国际汽联与AWS合作，AI将参与未来F1赛车设计
国际汽联与亚马逊云科技（AWS）达成新合作，更强大的人工智能工具将在F1空气动力学规则制定中发挥更大作用。AWS将成为国际汽联官方人工智能、云计算、机器学习和云基础设施合作伙伴。
作为国际汽联与亚马逊云科技（AWS）新合作的一部分，更强大的人工智能工具将在F1空气动力学规则的制定中发挥更大作用。
AWS将成为国际汽联官方的人工智能、云计算、机器学习和云基础设施合作伙伴，其职责范围包括以更好的空气动力学模拟来塑造未来规则、协助改进国际汽联在全球的许可系统，以及更深入地融入国际汽联的内部运作方式。
与国际汽联合作开展计算流体力学（CFD）测试，旨在借助更新的硬件实现更快、更高保真度的模拟，并在模拟之上叠加人工智能加速手段，再由人工智能核查规则手册本身。
不过，生成式人工智能暂不会用于提出设计思路，因为当前的优先事项是用这些工具加速并支持人类工作。
在CFD工作的核心使用AWS，将其纳入国际汽联自己的虚拟风洞，并支持概念车的创建，这些概念车可以检查诸如跟随另一辆车时损失多少下压力，以及规则是否产生国际汽联想要的比赛。
虚拟过程涉及将汽车零件的3D数字图纸组装成模型，将汽车周围的空气分成数百万个微小单元，以便逐单元求解方程，然后后处理将原始数字转化为压力图、流动可视化和国际汽联可以分析的力。
国际汽联F1副技术总监多米尼克·哈洛解释说：“AWS硬件构建了那个简报，所以这是一个传统的CFD过程，使用OpenFOAM[一个开源软件包]和各种后处理工具运行。
“然后除此之外，CFD数据查看和操作的增强可能来自人工智能。
“因此，我们可能会考虑使用一些基于我们计算流体动力学（CFD）数据训练出的代理方法，来简化模拟过程，从而以更少的计算时间、更少的工程时间、更少的能源消耗和更少的资金，从整个流程中获得更高的生产力。”
尽管当前规则集才进入第一年，但针对可能于2030年推出的全新赛车和引擎的工作已经在进行中。
哈洛认为，风洞、计算流体动力学（CFD）和人工智能工作将相互关联，以制定这些规则，要么填补赛道数据的空白，要么单纯增强虚拟开发工作量。
完整的计算流体动力学（CFD）运行可能需要数小时，而哈洛提到的代理模型则是一种基于数百个过往CFD结果训练的人工智能，它学习形状与结果之间的关系，然后在几秒或几分钟内预测新形状的结果，而无需求解物理方程。
对国际汽联而言，这意味着要在相同时间内测试更多规则草案的变体——例如翼片形状和底板几何结构——但好处也延伸到了非空气动力学方面，国际汽联的性能部门正在研究电子设备和发动机法规。
“所以在性能方面，他们拥有计算机视觉[图像/视频分析]等工具，这些工具使我们能够将收集到的许多指标仪表化展示，其中一些用于安全，一些用于能量管理，一些用于合法性检查等等，”哈洛说。
“因此，我们收到的许多数据，可以利用人工智能更快地处理，用于趋势监测、发现模式，基本上，这是它擅长的事情之一。
“同样，对于动力单元，他们在动力单元[PU]小组中使用这类工具进行异常检测，因此神经网络可以表示你想要的任何连续函数。
如果你给它足够的神经元，你就能绘制出最复杂的波浪线，然后将这些波浪线与赛道上采集到的数据进行比较，从而看出某人的油压何时下降或其他情况。
神经网络是基本的人工智能结构，其规模足以对任何模式进行建模——而异常检测的工作原理是学习正常轨迹的样子，然后标记任何偏离该轨迹的情况。
这对于为国际汽联的分析提供信息更有用，因为人工智能工具本质上擅长根据对输入数据的查询来预测不同事件的结果。
构建能够边运行边学习的模型，还将使这一过程从一次性的规则制定转变为持续循环，因为这些工具会学习任何新法规在实际运行中的运作方式并不断演进。
当规则发生重大变化、以往数据无法准确反映新赛车时，这一点很有用；但需要注意的是，神经网络可能是所谓的黑箱，其内部决策过程对人类来说难以或无法追溯，因此国际汽联需要能够理解其工具为何标记了某些问题。
而且由于F1赛车如此复杂，目前人工智能工具在创造全新事物方面能力较弱，不过“提示工程”这一概念有可能在更远的将来成为现实。
“基于对这些[2026年]赛车的认识，已经开始为未来赛车研究新的空气动力学概念，”哈洛说。“还有一些我们可以尝试和调查的不同想法。
“我想人工智能的生成部分是我们未来可以想象的，但目前它是由人驱动、由工程师驱动的。”
“对你的赛车进行提示词工程，这显然全都关乎训练和模型的演进，而且要做这些事情肯定需要大量算力和大量资金。”
“我相信这一定会实现。”
国际汽联应当能够通过其与AWS的合作关系获益的一个领域，就是获取最新的图形处理器硬件。
计算流体动力学传统上在CPU上运行，但相关软件现在正被改写为在GPU上运行，速度往往快上数倍，并且能够同时处理海量简单计算，这既适合人工智能，也适合计算流体动力学。更好的硬件意味着更快、更准确的结果。
哈洛举了2022年地效时代初期让多支车队和国际汽联措手不及的海豚跳效应为例，说明这类现象有可能从中受益。
“因此，不必只运行一轮稳态计算流体力学模拟，你可以在非稳态条件下运行对时间敏感的模拟，”哈洛说。
“也许对于像海豚跳这样的现象，可以在它成为问题之前，或在全尺寸上看到它之前，获得更多理解。”
人工智能可能融入国际汽联流程的另一个领域，是使用更贴合用途的大语言模型，来支持车队对合法性问题的咨询以及对法规的压力测试。
