower overall manufacturing costs.

The initial goal was to analyze ribbing structures against 2D static loads within a basic overall shape we derived from an existing bracket that holds an Advanced Driving Assistance System (ADAS) front radar. Nifco created several 2D, 200 Newton static-loading cases using Finite Element Method (FEM) simulations (which we’ll explain later).

Figure 3 shows a honeycomb ribbing structure to mimic various loading situations that a similar bracket could encounter during the lifetime of the vehicle. The various loading conditions included horizontal, vertical, and angle forces with different fixed boundary conditions that represent how the bracket could be mounted to the vehicle body or pushed to an extreme situation.

The ideal microstructural ribbing pattern will perform well across all of these rather than specializing in one particular loading case, which we’ll capture in a stress metric which you’ll see soon.

In addition, injection molding plastics used in bracket design come in variety of material grades and properties. We used a copolymer POM plastic material for our investigation because it’s a relatively mid-stiffness plastic common in the automotive industry.

Figure 3 – Static loading scenarios for benchmarking ribbing structures with a representative automotive component

Computational architecture

The overall compute workflow of the solution is outlined in Figure 4. The key underlying technology is a diffusion model, TopoDiff, which we trained on several input-output pairs of applied loads and boundary conditions, and corresponding optimized 2D space-filling structures.

The input channels included: (a) a uniform channel with the desired volume fraction; (b) binary channels for x and for y indicating fixed boundary points; and (c) physical stress fields associated with point loads applied on the system computed using a FEM calculation (we’ve included further details later). The output channel has a binary variable corresponding to where material is present. To obtain training data, we used the open-source ToPy library for creating SIMP results as ground truth. For FEM calculations, we used the SolidsPy library to obtain physical fields.

The AWS team used a total of 180,000 ground truth samples for training, and we’ve shown the training progress in Figure 4 for the different neural networks: the main diffusion model, a conditioning regression model, and a floating material classifier – which took around 24 hours on a ml.g5.12xlarge. After the model was trained, we used the model to run inferences with desired volume fraction and applied loading conditions as the ML input channels, and then perturbations are applied randomly within the domain as small applied internal forces, to generate variant patterns. The magnitude of the perturbations, as well as the parameters used for the main diffusion model, determine the degree of variation in patterns, which is chosen empirically for this study.

We then converted the generated 2D patterns into a 3D CAD geometry of an angular bracket using linear patterning, vectorization, and extrusion. After meshing these geometries, we ran a Finite Element Method (FEM) calculation to compute the stresses and strains experienced by the bracket, from which we extracted the von Mises stress field of all the elements. We then averaged the peak 0.1% to derive the stress metric.

Figure 4 – Overall workflow combining machine learning, data processing and simulation steps

AWS also architected the corresponding workflow for the cloud, and we’ve shown that in Figure 5. The remote client machine running on the customer side logs in to an instance for visualization and processing (in this case, a g5.24xlarge instance), which is powerful enough to render 3D geometries. We used Amazon DCV to the remote desktop connection (but you could also use a regular RDP client). Another option is to provision an Amazon Workspaces desktop machine to connect to your Amazon Elastic Compute Cloud (Amazon EC2) instance.

We trained our ML models (diffusion, regression, and classification) using Amazon SageMaker, using the FEM and prior-topology optimization ground truth data. We stored the trained model weights in an Amazon Simple Storage Service (Amazon S3) bucket.

At inference time, we ran this model multiple times to obtain different geometries in parallel, running within a containerized GPU environment on AWS Batch (e.g. with g5 family instances), with slight perturbations on the loading conditions.

Once we obtained results, we convert them into a 3D geometry using AWS Lambda calls, and pushed them to another Amazon S3 bucket.

Finally, we used these as input geometries for FEM simulations which ran on CPU instances in AWS Batch.

All of these results were post-processed to obtain curves of stress-metric vs volume fraction, and visualized to understand where the highest stress concentration and failure points occurred within the geometry.

Figure 5 – Cloud architecture for combined workflow of ML training, inference and HPC FEM simulations

AI methodology

As mentioned earlier, AWS used the TopoDiff algorithm [1] to generate optimized topology designs for generative AI diffusion models – combined with guidance strategies – to perform topology optimization.

At its core, TopoDiff relies on a U-Net convolutional neural network (CNN)-based conditional diffusion model which learns to denoise an initial noise image into the final topology design.

1. The regressor guidance penalizes samples with high compliance by adjusting the diffusion output based on gradients from a compliance predictor.
2. The classifier guidance detects and penalizes any floating disconnected material regions using a binary classifier.

We’ve heuristically chosen tunable hyperparameters to control the influence of each guidance method at different noise levels to produce reasonable structures.

Using models trained on optimized structures using FEM and SIMP, we can perform inference using the following inputs: 2 channels for the boundary conditions (x & y directions); 2 channels for applied loads (x & y directions); and 3 channels for physical fields based on volume fraction, strain energy density and von Mises stress of a homogeneous plate. To convert the 2D microstructures into a repeating unit cell, we used 4-fold symmetry [2] (and in future we may explore of 8-fold and other symmetries, too).

In this study, we applied the loading conditions illustrated in Figure 3, with a small added noise to induce variations in the patterns. As a result, Figure 6 shows the geometric variations we obtained by adding noise and varying the volume fraction between 0.35-0.42, which is similar to existing bracket-ribbing structures.

From these designs the AWS and Nifco joint team shortlisted T1, T4, and T5 for their relatively even thickness of support structures (which makes them easier to manufacture) while avoiding sharp edges (which might cause stress concentrations).

