Between the chequered flag at Montreal on the evening of 24 May and first practice in Monte Carlo on Thursday 4 June, eleven Formula 1 teams have to do something that no other engineering organisation on earth does on this kind of timetable. They have to design, validate, manufacture, ship, fit and dial-in a substantively different aerodynamic package from the one they raced ten days earlier.
The new package will be the most extreme high-downforce specification of their year — different rear wing, different front wing, different floor edge geometry, in some cases different bargeboard treatment, all of it tuned for a circuit where average speed is the lowest of the season and where every kilogram of downforce a car can generate is worth lap time the straights are too short to recover.
That two-week window is the cleanest live test on the calendar of what has quietly become the most consequential capability in modern F1 engineering: the integrated stack of aerodynamics, machine learning, and high-performance computing that decides which configurations get developed, which never make it off the screen, and which arrive at the race in time to be a tenth quicker than the rivals’.
This piece is about that stack — what it actually does, why the 2026 regulatory environment has made it more important than at any point in F1 history, and the engineering profiles that the teams hiring most aggressively in this space are advertising for, right now, while the cars are being trucked from Heathrow to Nice.
Why the Two-Week Turnaround Is the Stress Test
A Monaco aero package is not a marginal update. It is a fundamentally different aerodynamic philosophy from the one used at Montreal a fortnight earlier. Montreal rewards low-drag setups — slim rear wings, efficient straight-line speed, active aero used aggressively down the back straight. Monaco rewards the opposite — maximum downforce regardless of drag penalty, aggressive front wing flaps, rear wing endplate geometry that prioritises corner load over straight-line efficiency. The wing surfaces involved are not the same as the ones the team raced two weeks ago, and on some teams the underfloor specification changes too.
That work has to be designed against the FIA’s Aerodynamic Testing Restrictions, which is where the engineering challenge becomes a computational one rather than a craftsmanship one. Under the current ATR regime, the baseline allowance for any two-month Aerodynamic Testing Period is 320 wind tunnel runs and 2,000 CFD items, with a sliding scale based on the prior season’s constructors’ championship position. McLaren, as 2025 champions, operate at 70% of baseline — 224 wind tunnel runs and 1,400 CFD items per period. Alpine and the new Cadillac entry, at the bottom of last year’s table, run at 115% — 368 runs and 2,300 CFD items. Every team is testing under a hard cap. And every CFD item spent on a Monaco wing is one not spent on developing the next chassis update for Silverstone.
This is the regulatory pressure that has, over the last five years, transformed how the best aerodynamics groups in F1 actually do their work. You cannot afford to run a CFD simulation on a configuration that will not work. You cannot afford to fill a wind tunnel slot with a part that is not going to be on the car. Every test has to be the right test. And the technology that decides what the right test is, increasingly, is a machine learning model.
The Stack: Aero, ML, HPC
The modern F1 aerodynamics workflow, at the front of the grid, looks something like this.
A senior aerodynamicist or principal aerodynamicist defines a design intent — for example, “find an additional five points of downforce at the rear of the car at Monaco’s representative ride heights and yaw angles, without exceeding our drag budget at peak speed in the tunnel section.” That intent gets parameterised into a design space, typically with somewhere between five and twenty geometric variables — wing chord lengths, twist distributions, endplate angles, gurney heights, flap positions.
At this point the workflow used to be: run CFD on a coarse sweep of the design space, identify the best two or three candidates, refine them, take the best one to the wind tunnel, manufacture. That worked when CFD was cheap. Under ATR, it no longer is — and the cost-per-test of wind tunnel time is even higher.
The current stack works differently. The design space is sampled by an active-learning algorithm — typically Bayesian optimisation built on a Gaussian process or neural network surrogate — that has been trained on the team’s accumulated CFD and wind tunnel history. The surrogate predicts the aerodynamic performance of any geometry in the design space in seconds, not hours. The optimisation algorithm uses the surrogate’s predictions, together with its quantified uncertainty, to choose which small number of full CFD runs are actually worth doing. Those CFD runs then update the surrogate, the design space narrows, and the cycle repeats. By the end of the loop, the team has converged on a small set of high-confidence candidate geometries that have been validated by real CFD — but using a tiny fraction of the ATR allocation that a brute-force sweep would have consumed.
The wind tunnel programme then validates the strongest CFD candidates. The track data feeds back into the surrogate, closing the correlation loop between simulated performance, tunnel performance, and real-world performance under race conditions. The same machine learning models that selected the candidates in the first place get steadily more accurate as the season progresses.
