From Dataset to Design Impact: How Neural Concept Set a New Benchmark on MIT’s DrivAerNet++
Neural Concept set a new state of the art on MIT’s DrivAerNet++ dataset, achieving the most accurate academic results to date across surface fields, volumetric velocity, and drag. The real breakthrough for industry lies beyond the benchmark: in less than two weeks, Neural Concept transformed 39 TB of CFD data into an end-to-end, production-ready workflow that supports the entire aerodynamic design process – from data preparation and model training to real-time deployment and collaborative design exploration. Powered by Microsoft Azure, this achievement shows how Engineering Intelligence can compress design cycles, cut costs, and give automotive OEMs faster, more confident decisions at every stage of vehicle development.
Introduction
In automotive design, external aerodynamics plays a critical role in shaping performance, energy efficiency, and cost. A small reduction in drag can translate into significant fuel savings or extended EV range over a vehicle’s lifetime. As design timelines get tighter, engineering teams are increasingly turning to data-driven tools to accelerate aerodynamic development.
Within this landscape, MIT’s DrivAerNet++ dataset has emerged as a widely used public benchmark. It combines realistic automotive geometry variations with rich flow information, which makes it a credible way to test whether a learning-based approach can capture the signals that matter to engineers. Evaluating on this dataset is not only a research exercise. It is a practical check on readiness for real-world engineering work.
Using Neural Concept, we trained and validated our Geometric Regressor on the DrivAerNet++ dataset, a geometry native model designed for engineering data. In 2 weeks, we achieved the most accurate predictions reported to date across surface fields, volumetric velocity, and scalar targets such as drag coefficient. The speed of this cycle comes from codified best practices executed by an automated platform.
Why Results on DrivAerNet++ Matters
DrivAerNet++ is the largest and most comprehensive multimodal dataset for aerodynamic car design, comprising 8,000 car geometries across fastback, notchback, and estate back variants and 39 TB of CFD data spanning surface fields, volumetric flow, and drag. Its scale and diversity make it a credible proving ground for AI surrogate models, and performance on this benchmark is considered state-of-the-art.
We built DrivAerNet++ as an academic foundation to accelerate automotive’s transformation in the AI era.
It’s thrilling to see it quickly picked up by companies like Neural Concept to redefine the speed of real industrial design cycles.
Dr. Faez Ahmed, Associate Professor, Department of Mechanical Engineering, Massachusetts Institute of Technology (MIT)
DrivAerNet++ provides a common ground for benchmarking and enabling the exchange of insights and transfer of innovation across the global engineering community. By overcoming the typical confidentiality constraints limiting collaborations on concrete product development challenges, it unlocks tremendous innovation opportunities in both academic and industrial contexts.
In the sections that follow, we share the results obtained by applying Neural Concept’s aerodynamics predictive model training template to Drivaernet++. Our work demonstrates how an industry-ready platform turns a public benchmark into a repeatable workflow for real design decisions. Beyond model innovation, real engineering value comes from what happens after: deploying those models in collaborative, auditable workflows that evolve across teams and projects.
Defining A New State of the Art on the DrivAerNet++ Dataset
Under the official evaluation protocol and train/test split, our Geometric Regressor delivers the most accurate predictions across all measured outputs on the dataset to date. For baselines we use metrics from the public leaderboard, and report Mean Squared Error (MSE) and Mean Absolute Error (MAE), where lower values indicate closer agreement with the CFD ground truth. To keep comparisons fair, the model is trained on geometry alone without the dataset’s parametric inputs.
In the next subsections, we break down the results and show representative examples.
1. Surface Pressure Prediction
Neural Concept’s model attains the most accurate predictions for both MSE and MAE on the official split, outperforming all published methods, including GAOT.
Below, we present a quantitative assessment of our model’s prediction for this physical quantity.
The above qualitative example supports the metrics - the predictions from our surrogate model are remarkably close to the CFD ground truth. By accurately predicting surface pressure, our model reveals to engineers where high- or low-pressure zones are forming and how they contribute to drag.
2. Wall Shear Stress Prediction
Our Geometric Regressor also delivers top tier results and outperforms all competing methods in predicting wall shear stress on the car surface. Reliable wall shear stress prediction enables engineers to identify frictional effects and understand whether the flow is staying attached or separating, critical for managing drag and stability.
Across both surface fields, pressure and wall shear stress, our model achieves the lowest MSE and MAE by a clear margin over all published methods. Even against recent, high quality academic baselines, it sets a new state of the art in predictive performance.
3. Volumetric Velocity Field Prediction
In line with common practice for external aerodynamics datasets, DrivaerNet++ provides 3D velocity fields in the domain around the vehicle. Using the same Geometric Regressor architecture, we again achieve SOTA results, this time for cutting MSE by more than 50 percent compared to TripNet:
Accurate velocity field prediction gives a full picture of the flow structure in the volume and proves useful for analyzing wake stability.
