Driving innovation: high-performance vehicle design for maximum efficiency with AI-powered design optimization

Multidisciplinary design optimization (MDO) has emerged as a transformative approach in high-performance automotive design, allowing engineers to simultaneously address competing requirements across various engineering disciplines, such as structural mechanics, thermodynamics, fluid dynamics, and crashworthiness. Emerging technologies like reduced order models (ROM) are further enhancing the capabilities of MDO by enabling real-time decision-making during early design stages. These advancements promise even greater reductions in development time and costs while pushing the boundaries of vehicle performance. This was the key theme of our recent automotive event where engineers from Ferrari and Dallara shared insights into the use of design optimization techniques to improve their racing vehicle designs while minimizing the need for expensive physical prototypes.

Ferrari relies on design optimization to improve F1 power unit efficiency

Enrico Gualtieri, Technical Director Power Unit at Ferrari SF, illustrated that the need to adapt quickly to changes, with a holistic approach to complexity – together with the focus on details and constant learning – is the core of design optimization approach at Ferrari F1 racing team. In this context, parametric design optimization technology plays a crucial role in optimizing 3D CFD port and intake simulations – F1 start scenario – and power unit energy flows.

Dallara speeds up fluid structure interaction analysis with reduced order models

Dallara is one of the most important companies specializing in the design, development and production of high-performance racing cars. Mattia Murari, Lead Engineer CFD Methodology at Dallara, explained why at the Aerodynamic Department they created an automatic simulation process to overcome typical multiphysics problems related to fluid structure interaction (FSI). By doing so, the interaction between aerodynamic loads and structural deformations can be captured, increasing awareness during the development phase and meeting specific performance requirements for racing cars:

• Low weight structures vs safety.
• Aerodynamic performance on real deformed geometries.
• Compliance with racing series regulations.
• Cost minimization.

FSI simulations are nowadays fundamental for an accurate and holistic prediction of the car performance. However, the classical direct coupling of the computational fluid dynamics (CFD) and structural FEM workflows can be very time consuming when dealing with the complexity of a full car geometry. In this context, how is it possible to reduce the time and computational costs associated with the FSI simulations? The aerodynamic team at Dallara is now exploring replacing costly CFD simulations with reduced order models (ROM) capable of rapidly predicting pressure distributions on deformed geometries. This decision led to the adoption of modeFRONTIER, leveraging its AI data-driven modeling technology for ROM. The new approach was successfully validated through a proof-of-concept application using the complete front wing geometry of a Dallara Open Wheels race car.

The development of a non-intrusive reduced-order model begins with the generation of high-fidelity simulation data, which serves as the training data set. In this project, 30 complete fluid-structure interaction cycles were performed, each characterized by different material properties of the front wing. Specifically, the finite element model was parameterized by varying the structural stiffness of three key components: the mainplane, the flap, and the fishplate. The FSI coupling was addressed through a partitioned two-way approach, whereby the fluid and structural domains were solved separately, but iteratively exchanged data to ensure physical consistency. This approach required an automated simulation workflow capable of managing the iterative loop between the fluid and structural solvers, collecting the required data from each domain at each step. The implementation of this automation was facilitated by modeFRONTIER, which enabled the seamless integration of the various simulation tools involved. These tools encompassed structural analysis, mesh morphing and computational fluid dynamics. The flexibility of the system is such that it facilitates effortless file exchanges and result tracking across the FSI loop.

The modeFRONTIER workflow shown in figure 1 highlights how the various stages of the FSI cycle are integrated. The process starts with a structural simulation where a pressure load, derived from a preliminary CFD analysis of the undeformed wing, is applied to the structure. The resulting deformation field becomes the input for the mesh morphing phase, where the front wing surface mesh is updated to reflect the structural deflection. The morphed CFD mesh is the input for the fluid dynamic simulation that produces a new pressure distribution. The updated pressure field and the deformed FEM mesh are used in the next structural iteration. This iterative process continues until convergence is achieved. Each complete cycle contributes to the creation of a representative dataset for the purpose of training the ROM. To this end both the deformation and the pressure field relative to the deformed geometry for each iteration are collected.

modeFRONTIER for ROM training and prediction

In order to build a ROM capable of predicting the pressure field as a function of the deformed geometry, it is required to characterize each deformation field through the geometry encoding. The data set solutions, consisting of the displacement field for each front wing design, are rearranged into a matrix that can be expressed by proper orthogonal decomposition (POD) as a linear combination of coefficients and modes. These modes are ranked based on their energy contribution, with the first few modes capturing most of the dataset’s variance. By truncating the basis and keeping only the most significant modes, we obtain a compact yet informative representation of the deformation behavior. The resulting POD coefficients serve as the encoded parameters, which are essential for the ROM training phase. This encoding step is carried out within modeFRONTIER using the ROM encode connector (as shown in figure 2). The pressure fields and the parameters derived from the encoding process, constitute in fact the fundamental data set for training the reduced model. The training process starts with applying POD again, this time to the pressure fields, in order to approximate the high-dimensional pressure data. The pressure POD coefficients can be interpreted as functions of the previously derived encoding parameters. This relationship allows us to train a regression model to predict the pressure POD coefficients for unseen wing geometries. Common regression techniques used in this context include radial basis functions, Gaussian processes, and neural networks. Once the coefficients are predicted, the pressure field can be reconstructed as a linear combination of the pressure POD modes, weighted by the predicted coefficients. This process can be entirely implemented in modeFRONTIER by using the ROM train connector, which also includes functionality for model quality validation. The dataset is split into a training and a validation set and the accuracy of the ROM is then assessed by comparing the predicted pressure fields with the original ones.

Once the ROM quality is deemed satisfactory, it can be deployed using the pPredict connector to efficiently estimate pressure fields for new wing configurations. The connector can be integrated in the original workflow, substituting the computationally intensive CFD simulation.

A new combination of wing material properties is created in order to test the ROM-FEM coupling. The predicted pressure fields for the two first iterations of the cycle are visible in figure 4, where it is possible to observe the good agreement among the pressure distribution computed by the CFD model and the one predicted by ROM.

This case study has demonstrated the feasibility of replacing the CFD process with a reduced-order model within a fluid-structure interaction cycle. The proposed approach leads to a significant gain in both computational time and efficiency: the ROM developed in modeFRONTIER is capable of replacing a 15-minute FSI cycle running on 350 cores with a near-instantaneous process on a single core. The next step is to extend the reduced-order model integration to the entire open-wheel race car, enabling the efficient simulation of the complete FSI process.