The HPC infrastructure underneath all of this is where the picture gets genuinely industrial. F1’s chief technical officer Pat Symonds described current CFD setups using over 550 million computational cells across 1,152 cores, with the latest generation moving towards more than double that. The teams running their CFD pipelines on the largest GPU/CPU clusters in the sport can iterate more design candidates per ATR period than the teams with smaller infrastructure. Cloud bursting — the ability to scale up to public-cloud HPC for peak workloads, on top of on-prem capacity — is now a genuine competitive variable.
This is the stack that, for the Monaco weekend, decides whether the rear wing on the car at FP1 is the right rear wing or the second-best one. And the engineers who build, maintain, and operate that stack are some of the most specifically recruited profiles in motorsport today.
What the Teams Are Actually Advertising For
Three observations from the current job market in F1 aerodynamics-meets-ML, drawn from publicly advertised vacancies at the time of writing:
The aero-ML role has become its own team, not an outsourced data science capability
The clearest signal of how seriously the leading teams take this stack is in their org structures. Mercedes-AMG Petronas advertise a Graduate Machine Learning Engineer to join their Aerodynamics Department, reporting to a Senior Data Scientist – Aerodynamics. That is a dedicated data science function sitting inside the aerodynamics group, not in a separate IT or innovation team. The work, as the role describes, is developing and supporting ML models to improve the aerodynamic development process, building automated pipelines for data gathering and wrangling, and using data from CFD, wind tunnel measurements and vehicle testing — the closed correlation loop in a single job description.
Williams have gone further. Their current Senior Aerodynamic AI Engineer vacancy at Grove describes the work as transforming aerodynamic data from CFD simulations, wind tunnel measurements, track telemetry, and geometric information into actionable insights using advanced AI and machine learning techniques, with explicit responsibility for leading the creation of advanced surrogate models for aerodynamic predictions, particularly in fluid dynamics applications. That is, almost word-for-word, the active-learning-meets-surrogate-modelling capability described above — and Williams are willing to publicly attach the words “Senior” and “Engineer” to the role, which signals real seniority and budget.
The wider AI infrastructure roles around aerodynamics are scaling fast
Williams alone are concurrently advertising a Senior AI Engineer role at Grove with explicit collaboration across Aero Development, Vehicle Dynamics and Performance, and Operations, and a separate AI Engineer position focused on real-time and offline ML applications, real-time simulations, anomaly detection and data-driven optimisation of vehicle dynamics. The technology stack named in the role descriptions — Python with TensorFlow, PyTorch and Scikit-learn, with C++ as a plus — is exactly the modern industrial ML stack, not a motorsport-bespoke environment.
Mercedes have a parallel Machine Learning Scientist role at Brackley, describing the work as research, design and development of machine learning models and methodologies for simulation acceleration, surrogate modelling, and performance prediction, owned end-to-end from problem definition through to deployment into engineering workflows. The line that stands out for anyone hiring in this space is the explicit ask for someone comfortable moving between research and engineering, taking ideas from early experimentation through to production systems used by engineers and trackside teams. That is a research-engineer hybrid — a profile that is rare in any sector and contested in every sector.
Below the headline roles, the broader aerodynamics organisation is hiring
The Williams Senior Aerodynamicist vacancy at Grove describes a role in the Aerodynamic Development team, working closely with the Principal Aerodynamicist and Project Leaders, helping to outperform rivals by maximising aerodynamic understanding. The Williams Experimental Analysis and Correlation Team Leader role supports aerodynamicists in the wind tunnel by pushing the boundaries of experimentation, analysis and correlation — the wind-tunnel side of the same correlation loop the AI Engineer is working on the simulation side. Aston Martin advertise a non-F1 aerodynamicist role in Brackley working on innovative non-F1 projects, developing aerodynamic concepts and advancing understanding of flow physics. The new Cadillac entry, via GM Performance Power Units in Mooresville, is hiring CFD engineers to support its F1 power unit programme.
What this paints is a sector-wide picture: the established teams are building dedicated aero-ML capabilities; the rebuilding teams (Williams, Aston Martin) are hiring across the full aerodynamics stack including ML; the new entrants (Cadillac, Audi) are building from scratch and competing for the same profile against the incumbents.
Why This Stack Matters at Monaco Specifically
Monaco is the race that tests whether the aero-ML pipeline actually delivers, for three reasons.
