KYAMOS research portfolio
BRIDGE2HORIZON/0823 · Research & development

OPTI-WIND-AI

Explore wind-farm wakes, turbine layouts and terrain-aware modelling research

OPTI-WIND-AI brings together fast wind-farm calculations, learning-based turbine placement and a separate graphics-processor airflow research tool. Explore the saved demonstrations, understand their limits and discuss the next steps with the KYAMOS team.

Reported OPTI software checks
35
The November 2026 report records 35 passing OPTI tests. These are software checks; they were not rerun for this website update.
Turbines in the saved wake examples
4–24
Four predefined layouts use a flat-terrain engineering model. They are separate from the airflow research calculations.
Saved V100 solver time
~265 s
Each of two specific two-turbine runs completed 10,240 steps in about 4.4 minutes. This is a case-specific timing.

Evidence checked 20 November 2026. Numerical examples and software checks have the limits stated below.

What the project does

The project explores how turbine spacing and wakes affect wind-farm output. A wake is the slower air behind a turbine: like a boat leaving disturbed water, one turbine changes the flow reaching the next. The fast engineering model helps compare layouts under stated wind conditions.

A separate computational fluid dynamics (CFD) tool [airflow calculated on a grid] supports research into turbine forces and different cell sizes. It runs on an NVIDIA Tesla V100 graphics processing unit (GPU) [a chip for parallel calculations]. Synthetic hill inputs extend this work towards terrain-aware modelling; real-site accuracy remains to be established.

Learning-based placement is another research strand. The project compares learned choices with regular layouts, random search, greedy search and genetic search [breeding improved candidate layouts]. A useful comparison gives methods the same evaluation allowance and reports weaker outcomes as well as stronger ones.

The November software report records 248 passing tests across the two projects: a reported 207 SMART tests, 35 OPTI tests and six solar tests. The SMART component counts do not reconcile with its reported total. This is a reported software-check count, not 248 wind-flow validations or proof that every original project requirement has been accepted.

Multi-turbine wake examples

Four saved layouts show how downstream wakes reduce modelled power.

The demonstration evaluates predefined regular layouts using the Jensen engineering wake model [a simplified downstream wind-loss model]. It combines overlapping wakes and uses reference turbine power and thrust curves. The demonstration itself does not train a model or optimise these layouts.

All four examples use flat terrain, steady 8.0 m/s wind, identical 5 MW reference turbines, air density of 1.225 kg/m³ and 500 m grid spacing. Power is the calculated output at this wind speed, rather than rated capacity or a yearly energy forecast.

Saved engineering examples · 15 November 2026
TurbinesTotal powerAverage per turbineWake loss
44.88 MW1.22 MW31.1%
88.16 MW1.02 MW42.4%
1612.66 MW0.79 MW55.3%
2416.65 MW0.69 MW60.8%

Wake loss compares each layout with the output its turbines would produce without wakes under the same wind conditions. These are modelled examples, not measurements at an operating farm. The engineering model accepts up to 256 turbine positions; this demonstration covers up to 24. The separate inspected airflow tool supports one or two turbines.

Airflow research tool

Numerical airflow experiments complement the fast layout calculations.

  • D3Q19 lattice Boltzmann method [flow represented in 19 directions] with hybrid recursive regularisation (HRR) [a numerical collision treatment].
  • Two levels of fixed grid refinement [smaller cells in selected regions]. The saved cases use a 2:1 cell-size ratio.
  • Actuator-disk turbine representation [rotor force spread over a disk], with reference thrust and power curves.
  • Single-GPU CUDA execution [calculations run on an NVIDIA chip]. Independent cases can run on separate GPUs.
  • Saved numerical checks cover force calculations, mass and momentum accounting, positive values and selected repeated cases.

The airflow tool and the Jensen layout model answer different questions. A 24-turbine engineering result does not demonstrate 24-turbine CFD, and successful software checks do not establish measured turbine-power accuracy.

Terrain capability & remaining work

Synthetic hill inputs are working; realistic hill-flow prediction remains research work.

