Explore artificial intelligence (AI) models trained to approximate engineering simulations, alongside the reference calculations and their limits.
A surrogate is a trained AI model that approximates a simulation. It can help engineers explore alternatives without repeating the full calculation each time. Its accuracy depends on the training examples, the physics and the conditions being tested.
Compare each model with its numerical reference, then check the intended physical application separately. An error reduction, an accuracy percentage and a speed ratio are different measurements. None is a blanket guarantee for every demo on this page.
For engineering and research teams exploring early designs. Review the evidence and agree the test case before using a prediction for a design or operating decision.
See benchmarks and limitsDescribe a design in everyday language and inspect a structured design specification (YAML) and a three-dimensional surface preview (STL). Geometry checks do not prove thermal performance; simulation preparation and physical checks are separate steps.
Natural language engineering request.
A checked, structured design specification.
A surface-geometry preview for inspection and further preparation.
I can help describe an air-insulated substation (AIS), where air helps insulate the electrical equipment. I check the equipment groups, transformer connections, spacing and unclear instructions before building a 3D layout. This is a design aid, not an electrical safety approval.
Compare the TUNet AI model with a saved detailed simulation for a fixed region of 128 × 128 × 128 cells. Names such as scenario_01 identify saved test cases, not live equipment. A new layout generated above is a geometry preview only; this prediction does not analyse that new layout.
Explore saved heat-flow calculations inside a sand store, then try the separate geometry builder for its container and heater pipes. The saved results are not live AI predictions. The older AI temperature model remains offline and is not qualified for physical design or control.
Open a recorded temperature field, move the cut through the store, then compare 60 seconds of heating with the next 60 seconds after the heater switches off. Every coloured cell comes from the saved calculation; moving the controls does not run a new prediction.
Loads two recorded 96 × 96 × 96 fields (about 7 MB total). No account needed.
Saved on 5 October 2026. This is a conservative heat-conduction reference: heat exchanged between neighbouring cells is accounted for on both sides. Ten analytical and numerical checks passed in the original study. The heater supplies 339.084 W for 60 s; the second snapshot follows a further 60 s with it off. Starting temperature: 300 K (26.85 °C).
This is not a live simulation, an AI result, a test of your custom geometry, or an experimental measurement. The older AI thermal model is not physically qualified for equipment design or control. No heat-exchanger output, usable heat delivery, household-size scaling, or safe operating limits are established here.
Download the numerical evidence summaryThis optional tool makes parts, not a thermal prediction. Your design does not change the saved temperature example above. The public preview supports up to 36 straight heater pipes.
This rule-based guide explains what to include in a sand-storage description. It is not a conversational AI or safety approval. The geometry generator separately checks dimensions, heater layout and wall clearance before making a preview.
When available, the research model predicts one saved case on a 96 × 96 × 96 grid. Average errors can hide large local differences; these results are not certified equipment temperatures or a safe operating recommendation.
Why it matters: dispatch policies help decide when stored thermal energy should be charged, held, or released under changing solar generation and demand conditions.
Replay saved winter and summer tests of a reinforcement-learning controller (AI trained through trial and feedback). The implementation uses IMPALA/V-trace, a learning method, and compares a solar-only baseline with solar plus sand storage. This is recorded simulation evidence, not live plant control or measured energy savings.
KYAMOS develops time-aware neural models for physical fields. Architecture, input ranges and validation differ between applications; a result from one thermal or flow case does not establish the accuracy of another.
Shock-tube tests study abrupt changes in pressure and density. A reproducible comparison needs the reference solution, test split, error definition and runtime conditions. Contact us for the evidence matching your proposed use.
For engineering and research teams comparing design options or exploring faster simulation workflows. Tell us about your reference calculations, input ranges and the accuracy your decision requires.
Pricing: quoted per project. Scope, computing needs, validation checks and support determine the quote. Contact us to agree access and deliverables before work starts.
A model must be checked for its intended application. The saved sand examples are demonstrations, not certified designs or plant-control advice.