One flow solve per candidate
Make several candidate shapes, compute the airflow around each one, and keep the best.
- The more areas you want to change, the more candidates you need
- Every candidate costs a full flow solve
AOX Product
OPTIMIZATIONSAME LINES, MADE BETTER.
AOX reshapes your own geometry to cut drag or add downforce, only where and only as far as you allow. AOX finds the direction; designers decide how far.
Flow, sensitivity, shape. Repeat.
AOX alternates between computing the airflow and moving the surface, refining the shape a little every cycle.
- 01
Compute the flow
The space around the shape is split into tiny cells, and the velocity and pressure in each cell give the drag and lift on the shape.
- 02
Find the sensitivity
For every point on the surface at once, AOX works out which way to move it, and how far, to improve the goal.
- 03
Move the surface
The surface moves a little along the sensitivity, inside your regions and limits, with its surroundings following smoothly.
- 04
Repeat
The new shape goes back to the flow step, for as many cycles as you set.
Every point, in one extra solve.
Most shape optimization builds candidate shapes and computes the flow for each of them. AOX uses the adjoint method instead.
Adjoint sensitivity
After the flow solve, one more computation shows how every point on the surface should move, and how much the goal improves.
- A sensitivity map, painted on the surface
- The shape follows the map, cycle after cycle
From upload to run.
Bring the files as they are and skip the mesh. Recommended values are filled in along the way.
- 01
Create a project
Select New Project and give it a name.
- 02
Upload and check the shape
Drop in one file, or several files split by part. Every upload is checked automatically to see whether the shape is fully closed.
- STL (.stl, .stlb), STEP (.step, .stp), IGES (.iges, .igs), OBJ, and FBX
- Repair (keep parts) fills small gaps and simple holes without changing the shape
- Force watertight always closes the shape, but sharp character lines may round off and intended openings such as grilles may close
- Half geometry for one side of a symmetric model, mirrored across the symmetry plane
Both repairs are automatic and may not fix every shape. Checking the source file before upload gives the most accurate results.
- 03
See it in 3D
The shape opens in 3D, with streamlines generated around it automatically.
- Turn the streamlines on or off and change the flow speed
- 04
Choose what may change
Paint the regions to optimize right on the model. Selections run smoothly across part boundaries.
- Brush paints detail, with an adjustable brush size
- Lasso encloses a large area at once
- New, add, and remove selection modes
- Or move whole parts into the optimized set
- 05
Set the wheels
For a car, mark the front and rear wheels. They turn at road speed over a moving road, for results closer to real driving.
- Available when the wheels are uploaded as separate files
- Without wheels, set the height above the ground instead
- 06
Limit the deformation
Set how far, and in which direction, the optimized regions may move. These limits protect the design intent and production constraints.
- Normal-displacement limits: how far the surface may move inward and outward
- Coordinate limits: minimum and maximum positions on X, Y, and Z, for example keeping a front lip at least 160 mm above the ground
- Or an offset from the baseline shape
Even when the shape should only grow outward, leave about 2 mm of inward room. With no room at all, the optimization may not progress.
- 07
Set the parameters
Choose Designer mode for the goal and the essentials, or Engineer mode for mesh, solver, and compute settings.
- AI settings analysis reviews the setup and suggests better-suited values
AI suggestions are a reference, not a rule.
Lower drag. Or more downforce.
Pick one objective. Then watch other values, or keep them inside a range while AOX optimizes.
Minimize drag coefficient (Cd)
Lower air resistance for better fuel or energy efficiency and a higher top speed.
- Monitor lift (Cl) to check that downforce does not drop
- Monitor or bound lift and lift balance
Minimize lift coefficient (Cl)
Less lift means more downforce, for stability at high speed.
- Monitor drag (Cd) to see how much air resistance rises
- Monitor or bound drag and lift balance
Monitoring only records a value and does not change the result. A flow constraint keeps a value between lower and upper bounds, set as a share of the baseline shape (100%). Each added constraint needs one more computation, which adds time and credits.
Hands off. Get back to designing.
Once you start the run, AOX takes it from there.
- 01
Meshed for you
The mesh is built automatically. There is nothing to set.
- 02
Runs on its own
Cycle after cycle with no input needed. Live flow and logs are there whenever you check in.
- 03
Done
The optimized shape is ready to review and download.
Credits are shown before every run.
See what changed. From every angle.
The shape, and the numbers behind it.
Download the baseline or optimized geometry, plus every figure.
Proven on real projects.
Porsche 992.1 GT3
Downforce-oriented front splitter and rear wing
- Total downforce
- +132%
- Total drag
- +28%
- Downforce-to-drag ratio
- +81%
Compared with the stock car. Starting geometry: ADRO aero kits.
BMW M4 aero parts by ADRO
Human-designed, AI-refined front splitter and rear wing
- Total downforce
- +24%
- Total drag
- +13%
- Downforce-to-drag ratio
- +9.8%
Compared with ADRO's original design.