Development cases exercise Gaussian hills [smooth, rounded test hills], elevation-file input, slope forces and Manning roughness [a surface-drag setting]. A saved synthetic-terrain case completed 10,240 steps, and selected repeated cases produced identical saved final data.

These tests show that terrain inputs affect the calculation. They do not establish the correct acceleration of wind over a hill. The technical assessment records negligible hill speed-up and no completed comparison with measured terrain benchmarks.

  • Only synthetic Gaussian hills were tested in the inspected terrain evidence.
  • Real elevation-data accuracy and comparison with measured hill-flow references remain open.
  • The terrain-enabled airflow tool inspected here is limited to one or two turbines.
  • Terrain-aware placement remains a development goal; the multi-turbine results above use flat terrain.

A small reported speed-up factor, such as 1.0003, is not presented as validated hill-flow accuracy. Completing a long numerical run and matching a real terrain measurement are separate achievements.

Accuracy & cross-software comparisons

An exact mathematical example helps separate numerical accuracy from real-site accuracy.

The saved analytical test follows a smooth wave moving and fading through a periodic cube [air leaving one side returns opposite]. Its exact solution lets the team measure error as cell width halves. There are no turbines, terrain or walls in this test.

Uniform grids show approximately second-order convergence [halving cell width cuts error fourfold]. The joined-grid treatment in the same campaign did not pass that accuracy screen. Both outcomes are retained below.

Analytical wave test · observed convergence order · 12 November 2026
Grid arrangement16 to 32 cells32 to 64 cellsSecond-order screen
Uniform larger cells2.0002.000Passed
Uniform smaller cells2.0002.000Passed
Joined larger and smaller cells1.2760.977Failed

Selected analytical runs conserved mass and momentum close to numerical rounding limits. A literature comparison discusses OpenFOAM, but uses different test setups. It does not establish matched accuracy, faster performance or superiority over OpenFOAM. These numerical checks also do not establish turbulent wind-farm or terrain accuracy.

Learning-based placement

Compare a learning prototype with simple methods before choosing a layout.

The prototype reuses an IMPALA reinforcement-learning framework [workers learn from trial outcomes] with two workers. Bounded experiments use spacing constraints and an engineering wind evaluator, and compare methods using equal request budgets. These are controlled software experiments on constructed wind and layout conditions.

In a saved larger comparison, learning, random search and genetic search each received the same 512 requests per starting seed. All three methods found the same best modelled result in that constructed example. The saved learned policy missed the best layout in two of three seeds.

  • 35 reported OPTI checks cover placement rules, engineering wind calculations and annual-placement learning software.
  • Regular, random, greedy, genetic and learned approaches appear in the project comparison work; their budgets and test conditions matter when interpreting results.
  • A consistent advantage for learning has not been established.
  • Real-site terrain accuracy, annual farm energy and the final selected layout need separate qualification.

A better score in a simplified model is a useful research result. It is not a guaranteed increase in energy production at a real wind farm.

GPU timings & reproducibility

Saved timings describe specific cases and hardware, so comparisons stay useful.

The 14 November campaign used a Tesla V100 with 32 GB memory, two actuator-disk turbines, a 144 × 96 × 96 coarse grid with refinement, a refined-box setting of 102 and HRR sigma of 0.9. The two full runs each advanced 10,240 steps, representing 800 simulated seconds.

The executable is single-GPU. Running two independent simulations on two GPUs increases the number of cases completed together; it does not make one simulation use both GPUs. Available communication-library infrastructure does not establish a multi-GPU airflow implementation.

  • The two full runs recorded identical final coarse-grid airflow data. This repeat result applies to those saved cases and devices.
  • A separate 64-step cross-server check recorded 4.12 seconds wall time on one server and 4.80–5.71 seconds on another. This was a small single-GPU comparison, not a communication-library benchmark.
  • The campaign summary reports two independent full runs completing concurrently. Larger cluster throughput remains dependent on available hardware and workload.
Saved two-turbine V100 timing campaign · 14 November 2026
CaseStepsRecorded solver timeScope
Full run on V100 A10,240265.12 sOne independent GPU case
Full run on V100 B10,240265.19 sOne independent GPU case
Short run on V100 A1,02430.44 sSame stated case configuration

These are recorded development timings, not a fresh benchmark for this page. They exclude the complete user workflow and do not support a general speed-up claim against a CPU, another simulator or a production installation. No claim of identical results for every input or GPU is made.

