Spaider-AI Platform
How it works
The Spaider-AI platform makes our trained neural networks and ai models directly available in the web browser. The user can a) Upload or generate a geometry, b) Train a model or select an available model, c) set the operating conditions, d) run the inference, e) analyze the predicted fields. Everything in seconds: no meshing, no solver setup, no queue.
Each vertical is a physics problem trained end to end: it is created starting from a simulation database, then a custom mathematcal model uses those data to generate a surrogate. Moreover, where the design space allows it, a generative model produces new geometries on demand. Four are available today.
Behind every prediction is a pipeline the platform exposes as five workflows, each a step you can inspect, repeat and change:
- Data — Simulations enter the platform here. Browse the library — every sample with its parameters, convergence checks, mesh statistics and result files — and curate the named datasets that everything downstream consumes.
- Preprocessing — Assign computed field and global variables as a network input or output, and choose the normalisation for each. The configuration is saved onto the dataset.
- Training — Set the architecture and hyperparameters, define the train/validation split, and launch the run. The finished checkpoint and its loss curves are registered as a trained model.
- Inference — Define a geometry, set the operating conditions, choose a model, and read the predicted surface and volume fields in a 3D viewer — in seconds, from CAD, with no mesh in between.
An additional workflow is also available for some verticals:
- Latent Geometry — Build or load a generative model that compresses a family of validated designs into a handful of latent variables, and decodes any point in that space back into a usable geometry.
The four run in sequence, from raw simulations to a trained surrogate. Latent Geometry feeds Inference directly, so a newly generated shape can be evaluated the moment it exists — which is what turns design exploration from a queue of simulations into a search over shapes.
Available Verticals
Transformer · thermal analysis
Cooling a power transformer is a conjugate heat-transfer problem, and a slow one to simulate. The model predicts surface temperature and pressure on the enclosure, and velocity, pressure and temperature on any plane through the internal fluid domain — directly from the CAD geometry, with no mesh.
Geometry is parametric: the enclosure is re-morphed at runtime from length, width and height, so an engineer can sweep the design space in the browser and watch the thermal response follow.
- Simulation database: 100 simulations · Ansys Fluent
- Solver: Steady, buoyant conjugate heat transfer
- Trained models: transformer derived architecture (surf & vol models)
- Predicted fields: temperature · pressure · velocity
- Geometry: Parametric on enclosure length, width, height
- BC Conditions: dissipated power, external temperature
- Simulation time: ~1.0 h on 16 cores / 64 GB ram
- Inference time: ~14 s
Moto · external aerodynamics
Developed with Aprilia Racing on the RS-GP24 windshield. The model predicts surface pressure and wall shear stress over the full vehicle and integrates them into drag and lift coefficients, the numbers an aerodynamicist actually decides on, plus velocity and pressure on any slice through the domain.
Geometry can come from the design library, a morphing parametrisation, an uploaded STL, or from a generative model that produces entirely new windshield shapes: a Folding-VAE with a 128-dim latent space and a learned flow-matching prior, so new designs can be sampled rather than drawn.
- Simulation database: 242 simulations · OpenFOAM v2512 · ~4M cells
- Solver: steady, incompressible
- Trained models: transformer derived architecture (surf & vol models)
- Predicted fields: pressure · wall shear stress · velocity · lift and drag
- Generative model: diffusion-based model
- BC Conditions: fixed boundary conditions
- Simulation time: ~4.5 h on 32 cores / 256 GB ram
- Inference time: ~12 s
Hull · VOF hydrodynamics
Volume-of-fluid simulations resolve the interface between water and air, which makes them among the most demanding analyses in marine design — and among the slowest. The model returns the volume fraction of water, so the wave pattern itself, together with pressure and wall shear stress over the hull.
Three generative models are available for this vertical, each a Folding-VAE over a 32-dimensional latent space, working on the amplified residual from the mean hull so that millimetre-scale variation survives the encoding.
- Simulation database: 648 simulations · Simcenter STAR-CCM+
- Solver: transient VOF
- Trained models: transformer derived architecture (surf models)
- Predicted fields: volume fraction of water · pressure · wall shear stress
- Generative models: Folding-VAE, 32-dim latent
- BC Conditions: ship speed
- Inference time: ~6 s
Compressor · transonic blade
Built on NASA Rotor 37, the transonic axial compressor rotor that has served as a turbomachinery CFD benchmark for decades — so results can be compared against a case the whole field already knows.
Blades are generated from 32 PCA shape modes, and the model predicts density, pressure and temperature over the blade surface, conditioned on rotational speed and outlet pressure. A design sweep that would take a cluster overnight becomes an interactive study.
- Simulation database: 1200 samples · Rotor37 (Safran, CC-BY-SA)
- Solver: elsA · steady compressible RANS
- Trained models: transformer derived architecture (surf & vol models)
- Predicted fields: density · pressure · temperature
- Generative models: 32 PCA shape modes
- BC Conditions: rotational speed · outlet pressure
- Inference time: ~6 s
