Live Demo

Surrogate AI

Multiphysics solvers and time-aware neural surrogates with validated <1% error against full HPC simulation.

What Is Surrogate AI?

Surrogate AI replaces slow, high-fidelity numerical simulations with trained neural models that can predict simulation results in milliseconds. KYAMOS trains these models using validated HPC simulation data so engineers can explore design alternatives and optimize parameters without running a full classical solver every time.

Validation & Benchmarks

KYAMOS validates surrogate AI models against full classical solvers and benchmark simulation cases. Accuracy and speedup metrics are based on comparisons between neural surrogate predictions and high-fidelity numerical simulation outputs.

<1%
MSE vs Classical
10,000×
Inference Speedup
98.94%
MSE Reduction

Live Geometry Generation

The live geometry module converts natural language engineering requirements into structured AIS YAML and 3D geometry previews. The generated geometry can be used as input mesh for thermal CFD and downstream simulation workflows.

Input

Natural language engineering request.

Intermediate

Validated AIS YAML.

Output

STL geometry / simulation-ready mesh.

Design AdvisorClarify intent before YAML and STL generation

I can help turn a rough substation request into an engineering-sensible AIS prompt. I check bay logic, transformer connections, phase spacing, directions, and obvious ambiguities before geometry generation.

Agentic YAMLNatural language → validated AIS YAML → STL preview
TUNet-AI Thermal Predictionpreconfigured realistic AIS benchmark only

Displays the approved cache-backed 128^3 cropped AIS prediction. Arbitrary YAML/STL substations remain geometry-only until geometry-aware TUNet training is validated.

Live Sand Storage Generation

Describe a sand thermal store and generate validated YAML plus STL parts for the sand core, steel shell, insulation, and heater pipes. The same geometry can be used for LBM/TUNet training cases.

Sand Design AdvisorClarify storage intent before YAML and STL generation

I can help turn a rough sand-storage request into an engineering-sensible prompt. I check tank dimensions, heater layout, pipe spacing, wall clearance, power density, insulation, and boundary conditions before geometry generation.

Sand YAMLNatural language -> validated sand-storage YAML -> STL preview
TUNet benchmark: T42
The thermal prediction panel is intentionally locked to the approved benchmark scenarios `T01`, `T41`, and `T42`. If you edit the prompt manually, the site will still generate YAML and STL geometry, but it will not claim an accurate TUNet result for that arbitrary design.
TUNet-AI Thermal Predictionpreconfigured benchmark only

Displays the cache-backed `96^3` sand TUNet prediction at step 5000 for the selected approved preset.

Sand Storage RL Dispatch Evidence

Why it matters: dispatch policies help decide when stored thermal energy should be charged, held, or released under changing PV and demand conditions.

Cached replay of the IMPALA/V-trace sand-storage dispatcher using saved checkpoints, held-out winter/summer episodes, and PV-only baseline comparison. The relevant metric is policy performance versus baseline, not prediction error.

Loading cached IMPALA/V-trace evidence...

TUNet-AI Architecture

Proprietary KYAMOS-developed time-aware UNET architecture with sub-1% shock-capturing accuracy, non-autoregressive inference, and parametrized generalization across flow conditions.

Compressible Flow

Validated on Sod shock tube, Lax problem, and 3D engineering proxies. Handles discontinuities that defeat standard neural operators.