Deliverables & current status

Software delivery and scientific acceptance are separate milestones.

D12

Graphical platform software

Handover documents and application artifacts prepared

Linux and Windows application records are present. Retained checks cover specific released versions; complete feature parity and intended-user acceptance are not established by the inspected records.

D13

LB-MRT-LES wind and terrain simulator

Software package prepared; scientific qualification remains open

Numerical research cases, timings and selected repeat checks are documented. Underlying status records retain incomplete original requirements, failed scientific checks, unvalidated terrain accuracy and no formal acceptance.

D14

Learning & optimisation methods review

Review document prepared

The methods document is present. Document availability is confirmed; authorised acceptance of the exact version is not established by the inspected evidence.

D15

Prototype AI model for terrain-aware optimisation

Prototype and engineering demonstrations prepared

Placement experiments and the 4/8/16/24-turbine demonstration are available. A reliable learning advantage, qualified real-terrain layout and formal acceptance remain unestablished.

D16

Management deliverable listed in the website brief

Status to confirm

The website brief requests this item, but an OPTI D16 document or matching deliverable-register entry was not found in the inspected package. The team needs to confirm its title, scope and delivery status.

D6 · Product forum

The project discussion area is prepared. Public posting opens after a moderator is assigned and topic, reply and moderation checks pass. Creating a page alone does not complete D6.

Open the project forum

Evidence & remaining work

This summary distinguishes saved development evidence from measured performance at a real installation. Supporting technical records can be discussed with the project team; confidential project files are not published here.

Software checks across both projects

Report dated 14–15 November 2026

The report records 248 passing software tests across both projects, including 35 OPTI tests. The SMART total remains unreconciled with its component counts. This website work checked saved records and did not rerun the scientific programs.

Multi-turbine engineering demonstration

Saved results dated 15 November 2026

Saved turbine-level outputs reproduce the rounded power and wake-loss table. The demonstration evaluates four fixed layouts under flat, steady-wind assumptions.

Analytical airflow accuracy

Campaign dated 12 November 2026

Uniform-grid convergence is approximately second order. The joined-grid campaign fails the declared second-order screen. This is a mathematical verification case, not outdoor validation.

Terrain development assessment

Technical records dated 14 November 2026

Synthetic hill inputs and long development runs are documented. The underlying assessment explicitly retains missing hill speed-up, measured terrain validation and a one/two-turbine airflow limit.

Timing and selected repeated cases

Saved V100 campaign dated 14 November 2026

Native run records support the stated 10,240-step timings. Two cases record identical final coarse-grid data; both retain an explicit absence of physical validation.

Delivery and acceptance records

Underlying status snapshots dated 13 November 2026

Delivery documents use broad completion language, while supporting status records retain partial scientific scope and unaccepted outputs. This page reports the narrower supported results.

What the results do not establish

  • The engineering wake examples assume flat terrain, uniform steady wind and identical turbines. They omit turbulence, time-varying wind, turbine blockage and electrical collection losses.
  • The wake expansion setting is nominal rather than calibrated to a site. Modelled megawatts are not a measured farm output or a forecast of annual revenue.
  • The 256-position engineering input limit does not establish 256-turbine validation. The multi-turbine demonstration and the one/two-turbine airflow tool have different scopes.
  • Real terrain, measured wind-farm accuracy, turbulent-flow accuracy and turbine-power validation remain open. Stable numerical runs and conservation checks do not establish those results.
  • Learning experiments use bounded, constructed conditions. They do not demonstrate a general advantage over simpler placement methods.
  • Saved numerical checks, document delivery and formal project acceptance are separate. The evidence does not support a blanket production-readiness or all-requirements-complete claim.

Talk with the project team

Project coordinator: Antonis Papadakis, KYAMOS LIMITED. Programme: BRIDGE2HORIZON/0823.

For desktop demonstration access, installation instructions or a project-specific evaluation, contact KYAMOS. Keep customer data and confidential proposals out of public discussions.