QMikeAI is a research institution in development, centred on high-performance weather forecasting using quantum reservoir computing, within a broader mission in quantum AI, HPC and computational science.
High performance is an objective to measure. This research preview demonstrates a research interface, not an operational forecasting service.
Investigate whether reservoir dynamics provide useful temporal features. Distinguish simulated circuits from quantum hardware measurements. Input encoding, washout, replay, shots and measurement all contribute to cost.
Investigate whether QRC improves held-out temperature prediction or corrects residual errors in numerical forecasts for a bounded set of Norwegian stations, relative to tuned classical methods under controlled information and resource budgets.
Proposed protocol. Not a registered experiment or a reported result.
The synthetic fixture qmikeai-synthetic-norway-v1 uses deterministic formulae at illustrative Oslo, Bergen and Tromsø coordinates. Both reference and experimental scenarios are synthetic. Neither is a forecast, an observation or an executed QRC model.
Air temperature is represented at 2 m and wind speed at 10 m. Coordinates use WGS84 longitude/latitude. Two synthetic cycles initialise on 1 October 2026, at 00:00 and 12:00 UTC, with illustrative issue times one hour later. Values span 0–72 hours in six-hour steps. One missing experimental wind value at Tromsø, lead 72, exercises unavailable-data handling.
A real-data study will require verified observation quality, original forecast availability, station/grid matching and archived lead coverage. MET Nordic documents forecasts spanning 58–64 hours. MEPS documents 61-hour ensemble and 66-hour control output. Analyses must not be substituted for forecasts that were available at issue time.
The implemented car case is an original Ferrari-inspired, open-wheel F1 concept in a three-dimensional channel. Gmsh provides the closed solid geometry and volume mesh. A Mojo lattice-Boltzmann solver computes the surrounding fluid on a voxelized Cartesian grid.
The homepage replays three-component velocity and density fields from an executed classical 3D lattice-Boltzmann simulation. Pressure is derived from the lattice equation of state. Instantaneous streamlines are integrated through each stored velocity field. The next research question is whether a hybrid QRC surrogate improves held-out prediction of reduced flow states compared with strong classical alternatives.
DNS resolves the relevant turbulent dynamics without a turbulence model. A fine mesh alone is insufficient. LES or RANS can be appropriate when labelled and justified for the chosen regime. A surrogate trained on DNS data is not itself a DNS solver.
The Float64 Mojo CPU car solver uses D3Q19 BGK collision, Guo forcing and halfway bounceback on a 96 × 48 × 32 grid. Its independent references are a three-dimensional diagonal shear wave and Poiseuille channel flow, each with refinement. The car case begins at rest and spans 2,000 lattice timesteps. Conservation, density, low-Mach and three-component numerical-momentum checks are recorded with the executed inputs. Gmsh tetrahedra define the car solid; the fluid solver uses a separate Cartesian lattice. Both representations can be inspected. The streamwise and spanwise periodic channel, stationary wheels and floor, low Reynolds number and coarse voxel boundary limit this to a computational reference. Full-scale aerodynamic forces and QRC performance have not been established. The earlier 2D wing benchmarks remain available. A quantum pressure-Poisson solver is a separate research track; the dedication-page car remains a display asset.
Hou et al., Physical Review Letters 136, 120602 (2026). An experimental spin-reservoir study motivates further time-series research. It does not establish QMikeAI forecasting performance or end-to-end computational advantage.
Pfeffer, Heyder and Schumacher, Physical Review Research 5, 043242 (2023). This study reconstructs reduced flow modes from DNS-derived data with performance similar to a classical reservoir. Reconstruction is distinct from autonomous forecasting or solving full-car DNS.
Dataset documentation guides future archive selection. This website contains no downloaded meteorological observations or archived forecasts. Archive inventory and quality checks remain necessary before fixing the study.