The illustration below shows the velocity magnitude for two test samples. Note that only a single 2D slice of the 3D volumetric domain is shown here, focusing on the wake region behind the car. In practice, the network predicts velocity at any location within the full 3D domain, not just on this slice.
4. Drag Coefficient Prediction
Drag coefficient (Cd) is the headline number in aerodynamics, as reducing it directly translates to lower fuel consumption in combustion vehicles and increased range in electric vehicles.
Using the same underlying architecture, our model achieves state-of-the-art performance in Cd prediction, outperforming the previous best approach by a significant margin.
On the official split, our model shows tight agreement with CFD (R² of 0.978) across the test set, which is sufficient for early design screening where engineers need to rank variants confidently and spot meaningful gains without running full simulations for every change.
5. Compute Efficiency
In this section we compare compute characteristics of our model and training process with previous benchmark leaders GAOT. Like GAOT, our model was trained on 4 x A100 GPUs, hosted on Azure.
Training completed in 24 hours, with the best model obtained after 16 hours. Although the training duration for GAOT is not reported, the similarity in GPU type and configuration suggests both approaches require a comparable order of magnitude in compute. The gains come from architecture and workflow rather than more hardware. The final model is compact. Real-time predictions can be served on a single 16 GB GPU for industrial use.
In the following section, we quantify how the results translate into measurable gains in car development programs.
Industry Impact
Model accuracy alone is necessary, but not sufficient for industrial impact. Transformative gains at scale and over time are only revealed once high-performing models are deployed into maintainable, and repeatable workflows across organizations.
Customers using Neural Concept’s platform have achieved:
- 30% shorter design cycles
- $20M in savings on a 100,000-unit vehicle program
These outcomes fundamentally result from a transformed, systematic approach to design, unlocking better and faster data-driven decisions. The Design Lab interface, described in the next section, is at the core of this transformation.
From Prediction to Production with The Design Lab
Within Neural Concept, validated geometry and physics models can be instantly deployed into the Design Lab interface. The Design Lab offers a collaborative environment where aerodynamicists and designers explore tradeoffs, assisted in real-time by AI copilots providing performance feedback and geometric improvement suggestions. Through live KPI updates, this copilot experience effectively reconnects aerodynamics analysis with the ever-increasing pace of the vehicle design process.
As part of our lasting technology partnership with NVIDIA, the Design Lab directly embeds Omniverse, providing high-fidelity renderings bridging the gap between designer expectations and aerodynamic post-processing capabilities.
This is Engineering Intelligence in practice; not just a model, but a co-pilot experience that delivers real time physics prediction embedded in a visual design environment, guided by the expertise of aerodynamicists and designers.
A platform supporting long-lasting impact at scale
The concrete outcomes outlined above were achieved through industrial deployment at the level of whole teams or departments spanning multiple car programs over time. Beyond individual model accuracies, such a scenario requires solid practices to guarantee the system supports effective collaboration and consistently meets the requirements of the development process over time.
The workflow described in this article embodies such practices, built on years of experience working with car OEMs across the world. It showcases how the modularity of Neural Concept enables engineering organizations to combine (1) end-to-end templates that have delivered proven AI value in production1 , (2) core technological components2 and (3) an orchestration and workflow management environment. It takes all three layers carefully assembled to deliver on the promises of AI at an industrial scale.
In short, model accuracy is only the beginning. What sets Neural Concept apart is the ability to turn those models into enterprise-grade systems — reliable, auditable, evolvable — deployed in the tools where engineers actually work.
This workflow is being made available to all customers as part of the next Neural Concept Community Kit release, as a starting point for their own industrial workflows in automotive and beyond. It will be maintained and benchmarked continuously - in the same way top engineering teams use the platform to deploy and maintain their own workflows in production. This ensures it stays state-of-the-art and empowers our users with the latest best practices as part of every new release of the platform.
To learn more about Neural Concept’s capabilities and workflow, apply to our Spark Session program.
Conclusion
Neural Concept’s Geometric Regressor achieves state of the art results for surface, volumetric and scalar aerodynamics predictions. Beyond pure performance and accuracy, it took just 2 weeks of work to apply a full end-to-end pipeline including data preparation, model training and deployment into a production-ready interactive interface, setting a new standard for how fast AI workflows can be built and deployed to accelerate industrial design cycles.
The same features that unlocked this breakthrough in Neural Concept, are enabling leading engineering organizations such as Visa Cash App Racing Bulls and Subaru to shape their AI future. By removing the guesswork and capitalizing on years of proven best practices, these teams focus on building long term differentiated value through the biggest technology revolution of our time.
Acknowledgments
Our work built on the MIT-developed DrivAerNet++ dataset, trained on Neural Concept deployed as SaaS on Azure with dedicated compute resources provided by Microsoft, and integrated with NVIDIA Omniverse for high fidelity 3D rendering.
Access Our Technical White Paper
