TesboAI
TesboAI is our research-driven product line applying machine learning, deep learning, physics-informed methods and LLM agents to computational fluid dynamics — built on our self-developed GPU-native solver.
AI, built into the CFD loop
Most AI-for-CFD tools sit beside the solver and shuttle data across separate hardware. TesboAI is designed differently: because it is built on our own GPU-native solver, neural models are intended to run inside the solve loop on the same GPU, without cross-device data movement. From offline surrogates to physics-constrained, differentiable simulation, we treat AI as a native part of the solver rather than an afterthought.
AI is the umbrella, not the method
We use the word AI only as the field name. In practice we work across four distinct method classes, and we are explicit about which one applies.
Non-neural statistical and dimensionality-reduction methods such as POD.
Deep neural networks — neural operators, coordinate networks, tensor-basis closures, autoencoders.
Deep learning combined with numerical methods (adjoint, PDE residuals) — a category of its own, not pure ML.
Generative models for natural-language case setup and orchestration — not numerical models.
Where we push AI and CFD together
Grouped by method class — classical ML, deep learning, physics-informed, and LLM/agents — so it is always clear which technique is in play.
Deep Learning
Deep neural networks as fast surrogates and learned closures.
Classical ML + Reduced-Order
Statistical reduction (POD) combined with deep learning for compact, real-time models.
Physics-Informed & Differentiable
Learning with the governing equations and gradients in the loop.
LLM / Agents
Generative AI as the engineer's assistant around the solver.
On the horizon: physics foundation models pretrained across many geometries and operating conditions — a long-term direction that depends on large-scale simulation data.
Neural Operators
Neural operators learn the solution operator itself, so a model trained at one resolution can be evaluated at another. We use them for super-resolution (coarse to fine) and as fast field-to-field surrogates.
- Train coarse, infer fine — resolution independent
- Field-to-field super-resolution
- Generalizes to unseen resolutions
Implicit Neural Representations
Implicit neural representations encode a flow field as a continuous function of coordinates, decoupling the representation from any fixed mesh. They are memory-light and natural for complex 3D geometry.
- Mesh-free, continuous fields
- Sample any point at any resolution
- Analytically differentiable derived quantities
Data-Driven Turbulence Closure
Data-driven closures learn Reynolds-stress or subgrid terms from high-fidelity data, with physical invariances built into the network architecture so the model generalizes rather than memorizes.
- Tensor-basis networks with built-in invariances
- Trained on high-fidelity reference data
- Targets accuracy beyond baseline closures
Reduced-Order Models
Reduced-order models compress full simulations into a compact latent space and learn the latent dynamics, giving a near real-time, parametric surrogate for repeated queries.
- POD plus deep-learning latent dynamics
- Near real-time parametric evaluation
- Built for design sweeps and digital twins
Differentiable Physics & PINN
Making the solver differentiable lets gradients flow through the simulation itself. This enables training models against simulated trajectories and physics-constrained learning, working toward end-to-end differentiable CFD.
- Gradients propagate through the solve
- Physics-constrained, end-to-end training
- Foundation for posterior closure learning
Agentic CFD
Agentic CFD uses large language models and agents as an engineer's assistant — not as a numerical model. It translates natural language into a structured case, drives the solver, gates every result through verification oracles, and interprets the output in plain language. The physics is still solved by the solver; the LLM handles setup, orchestration and explanation.
- Natural language to set-up, run and interpretation
- Every result gated by verification oracles
- Designed to run fully on-premises and offline
Core Capabilities
Design-Speed Inference
Once trained, AI surrogates target near real-time evaluation, turning design iterations from hours into interactive cycles.
End-to-End Automation
From sampling to inference and visualization, the pipeline is designed to run without manual CFD setup.
Domain-Specific Models
Tailored models for different applications and physics, so outputs stay trustworthy and deployable in their target domain.
Physics-Consistent
Physics constraints such as divergence-free conditions are built into training to keep predictions physically coherent, not just visually plausible.
From surrogate to differentiable
Our work advances along increasing levels of integration between AI and the solver.
Design Input
Neural Network
Instant Results
Comparison with Traditional CFD
Traditional CFD
- Domain knowledge required
- Mesh generation and solver setup: hours to days
- Design space exploration is prohibitively expensive
AI‑Assisted CFD Design
- No manual setup; one‑click inference
- Per evaluation: milliseconds to seconds
- Interactive design cycles with instant feedback
Build with TesboAI
Exploring AI-assisted CFD design, or have a problem that needs a custom surrogate? We'd like to hear from you.