By Seth DeLand and Rashmi Gopala Rao
As engineers dive deeper into the integration of AI into vehicle systems and workflows, they encounter a series of formidable challenges. AI is no longer a standalone entity but a vital component that must seamlessly merge with adjacent systems. This necessitates the integration of AI algorithms within a simulation framework, a crucial step in comprehending their impact and functionality.
At a fundamental level, there exist three pivotal points of intersection between AI and simulation. Firstly, the quandary of inadequate data, which prompts the use of simulation to synthesise data that may be arduous or costly to obtain. Secondly, AI models stepping in as substitutes for complex, computationally demanding high-fidelity simulations. Lastly, the utilisation of AI models in control algorithms for embedded systems. In the following discussion, we will delve into how the synergy of simulation and AI offers innovative solutions to the challenges of development time, model reliability, and data quality.
Challenge: data for training and validating AI models
Gathering real-world data and curating it into a high-quality, well-organised dataset is a laborious and time-consuming process. Furthermore, the perpetual challenge lies in keeping the training set current with the latest field data. It’s worth noting that enhancing the training data can substantially boost the accuracy of AI models. Here, simulation emerges as a valuable ally, offering a range of advantages:
Computational simulation is a more cost-effective alternative to physical experiments.
Enables the simulation of scenarios that are either challenging or hazardous to replicate in the real world.
Provides access to internal states which helps debugging an AI model.
AI tools are adept at generating simulated data that mirrors real-world scenarios. This approach not only proves more economical but also empowers engineers to conduct data simulations within the same environment where they construct their AI models, promoting automation and seamless integration.
Challenge: approximating complex systems with AI
In the realm of designing algorithms that interact with physical systems, a simulation-based model of the system, often referred to as a “plant model,” plays a pivotal role in expediting the iterative design process. While these models offer high fidelity by being constructed from first principles, they entail considerable time investments in both building and simulating. This can make the evaluation of alternative designs a time-consuming endeavour.
Enter AI, which provides a solution to this quandary by approximating the high-fidelity model of the physical system with a reduced-order AI model. These AI-based models are computationally less demanding, reduce reliance on software dependencies, and mitigate the challenges faced during co-simulation with other engineering tools.
Recent advances in AI, such as neural Ordinary Differential Equations (ODEs), harmoniously blend AI training techniques with models infused with physics-based principles. These hybrid models prove invaluable when engineers aim to retain specific aspects of the physical system while approximating the remaining elements through a data-centric approach.
Challenge: AI for algorithm development
The domain of automotive engineering increasingly leverages simulations for algorithm development, including the creation of virtual sensors and the use of linear models and Kalman filters. However, these methods inherently exhibit limitations in capturing nonlinear behaviour. This is where AI steps in, offering the flexibility needed to model complex and nonlinear dynamics.
These AI models are deployed on performance and memory-constrained Electronic Control Units (ECUs), necessitating an evaluation of multiple models to strike the right trade-offs. Pioneering research takes this a step further by embracing reinforcement learning, not merely for learning estimators, but to comprehend and learn the entire control strategy. In addition to virtual sensors and reinforcement learning, AI algorithms find their place in embedded vision, audio processing, and signal analysis.
The Future of AI for Simulation
The future of AI and simulation in automotive engineering is nothing short of essential. These powerful tools are not just advantageous but absolutely indispensable. As the automotive industry marches towards increasingly sophisticated and complex models, the fusion of AI and simulation will become a defining factor in its success. The integration of techniques like synthetic data generation, reduced-order modelling, and embedded AI algorithms has already empowered engineers to streamline their workflows and slash development time. With the capability to develop and validate models accurately and affordably before introducing hardware into the equation, these methodologies will unquestionably dominate the landscape, setting a new standard for automotive engineering.
The authors Seth is Data Analytics Product Marketing Manager and Rashmi is Industry Marketing Manager at MathWorks.
Disclaimer: Views expressed are personal and do not reflect the official position or policy of Financial Express Online. Reproducing this content without permission is prohibited.