The two-week timeline compresses every weakness. Teams that have not invested in surrogate modelling have to rely on the team-of-aerodynamicists-running-CFD approach, which is bound by the number of CFD items the ATR allocates. Teams that have invested can sample the Monaco design space more aggressively, in less wall-clock time, with fewer ATR items consumed. Over a fortnight, those advantages compound.
The Monaco-specific aero kit is not on the chassis evolution path. A team’s main development trajectory points at the next regulatory generation or the second-half upgrade package. The Monaco kit is, by design, a one-weekend-a-year detour. The teams that can do that detour cheaply — without sacrificing CFD items and wind tunnel runs from their main development budget — are the teams whose ML pipelines are doing real work. The teams that can’t, run the detour and pay for it in the chassis development trajectory.
The 2026 active aero system has no historic dataset. The Z-mode and X-mode transitions of the 2026 active aero are new this year. There is no decade of CFD and wind tunnel data on how active aero behaves at Monaco’s specific yaw angles, ride heights, and speed range. The teams whose ML pipelines can extrapolate effectively into these undersampled regions of the design space — using physics-informed surrogates, geometric deep learning, or transfer learning from related conditions — are the teams that will arrive in Monte Carlo with a calibrated active aero strategy. The teams that cannot will be discovering active aero behaviour in FP1, FP2 and FP3, with the walls half a metre away.
What This Means for Hiring
Three implications, particularly for organisations outside F1 that should be paying attention.
The aero-ML engineer is the most contested profile in advanced engineering right now. Williams, Mercedes, the new entrants, and the rebuilding teams are all hiring concurrently. The candidates who fit the profile — strong fluid dynamics background plus production-grade ML engineering, or strong ML/scientific computing background plus enough fluid dynamics literacy to talk to aerodynamicists — are rare, mobile, and increasingly aware of the multiple bids on them. Mercedes’ job description for the Machine Learning Scientist role asks for engineers who can take ideas from early experimentation through to production systems. That is the same person an aerospace CFD group at Rolls-Royce wants, the same person an autonomous-vehicle simulation programme wants, and the same person a quant fund building physics-informed signal models wants. F1 is one bid among several, not the highest-paying.
The transfer story is strong from four specific places. Aerospace CFD groups (Airbus, Rolls-Royce, BAE Systems, GKN Aerospace, DSTL) carry exactly the right combination of fluid dynamics depth and computational rigour. Academic ML-for-physics research groups, particularly in geometric deep learning and physics-informed neural networks, produce candidates whose work directly maps to surrogate-model engineering. Industrial turbomachinery (Siemens Energy, GE Vernova) is an underused source of mid-career CFD engineers with corporate ML exposure. Cloud HPC providers (AWS HPC, Azure HPC, Google Cloud) increasingly produce engineers who have built scientific-computing pipelines at scale and want to apply them to a domain they care about. None of these candidates appear in motorsport LinkedIn searches. All of them are sourceable.
Retention is now the harder problem than acquisition. Once a team has built its aero-ML capability, the engineers who built it are visible, well-paid by motorsport standards, and well-paid by ML standards. The cross-sector portability of their skills means a senior ML engineer at an F1 team is one conversation away from a senior role at a hedge fund, an AV programme, or a national lab. The teams that have invested in this capability are the teams that now have to invest in keeping it. That is a different management challenge from the one F1 organisations are historically built to solve.
The Window
The week between Montreal and Monaco is the year’s most concentrated example of why this stack matters. By the time the chequered flag drops on Sunday 7 June, every team on the grid will know which of their Monaco upgrades worked and which didn’t. The teams whose ML pipelines selected the right configurations will quietly add to their lead in the underlying development race. The teams whose pipelines selected the wrong ones will go back to Brackley, Grove, Maranello, Milton Keynes, Enstone, and Hinwil and have a serious conversation about what their aerodynamics-meets-AI organisation should look like by 2027.
Those conversations are also hiring conversations. And for engineers in aerospace, academic ML research, industrial turbomachinery, and HPC who have been quietly curious about whether motorsport’s hype around AI is real, the answer this fortnight is: at the front of the grid, yes — and it is being advertised in the open.
| Working with Tiro
Tiro Associates places engineers across the aerodynamics, CFD, machine learning, and scientific computing disciplines that the modern F1 organisation depends on. We work across the F1 grid, the wider motorsport ecosystem, and the aerospace, defence, and industrial CFD/HPC sectors that increasingly compete for the same talent. If you are an engineer at the intersection of fluid dynamics and machine learning thinking about your next move — or a team building this capability and finding the conventional aerodynamics search does not return the right profiles — we would welcome a conversation |
