How AI is falling short at a time when the automotive industry needs it most

In this guest post, Gary Brotman, CEO of automotive-focused machine learning company Secondmind, sets out why current AI methods are not meeting the challenge of supporting the transition to EV.

In 2020 the EU set an ambitious goal of achieving a 30% reduction in carbon emissions from transport by 2030 as part of its pledge to bring about climate neutrality by 2050.

For an industry that is responsible for roughly one-fifth of global carbon dioxide (CO2) emissions, this is a herculean challenge, and one that is driving unparalleled transformation as automakers reimagine every aspect of the design, development and experience of driving a car.

Recent innovations in autonomous driving, 5G connectivity and electrification are positive signs that car makers are committed to this transformation and hastening the dawn of the software-defined, zero-emissions car. But it comes with a hefty price tag. It’s estimated that $50 billion will be required over the next five years just to meet sustainability targets.

From sustainable research and development (R&D) to sustainable manufacturing, the drive to carbon neutrality is putting pressure on automotive engineering organisations to identify and develop new practices across the value chain at a time when product complexity is also on the rise.

Artificial Intelligence (AI) has been heralded as the key ingredient for building the car of the future. Advanced Driver Assistance Systems (ADAS) and autonomous driving would remain in the realm of science fiction if it weren’t for AI techniques, like deep learning, which today must have vast amounts of data and compute available to ensure their efficacy.

But this “state of the art” AI has fallen short in effectively addressing complex high-dimensional engineering challenges. AI that flourishes in autonomous driving presents massive challenges in vehicle design and development where high-volumes of data create time and resource-intensive roadblocks in virtualising and optimising the design and calibration of increasingly more complex systems.

Machine learning must be practical and active

To address the complexity challenges in the early stages of design and development, automotive engineering organisations need more practical machine learning solutions that can operate efficiently at cloud-scale and deliver high-precision results with a fraction of data required by today’s design and calibration tools.

Today, the powertrain in the average passenger car has more than 70 electronic control units (ECUs) that govern the performance of the engine, electric motors, hybrid and other complex controls. With current approaches, experiments are time-consuming and expensive. The calibration of just one ECU takes many months to complete and costly prototypes are damaged and wasted in the calibration process. New machine learning techniques, such as Active Learning, can intelligently automate the experimentation process, resulting in significantly reduced time and energy needed to run large testing and calibration systems.

This increased efficiency translates into a reduction in raw material use and manufacturing costs because fewer costly prototypes are needed or at risk of being damaged.

Experts and algorithms produce better outcomes

Another shortcoming of AI is its inability to effectively leverage subject matter expertise in modelling and decision making that isn’t already captured in the data. Human expertise is critical to making optimal predictions and decisions especially in low-data scenarios. Active Learning that blends expert engineering knowledge with algorithmic intelligence will achieve better outcomes  than either could achieve alone. This leads to increased confidence and trust and will spur accelerated adoption of this new breed of software that is essential for sustainable engineering in the future.

Transition today to achieve the net-zero targets of tomorrow

Automotive engineering today is in a catch-22 situation. Traditional car makers are under huge pressure to invest in new zero-emissions innovations, satisfy growing consumer demands, and get to market faster. They must also continue to innovate in the design and development of conventional powertrain systems, including internal combustion engines, which will be necessary components in hybrid vehicles for another decade or more at the current rate of electric vehicle infrastructure investment, cost optimisation and consumer adoption.

To get the cleaner cars of tomorrow on the road faster, a more swift and aggressive transition is needed now. Broad and rapid adoption of data-efficient, human-centred machine learning capable of breaking through complexity barriers in design and development will help the automotive industry slash well-to-wheel emissions, accelerate the transition to carbon-neutral mobility, and achieve industry-wide business and environmental sustainability.

Data Center
Data Management