From 15 hours to one minute: How AI/ML is speeding up GM’s development

从15小时缩短至1分钟：AI/机器学习如何加速通用汽车的研发进程

When we met Sterling Anderson in 2024, he was the chief product officer of Aurora, the self-driving startup he cofounded in 2016 after several years at Tesla. Just over a year ago, though, Anderson decamped from the startup world for something a little more established, taking over as chief product officer at General Motors, the nation’s largest automaker. 2024年我们见到斯特林·安德森（Sterling Anderson）时，他还是自动驾驶初创公司Aurora的首席产品官，该公司由他在特斯拉工作多年后于2016年联合创立。然而，就在一年多前，安德森离开了初创圈，转投一家更为成熟的企业，出任美国最大汽车制造商——通用汽车（GM）的首席产品官。

Since then, he’s had a good view of how GM is entering what he calls the third epoch of engineering and design. “There was a time when humans looked at birds and were like, ‘OK, those wings seem to work pretty well. Let’s go and design something that looks like them.’” Anderson said, describing the first age of engineering. “And they just kind of iterated their way to something that was marginally feasible.” 自那时起，他亲眼见证了通用汽车如何进入他所称的“工程与设计的第三纪元”。“曾几何时，人类观察鸟类，心想：‘好吧，这些翅膀看起来效果不错，我们也去设计一些看起来像它们的东西吧。’”安德森在描述工程学的第一纪元时说道，“然后他们通过不断的迭代，最终做出了一些勉强可行的东西。”

The first few hundred years of inventing “was this era of highly empirical iterative design development and engineering,” he said. “And by that I mean humans largely started with what we know or had seen, built prototypes of something that kind of looked like it and maybe tweaked some things, hoping to make it perform better, tested it, iterated, and kind of went through this slow guess-and-check process until we got to something that marginally worked.” 他表示，最初的几百年发明史“是一个高度依赖经验、迭代式设计开发与工程的时代”。“我的意思是，人类很大程度上是从已知或见过的事物出发，制造出看起来相似的原型，或许再做些微调，希望能提升性能，然后进行测试、迭代，经历这种缓慢的‘猜测与验证’过程，直到最终得到一个勉强能用的产品。”

The second age began as computers became powerful enough to do some of the early work. “We started to see virtual development tools, in functionally specific ways, improve the work that people did so they didn’t have to go to empirical prototypical development,” Anderson said. “For instance, we started to see CFD [computational fluid dynamics] start to inform aero engineers,” he said. “We saw FEA [finite element analysis] inform structural engineers. We saw any number of other virtual tools… but the relay race that was development remained the same, which is to say design passed the baton to aero which passed the baton to structures, just always passed the baton back when they found something that the other guys had to fix.” 第二纪元始于计算机变得足够强大，能够承担部分早期工作。“我们开始看到虚拟开发工具在特定功能领域改善了人们的工作，使他们不必完全依赖经验性的原型开发，”安德森说。“例如，我们开始看到CFD（计算流体动力学）为航空工程师提供参考；我们看到FEA（有限元分析）为结构工程师提供参考。我们看到了许多其他虚拟工具……但研发过程中的‘接力赛’模式依然存在，也就是说，设计部门把接力棒交给空气动力学部门，再交给结构部门，一旦发现问题，又总是把接力棒传回去，让对方去修复。”

But Anderson’s world recently moved into its third epoch, “which is where GM has really been pushing, which is a collapse of those functions into a single broadly informed, largely probabilistic method for design, development and manufacturing of these assets,” he explained. And yes, by probabilistic, he means AI/machine learning. 但安德森的世界最近进入了第三纪元，“这正是通用汽车一直在推动的方向，即将这些职能整合为一种单一的、信息广泛的、很大程度上基于概率的方法，用于这些资产的设计、开发和制造，”他解释道。没错，他所说的“基于概率”，指的就是人工智能/机器学习。

Probable cause

概率因果

Using simulation for engineering work like CFD—versus using physical models in a physical wind tunnel—sped up that work, but the complexities of simulation mean it’s very computationally demanding in terms of resources and time. But you can teach a computer how to virtualize that analysis and then run multiple virtualizations in parallel using AI/ML; last month, we reported on just such an example, when IBM and the race car manufacturer Dallara published research showing how the approach produces data that’s well-correlated enough to use. 将仿真技术用于CFD等工程工作（而非在物理风洞中使用物理模型）确实加快了进度，但仿真的复杂性意味着它在资源和时间上对计算能力要求极高。不过，你可以教会计算机如何虚拟化这种分析，然后利用AI/机器学习并行运行多个虚拟化任务；上个月，我们报道过这样一个案例：IBM与赛车制造商Dallara发布的研究显示，这种方法产生的数据具有足够高的相关性，完全可以投入使用。