Class 8 long-haul truck
Drag-oriented cab roof, fender, rear cab, and container
- Total drag
- −7.4%
- Fuel efficiency
- +3.4%
Fuel efficiency estimated with the VECTO model and ICCT reports.
Simple when you want it. Deep when you need it.
Designer mode asks only for the goal and the essentials and applies verified defaults to the rest. Engineer mode opens every setting below.
Designer mode
- Objective
- Minimize Cd or Cl, with monitoring and flow constraints
- Geometry
- Half geometry and symmetry plane
- Optimization cycles
- 5 to 10 recommended
Compute resources
- CPU cores for optimizing
- 384 (up to 384)
- CPU cores for meshing
- 128 (up to 128)
Solver
- Flow iterations per cycle
- 4,000
- Averaging starts at iteration
- 2,000
- Precision
- Mixed, or Double
Adjoint setup
- Preset
- Default, Balanced, or Conservative
- Adjoint iterations per cycle
- 4,000
Optimization
- Initial-cycle displacement limit
- 15 mm
- Deformation spread radius
- 150 mm
- Proximity radius
- 150 mm
- Post-processing
- Shape images and convergence graphs every cycle
Background grid and flow
- Grid cells X / Y / Z
- 278 / 40 / 40
- Freestream velocity
- 30 m/s (about 108 km/h)
Mesh
- Surface refinement
- 6
- High-curvature refinement
- 6
- Near refinement (within 0.1 m)
- 5
- Far refinement (within 0.3 m)
- 4
- Main layers
- 5
- Wheel layers
- 5
- Wheel expansion ratio
- 1.6
- Wheel final thickness
- 0.8
These defaults are verified in AOX. Radius and displacement values are tuned for a typical car, so adjust them when your shape is much smaller or larger.
What every setting controls.
Each Engineer setting tunes one stage of the cycle.
- 01
Mesh
The space around the shape is split into cells. The mesh deforms with the shape and is rebuilt automatically when its cells distort too much.
Background grid · Mesh · CPU cores for meshing
- 02
Flow
The airflow around the current shape and the goal value, such as drag or lift, are computed.
Flow iterations · Averaging start · Precision
- 03
Sensitivity
AOX computes how the goal changes when each point on the surface moves. Each flow constraint adds one more of these computations.
Adjoint preset · Adjoint iterations
- 04
Deformation
The surface moves along the sensitivity map, stays smooth, blends into fixed areas, and keeps inside your limits. The first cycle moves only a little.
Spread radius · Proximity radius · Limits · Initial-cycle limit
The improvement gets smaller with each cycle until it converges. The optimization mesh is lighter than a single precise analysis so it can cycle quickly, so read the result as the direction of improvement against the baseline and treat absolute forces as a reference.
Read the results with confidence.
The ideas behind every AOX setting and result.
- Pressure
- Air presses hard where it hits the front of a shape and less where it speeds up around it. These differences add up to drag and lift.
- Boundary layer
- A layer of air a few millimetres thick clings to the surface and speeds up sharply away from it. Skin friction and separation are both decided here.The layer settings in Engineer mode resolve this region.
- Separation
- Air follows gentle curves but breaks away where the shape turns sharply or ends, such as a roof edge, the end of the trunk, or a bumper corner.
- Wake
- Separated air leaves a low-pressure region of vortices behind the shape. High pressure in front and low pressure behind make pressure drag, most of a passenger car’s drag.Drag optimization usually delays separation and shrinks the wake. You can see the flow tidy up in the streamlines.
- Lift and downforce
- Pressure differences above and below a moving shape create a vertical force: lift when it raises the shape, downforce when it presses it down. More downforce means more grip at speed, usually with more drag.Lift (Cl) is the size of this force, and lift balance is how it splits between front and rear.
- Coefficients
- Cd and Cl divide the force by the speed and the size of the shape, so different cars and speeds can be compared directly.
Two engines, each built for its job.
Fast enough for every iteration.
Adjoint CFD on HPC
One extra solve shows how every surface point pushes drag or lift.
- Node-level shape control: each surface node moves on its own, inside your limits, so changes stay smooth
- A mesh that keeps up: integrated meshing refines the surface through every loop
- Example run: 384 cores, 5 cycles in 1 h 27 min
Lattice Boltzmann LES on GPU
No body-fitted mesh: the grid builds and refines itself where the flow needs it.
- Wall-modelled LES handles the thin surface layer, so a whole car fits on one GPU
- Custom GPU kernels give the same answer 5.6 × faster than our first build
- 187 million cells on one GPU; a car-body LES in 73 minutes, down from 409
Two steps to a final number.
Explore candidate shapes with optimization, then compare them or confirm the final one with an LES validation in Lab.
| Optimization | Validation (LES) | |
|---|---|---|
| Use it to | Find a design direction | Compare candidates or confirm final values |
| What it does | Reshapes the geometry toward your goal | Measures the forces on the shape as it is |
| You get | The improved direction and the reshaped geometry | Absolute Cd and Cl, forces in N, flow images and videos |
| Time | Minutes to over an hour, by size and settings | About 15 minutes to 3 hours, by grade |