Every experimental campaign begins with the same negotiation: how many points can we afford to take, how many sensors can we afford to install, and how much of the operating map can we afford to cover. Some of these answers are set by budget and schedule, some of these are set by mechanical constraints, and some of these are set by physics. The prioritisation is not always physics-first.
This paper sets out where Hendon fits: what the physics layer does, which facilities it applies to, and what changes commercially once it is in place.
The shape of the problem
A modern test facility is instrumented far more sparsely than the flow it is trying to characterise. A handful of pressure taps, a few thermocouples, a traverse that takes hours, a force balance — and from these, engineers are often tasked with reporting quantities that were not measured directly: integrated loads, efficiencies, flow-field averages, and coefficients.
Two things follow. First, the data-acquisition stack aggregates channels but does not reason about them: it will show you a drift, but not what caused it, and not which sensor to trust. Second, the uncertainty attached to the final number is typically a legacy error-propagation calculation — a single figure at the end of the chain, with no way to see which measurement dominates it or what would reduce it. The uncertainty budget describes the answer; it does not improve it.
What the physics layer does
Hendon builds one probabilistic model of the experiment. The governing physics for the asset — its geometry, the conservation laws, the constraints the flow must obey — defines a prior over possible fields. Your measurements condition it. What comes back is a posterior: the full field, the integrated quantities derived from it, and calibrated uncertainty on every one of them.
Because the prior carries the physics, the model does not need many measurements to become useful, and it does not produce fields that violate the conservation laws when the data are sparse. Because the priors are mesh-free, inference is fast enough to run alongside a campaign rather than after it — including for unsteady flows.
The practical consequence is that uncertainty becomes actionable. The same model that reports ±2σ on a coefficient can be asked where the next probe should go to reduce it most, or which channel is inconsistent with the rest of the physics.
Market one: wind tunnels
Aerospace, motorsport, automotive and defence tunnels share a cost structure: occupancy is the scarce resource, and instrumentation is expensive to install and intrusive to run. Runs are budgeted in tunnel-hours, and correlation between facilities — or between tunnel and CFD — is a recurring source of dispute.
The physics layer addresses this on three fronts. It recovers full fields and surface quantities from a handful of taps and a balance, so fewer points buy the same answer. It fills the operating map between the conditions actually run, so fewer conditions need to be swept. And because every result carries physics-consistent uncertainty, disagreements between facilities can be assessed against a stated confidence rather than argued about.
Market two: turbomachinery rigs
Compressor, turbine and combustor rigs are where sparse instrumentation bites hardest. Traverses are slow, probes are intrusive, and the quantities that matter commercially — stage efficiency, loss breakdowns, margin against a compliance limit — are inferred rather than measured.
For propulsion programmes, the same machinery supports uncertainty budgets and compliance-reliability calculations on quantities — efficiency, lift-to-drag, margin — that no sensor reports directly.
Market three: everything else on the test floor
The method is not specific to a facility type. It applies wherever a physical asset is instrumented sparsely and the answer of interest is a field or an integral of one: cascade rigs and water tunnels, engine and altitude test cells, hypersonic and shock facilities, thermal and structural rigs where the governing equations are known and the sensors are few.
The requirements are modest: the geometry, where the sensors sit, and the measurements themselves. Primary in-situ measurements — pressures and temperatures along the surface or flow path — carry most of the information; secondary channels such as force transducers, strain gauges and PIV fold in where they exist.
What changes commercially
Four things, in the order customers usually notice them:
- Fewer measurements for the same quality of result — traverses and surveys that consumed hours collapse to minutes.
- Fewer operating points, because the map between tested conditions is inferred rather than swept.
- Fewer expensive sensors, because quantities that would have required costly or intrusive instrumentation are recovered from what is already installed.
- Physics-aware uncertainty analysis, where the error bar is part of the model rather than a calculation bolted on afterwards — and where the model can say what would reduce it.
Working with us
Deployment starts with a rig, its geometry, and a historical campaign to validate against: we condition on a fraction of the data and check the recovered field against what was held out. From there, the same model runs live against streamed test-cell data or in post-test analysis.
If you run a wind tunnel, a test cell or a rig and want to see this on your own data, get in touch.