When you realize just how much faster these new tools are, it becomes extremely clear why GM is embracing them. “Our FEA runs that historically were 15 hours per run? They’re now one minute,” Anderson told me. Rather than setting up a simulation to run overnight and hoping nothing goes wrong, “when you run this thing in one minute, you’re just pumping through iterations at a much faster clip and you can run a much broader set of tests than you could ever have done before, just given the time available to you,” Anderson said. 当你意识到这些新工具的速度有多快时，通用汽车为何拥抱它们就显而易见了。“我们过去每次运行需要15小时的FEA任务，现在只需要1分钟，”安德森告诉我。与其设置一个仿真任务运行整晚并祈祷不出错，“当你能在1分钟内完成时，你就能以快得多的速度进行迭代，并且在有限的时间内，你可以运行比以往任何时候都广泛得多的测试，”安德森说。

But the reach of these new virtualization tools goes well beyond early engineering analyses and the domains of aerodynamics or structural design, reaching into GM’s other businesses: motorsport, energy and batteries, defense, and even its lunar program. 但这些新虚拟化工具的影响力远不止于早期的工程分析和空气动力学或结构设计领域，它们还延伸到了通用汽车的其他业务：赛车运动、能源与电池、国防，甚至是月球探测项目。

“We’re not using virtual tools just to check our work after we’ve done vehicle design, but we’re actually giving our engineers a virtual environment where they can simultaneously optimize the hardware and the software and inform hardware design or software design or vehicle performance in a way that nobody in the industry is doing, especially at the scale and the speed of what we’re doing,” said Jason Fischer, executive director of virtual integration engineering at GM. “我们使用虚拟工具不仅仅是为了在完成车辆设计后检查工作，而是实际上为工程师提供了一个虚拟环境，让他们可以同时优化硬件和软件，并以业内无人能及的方式为硬件设计、软件设计或车辆性能提供参考，特别是在我们所达到的规模和速度下，”通用汽车虚拟集成工程执行总监杰森·费舍尔（Jason Fischer）表示。

“The beauty of these virtual tools is our collaboration with our motorsports team with NASCAR and Formula One,” Fischer continued. “We co-develop a lot of these tools together and then we independently develop tools depending on who’s got the strength and the bandwidth between the organizations to do that. And as one outpaces the other, we actually sit down and we have a monthly technology transfer between motorsports, and I’ll say the production side of things to ensure that we’re all seeing the latest and greatest technology and using the latest techniques.” “这些虚拟工具的妙处在于我们与NASCAR和一级方程式（F1）赛车团队的合作，”费舍尔继续说道。“我们共同开发了许多此类工具，然后根据各组织的能力和资源独立开发其他工具。当一方领先时，我们会坐下来进行月度技术交流，在赛车部门和生产部门之间确保我们都能看到最新、最尖端的技术，并使用最新的工艺。”

Drive it before you build it

先虚拟驾驶，再制造实车

One example Anderson and Fischer walked me through was using virtualization to perform a handling test for a vehicle in development, specifically Consumer Reports’ avoidance test, where a car has to swerve at speed to avoid an obstacle. Instead of connecting all the various subcomponents of a car’s electronics on a test bench to see if they talk to each other without errors, GM now models all the sensors, electronic control units, domain controllers, and so on. 安德森和费舍尔向我展示的一个例子是，利用虚拟化技术对开发中的车辆进行操控测试，特别是《消费者报告》（Consumer Reports）的避障测试——即车辆必须在高速行驶中转向以避开障碍物。通用汽车现在不再需要将汽车电子设备的所有子组件连接到测试台上以检查它们是否能无误地通信，而是对所有的传感器、电子控制单元（ECU）、域控制器等进行建模。

“We actually have IP protection on how we’ve set this system up at General Motors where we can put together the vehicle behavior from a physics perspective,” Fischer said. “So [we can now] run vehicle performance, electronic control units, and software simultaneously in this virtual environment, and we can really open up our design space exploration. This allows us to actually change physical parameters and run thousands of designs of experiments to see how the control logic handles that,” Fischer said. “我们实际上拥有通用汽车建立该系统的知识产权，我们可以从物理角度整合车辆行为，”费舍尔说。“因此，[我们现在可以]在这个虚拟环境中同时运行车辆性能、电子控制单元和软件，我们真正拓宽了设计空间的探索。这使我们能够实际改变物理参数，并运行数千次实验设计，以观察控制逻辑如何处理这些变化，”费舍尔说。