这意味着要打造一个比 ChatGPT 或 Claude 类人工智能更私密、更安全的版本，以圈护知识产权，使机密材料留在国际汽联内部，但又能用来核查如此庞大、复杂的规则手册及其相互关联的条款。
用途之广还包括协助管理安全、能源管理和赛道合法性检查的仪表板，并支持日内瓦增强型远程运营中心，该中心在赛事期间远程支持赛事控制。
这可能包括协助处理赛道限制，因为当前的自动化系统会抛出数百起事件，需要人工核查，而人工智能可以提供一份按优先级排序的清单，或在做出判断前提示进行一系列检查，这样仍由人类做决定，但能更快、更一致地完成。
英文原文
More powerful AI tools will play a bigger part in the development of Formula 1's aerodynamic rules as part of a new partnership between the FIA and Amazon Web Services.
AWS will be the FIA's official artificial intelligence, cloud computing, machine learning and cloud infrastructure partner with the scope to shape future regulations with better aero simulations, assist with improvements to the FIA's licensing systems globally, and be more integrated in how the FIA operates internally.
Running CFD testing with AWS is intended to unlock faster, higher-fidelity simulation on newer hardware, with AI shortcuts layered on top of that simulation and then AI checking of the rulebook itself.
Generative AI will not be used to prompt design ideas yet, though, as the priority is to use the tools to accelerate and support human work.
Using AWS at the core of the CFD work brings it into the FIA's own virtual windtunnel and supports the creation of concept cars that can check things such as how much downforce is lost when following another car and whether the rules produce the racing the FIA wants.
The virtual process involved 3D digital drawings of car parts that are assembled into models, with the air around the car divided into millions of tiny cells so equations can be solved cell by cell, then post-processing turns raw numbers into pressure maps, flow visualisations and forces the FIA can analyse.
FIA deputy technical director for F1 Dominic Harlow explained: "AWS hardware builds that brief, so that's a traditional CFD process run using OpenFOAM [an open-source software package] and various post-processing tools.
"Then alongside that, the enhancements to CFD data viewing and manipulation could come from AI.
"So we might look at using some surrogate approaches trained off our CFD data to shortcut the simulations, to get more productivity out of the whole process for less compute time, less engineering time, less energy consumption and less money."
