15時間から1分へ：AI/MLがGMの開発プロセスを劇的に加速 From 15 hours to one minute: How AI/ML is speeding up GM's development

General Motors is making a striking claim: simulation tasks that once took 15 hours can now be completed in roughly one minute, thanks to the integration of artificial intelligence and machine learning into its core engineering workflows. The development, reported by Ars Technica, signals a broader transformation underway in how the auto industry designs and validates vehicles.

At the heart of the shift are two workhorses of engineering simulation: computational fluid dynamics (CFD) and finite element analysis (FEA). CFD models how air and heat flow around and through a vehicle, while FEA evaluates structural integrity under stress conditions like crashes. Both disciplines are indispensable, but they have historically demanded enormous compute time — sometimes many hours per run — which bottlenecks the iterative design process engineers depend on.

GM's approach appears to center on so-called surrogate models: AI systems trained on large libraries of existing simulation results that can then predict outcomes for new design configurations in near real-time. Rather than replacing physics-based solvers entirely, these models act as fast approximators, allowing engineers to explore a much wider design space before committing to full-resolution simulations. The productivity gains, if validated at scale, could meaningfully compress vehicle development timelines.

Digital twins are another pillar of GM's virtualization strategy. By creating dynamic virtual replicas of vehicles and manufacturing systems — continuously updated with real-world sensor data — the company can monitor performance, detect anomalies, and run predictive maintenance scenarios without taking physical assets offline. Applied to production lines, this could reduce quality-control costs and unplanned downtime, though the full operational benefits would depend on implementation maturity.

GM is far from alone in pursuing this direction. BMW has been running experiments with generative AI in design, Ford has explored AI-assisted crash simulation, and NVIDIA's Omniverse platform has become a common infrastructure layer for several automakers building digital twin environments. The competitive pressure to cut development costs and shrink time-to-market is universal, and AI is increasingly seen as the lever that makes both achievable simultaneously.

That said, surrogate models carry inherent limitations worth noting. Their accuracy is bounded by the quality and diversity of training data, meaning performance on novel or edge-case designs may degrade in ways that aren't immediately obvious. Establishing robust validation pipelines — ensuring AI-recommended designs still meet regulatory and safety standards — remains an open engineering challenge. There are also softer concerns about how heavy reliance on AI tools affects the long-term expertise of human engineers.

GM's reported results are compelling, and the trajectory is consistent with where the industry appears to be heading. Whether the one-minute benchmark holds across the full complexity of real-world vehicle programs is a question that will likely be answered over the next few development cycles. For now, the company's public commitment to AI-driven virtualization marks it as one of the more aggressive adopters among legacy automakers navigating a rapidly changing technology landscape.