Automotive design is a complex, iterative, and highly proprietary process. Automotive manufacturers invest considerable time and resources in the design phase, refining 3D models through simulations before selecting the most promising prototypes to undergo rigorous physical testing. While the intricacies of check specifications and automotive aerodynamics are typically kept confidential. Significant breakthroughs in efficiency, akin to those achieved in gasoline efficacy or electric vehicle innovation, can manifest gradually and distinctively across companies.
According to MIT engineers, harnessing generative synthetic intelligence tools can accelerate the quest for innovative automotive designs at an exponential rate, as they rapidly sift through vast datasets and uncover hidden connections to produce a unique concept. While AI tools are available, the data they aim to analyze hasn’t been readily accessible, let alone consolidated into a centralized format.
However, the engineers have now made this valuable dataset publicly available for the first time. Dubbed DrivAerNet++, this dataset comprises over 8,000 automobile designs, primarily generated by engineers from the most common types of vehicles found globally today. Each design is visualized in three-dimensional form, incorporating data on the automobile’s aerodynamics – how air would ideally flow around a particular design, informed by sophisticated simulations of fluid dynamics conducted for each concept.
A credit score that accurately reflects your financial health? Yes!
Each of the dataset’s 8,000 designs exists in multiple forms, including mesh, level cloud, or a straightforward list of the design’s parameters and dimensions. As such, this versatile dataset can be leveraged by various AI architectures that can be fine-tuned to process information across distinct modalities.
The DrivAerNet++ dataset represents a significant milestone in the field of open-source automobile aerodynamics research, standing out as the most comprehensive and influential collection of data to date. Engineers aim to develop a comprehensive database of innovative automotive designs, featuring intricate aerodynamic analysis that can swiftly inform the development of AI models. These advancements enable rapid prototyping of innovative designs, potentially leading to more fuel-efficient vehicles and electric cars with extended ranges in a significantly shorter timeframe than the current industry standard.
“This dataset lays the muse for the following era of AI functions in engineering, selling environment friendly design processes, slicing R&D prices, and driving developments towards a extra sustainable automotive future,” says Mohamed Elrefaie, a mechanical engineering graduate scholar at MIT.
Elrefaie and his colleagues will present a paper detailing their brand-new dataset, and AI strategies that can be applied to it, at the NeurIPS conference in December. The paper’s co-authors comprise Faez Ahmed, an assistant professor of mechanical engineering at MIT, alongside Angela Dai, affiliate professor of computer science at the Technical University of Munich, and Florin Mazar from BETA CAE Systems.
Ahmed heads the Design Computation and Digital Engineering Lab (DeCoDE) at MIT, where his team investigates how artificial intelligence and machine learning tools can enhance the design of innovative engineering systems and products, including automotive applications.
“When designing an automobile, manufacturers often face significant upfront costs for the initial design process,” Ahmed notes, “making it impractical to make drastic changes between models.” “When dealing with larger datasets, it becomes clear which design variations are most effective. This allows you to train and refine machine-learning models quickly, ultimately yielding better designs.”
While velocity is crucial in automotive innovation, its significance has taken on new importance in recent times?
“Accordingly, the most opportune moment for propelling automotive advancements is now, as vehicles rank among the planet’s most significant polluters; prompt reductions in this area would significantly amplify our capacity to mitigate climate change,” Elrefaie explains.
Researchers exploring the cutting edge of automotive design found a crucial limitation: despite the existence of AI models capable of generating numerous designs, the actual available car data is limited in scope. Researchers had compiled preliminary datasets of simulated automotive designs, whereas automotive manufacturers rarely release detailed specifications on the specific models they test, refine, and ultimately produce.
The team aimed to bridge the knowledge gap, particularly in regards to an automobile’s aerodynamics, a crucial factor that significantly impacts the range of electric vehicles and the fuel efficiency of internal combustion engines? They discovered that the challenge lay in compiling a dataset comprising thousands of automobile designs, each meticulously detailed to accurately represent its function and type, without the benefit of physical testing or measurement to validate their performance.
Researchers constructed a dataset of automobile designs featuring accurate bodily representations of their aerodynamics by commencing with a set of baseline 3D models provided by Audi and BMW in 2014, respectively. Passenger vehicles can be broadly categorized into three distinct classes based on their rear profiles: fastbacks, characterized by a sloping rear finish; notches, featuring a subtle dip in the rear profile of sedans or coupes; and estatebacks, resembling station wagons with a flat, blunt rear end. The baseline fashions serve as a nexus between straightforward designs and more refined proprietary designs, offering a common starting point for various teams to experiment with innovative automotive design concepts.
The researchers applied a morphing process to each of the existing automotive designs in their study. The operation employed a systematic approach, introducing a subtle variation to each of the 26 parameters comprising the automobile design – including factors such as size, underbody configurations, windshield inclination, and wheel treads – before assigning it a distinct identifier and incorporating it into an ever-growing dataset. While the group ran an optimization algorithm in parallel, ensuring each novel design was uniquely distinct from any previously generated concept. To facilitate seamless translation, they transformed each 3D design into distinct modalities, ensuring that a single design could be portrayed as a mesh, point cloud, or inventory of dimensional specifications.
Researchers also employed advanced computational fluid dynamics simulations to predict airflow patterns around each generated automotive design. Subsequently, this endeavour yielded more than 8,000 unique, anatomically accurate 3D representations of automobiles, including the most prevalent types of passenger vehicles currently traversing our roads.
The researchers invested a staggering 3 million CPU hours on the MIT SuperCloud, yielding an astonishing 39 terabytes of data to comprise their comprehensive dataset. Estimates suggest that the entire printed collection of the Library of Congress would translate to approximately 10 terabytes of data.
Researchers can now leverage the dataset to train a specific AI model. A machine learning model could be trained on a subset of the dataset featuring automotive designs with intriguing aerodynamic properties. Within mere seconds, the AI can create a novel vehicle design with optimised aerodynamics, drawing upon insights gleaned from thousands of physically accurate designs within its vast dataset.
The researchers suggest that the dataset could potentially be utilized for an alternative objective. Following coaching on the dataset, designers can input a specific vehicle design into the AI model, which rapidly estimates the design’s aerodynamics, enabling the computation of fuel efficiency or electric range without incurring the costs associated with physically building and testing an actual car.
“What sets this dataset apart is its ability to enable the rapid development of generative AI models that can accomplish tasks in mere seconds, rather than hours,” Ahmed explains. “These fashion innovations can significantly reduce fuel consumption in internal combustion engines and boost the range of electric vehicles, ultimately paving the way for more sustainable, environmentally friendly transportation options.”
This work was partially supported by the German Education Exchange Service and the Massachusetts Institute of Technology’s (MIT) Department of Mechanical Engineering.