Figure 6 – Overview of designs produced using the diffusion model based on a superposition of homogeneous loading conditions with randomized internal forces.

We patterned these designs and extruded them to create a 3D angled bracket design and added an additional perimeter support structure.

We’ve shown the results of this process for the considered benchmark topologies (honeycomb, trusses, and arcs) and the generated microstructures (T1, T4, T5), together with the corresponding chosen volume fractions, in Figure 7. We chose these to understand performance against two key metrics: structural strength through FEM calculations, and their overall volume/weight.

Figure 7 – 3D bracket designs considered for benchmarking through structural simulations using FEMFigure 7 – 3D bracket designs considered for benchmarking through structural simulations using FEM

Evaluation of designs through simulations

Nifco meshed these designs using Dassault Systems 3DS CATIA, and then performed high-fidelity structural simulations using the finite element method (FEM) with Abaqus – for each of the loading conditions. The team took several steps to standardize the simulation and result-evaluation, including:

1. Using unstructured meshes for these simulations, evaluated using stress convergence studies.
2. To create a metric of structural strength, we used the peak von Mises stress. To avoid mesh sensitivity issues due to skewed elements – particularly with complex shapes produced by the diffusion model – we averaged the peak 0.1% von Mises stress points, and used this instead of the peak von Mises stress.
3. Across each of the loading scenarios, we normalized the peak 0.1% von Mises stress by the structure with the highest stress (generally, the highest stress tended towards the Arc structure from Figure 6), and averaged across all the loading scenarios to obtain a Normalized Stress Ratio (NSR).

Intentionally, these structures are designed with different volume fractions ranging from 0.25 to 0.5 and the heavier structures (diamond truss and triangular truss), perform very well on the stress metrics. A summary and comparison of the simulation results are shown in Figure 8. For example, the generative AI structure T5 performed well compared to the triangular truss – and was approximately 9% lighter.

Overall, the performance of these designs is plotted in Figure 9 through a Pareto multi-objective curve, where the designs form a front. By traversing this curve, it’s possible to tailor the unit cell patterns to compromise between weight and strength. This approach allows for automating this design space exploration using Generative AI.

In future studies, AWS and Nifco want to tune the volume fraction further by changing the characteristic thickness of the bars – and the length scale of the repeating units – in both the classical truss patterns and the generative designs. This might give a wholistic picture of a Pareto optimal front – which we’ll talk more about in a future post. Similarly, understanding how the front behaves in multiple further dimensions, like dynamic stiffness and in the vibration spectrum would be very interesting from a structural design perspective.

Figure 8 – Stress predictions per loading for each structure, together with the normalized stress ratio (NSR) metric.

Figure 9 – Evaluating designs on a Pareto front across the two objectives NSR and volume fraction.

Industry outcomes

The successful application of generative AI diffusion models to optimize the design of automotive components could have multiple downstream impacts on overall profitability and efficiency across the automotive industry. Enhancing the strength of component structures while also reducing vehicle weight can improve fuel efficiency and electric vehicle range – both crucial in influencing consumer purchasing decisions.

Additionally, the AI approach balances performance with ease of manufacturing, considering factors like even thickness for injection molding. This streamlines production processes, reduces waste, and lowers manufacturing costs. As a result, companies can either offer more competitive pricing to consumers or improve their own profit margins (or both).

From an engineering standpoint, the automated design exploration enabled by the generative AI models allows engineers to rapidly generate and evaluate a wide range of design options, significantly speeding up the design iteration process. This improvement can reduce time to market for new products, and reduce overall research and development cost per part.

By using generative AI diffusion models for structural optimization, companies can unlock a wide range of benefits – from enhanced product performance and reduced weight, to streamlined production processes and faster design iterations.

This technology holds promise to transform the way engineers approach component design, ultimately delivering value to both businesses and end-consumers through improved competitiveness, profitability and sustainability. The underlying principles showcased here can be applied far beyond the automotive industry, opening up exciting new possibilities for the future of manufacturing.

Conclusion

Exploration structural design space through generative AI hold promise using diffusion models which optimize volume fraction and average normalized von Mises stress under various loading scenarios.

This enables the investigation of quasistatic structural properties – as we showed in this post – and opens the door to studying dynamic vibration characteristics next. Both are crucial considerations in the development of efficient structures.

The natural extension of this approach to three-dimensional structures leverages the power of geometry-informed neural networks and other emerging AI technologies in conjunction with simulation techniques.

By combining AI and simulations, designers can navigate the intricate design space, capturing the complexities of real-world structures while evaluating their performance under various loading scenarios. Ultimately, this multidisciplinary approach paves the way for the creation of innovative and efficient structures that meet the demands of diverse applications, from cutting-edge materials to large-scale architectural projects.

Customers who want to discuss their business case or any of the technical details in this post can reach out to AWS or Nifco – we’d be happy to talk about it more.

References:

[1] Mazé, François, and Faez Ahmed. “Diffusion models beat gans on topology optimization.” Proceedings of the AAAI conference on artificial intelligence. Vol. 37. No. 8. (2023).

[2] Ivarsson, Niklas, Mathias Wallin, and Daniel A. Tortorelli. “Topology optimization for designing periodic microstructures based on finite strain viscoplasticity.” Structural and Multidisciplinary Optimization 61.6 (2020): 2501-2521.

[3] Gordon, Timothy, and Zevi Bareket. “Vibration transmission from road surface features: Vehicle measurement and detection.” University of Michigan, Ann Arbor, Transportation Research Institute (2007).