Though the current ruleset is only in its first year, work is already under way for what could be a 2030 introduction for brand new cars and engines.
Harlow believes windtunnel, CFD and AI work will be cross-correlated to develop these rules, either to fill in the gaps in track data or simply to enhance the virtual development workload.
A full CFD run can take hours whereas a surrogate model as Harlow mentions would be an AI trained on hundreds of past CFD results, which learns the relationship between shape and outcome, then predicts the result for a new shape in seconds or minutes without solving the physics.
For the FIA that means testing many more variations of a draft rule - such as wing shapes and floor geometries - in the same time but the benefits extend to the non-aero side with the FIA's performance department looking at electronics and engine regulations.
"So in performance, they've got tools like computer vision [image/video analysis], tools that enable us to dashboard a lot of metrics that we collect, some of the stuff for safety, some of the stuff for energy management, some of the stuff for legality checks and so on," said Harlow.
"So a lot of the data that's incoming to us, we can process more quickly using AI for trend monitoring, spotting patterns, basically, which is one of the things it's good at.
"Similarly for the power unit, they use those kinds of tools in the PU [power unit] group for anomaly detection, so a neural net can represent any continuous function you like.
"If you give it enough neurons, you can plot the most complex squiggly line and then compare those squiggly lines with what's coming off the track and you can see when somebody's oil pressure's dropped or whatever."
A neural net is the basic AI structure, which would be big enough to model any pattern - and anomaly detection works by learning what normal traces look like then flagging anything that deviates from that.
This is more useful for informing FIA analysis as the AI tools are good at essentially predicting an outcome for a different event based on interrogating the data fed into it.
Building the model to learn as it goes will also turn the process from one-off rule-making to a continuous loop as the tools learn how any new regulations work in action and keep evolving.
This is useful for big rule changes when previous data does not accurately reflect the new cars although the caveat is that neural nets can be so-called black boxes, where the internal decision-making process is difficult or impossible for humans to trace, so the FIA would need to be able to understand why its tools have flagged something.
And with F1 cars being so complex, AI tools would be weak at making something entirely new for now, although the idea of 'prompt engineering' is something that could materialise further down the line.
"Already there's work started on new aerodynamic concepts for future cars based on the learnings from these [2026] ones," said Harlow. “Some different ideas that we could try and investigate.
"I suppose the generative part of AI is something we could imagine in the future, but at the moment it's human-driven, engineer-driven.
"Prompt engineering your car, that's obviously all about the training and the evolution of the models, and it takes a lot of computing and a lot of money to try and do that stuff for sure.
"I'm sure it'll come."
One area the FIA should be able to gain in with its AWS partnership is accessing the latest GPU (graphics processing unit) hardware.
CFD traditionally ran on CPUs but the software is now being rewritten to run on GPUs, often several times faster and capable of handling huge numbers of simple calculations at once, which suits both AI and CFD. Better hardware means quicker, more accurate results.
Harlow gave the example of the porpoising effect that caught out several teams and the FIA at the start of the ground-effect era in 2022 as something that could potentially benefit from this.
"So instead of having to run just a round, steady-state CFD simulation, you can run a time-sensitive simulation in unsteady conditions," said Harlow.
"And perhaps with a phenomenon like porpoising, get a bit more of an understanding before it's a problem or before it's seen in full size."
Another area AI is likely to feed into the FIA's processes is using large language models that are more fit-for-purpose to support team enquiries into legalities and stress-testing of regulations.
This means making a more private, secure version than ChatGPT or Claude-type AIs to ring-fence intellectual property, so confidential material stays inside the FIA but can be used to check such a vast, complicated rulebook and its interrelated clauses.
The breadth of uses also includes assisting dashboards managing safety, energy management and legality checks at the track, and supporting the enhanced remote operations centre in Geneva that supports race control remotely during events.
This may include assisting with track limits, as the current automated system can throw up hundreds of incidents that a human has to check whereas AI could provide a prioritised list or prompt a series of checks before a judgement is made so humans still decide, but do so faster and more consistently.
The Race ·
内容来自公开信源，聚合内容版权归原作者与来源所有。