⚡ What is Kontinua?

Kontinua is an open-core physics AI platform built for engineers, researchers, and scientific ML practitioners. It replaces expensive, multi-day HPC simulation runs (CFD, FEA, MHD) with millisecond-scale neural operator and transformer surrogates.

🌟 Core Capabilities

• Open-Source Core (pip install kontinua): 8 production-ready architectures (FNO, CNextU-Net, TFNO, AViT, AFNO, DilatedResNet, ReFNO), standardized spatial/spectral/temporal evaluation metrics, physics-aware normalization, and local ONNX/TensorRT export.
• 15TB Multi-Physics Corpus: Native streaming and distributed loaders for 16 physical domains from The Well (NeurIPS 2024), covering fluid dynamics, astrophysics, acoustic scattering, viscoelasticity, and biological active matter.
• Kontinua Cloud: Auto-scaling NVIDIA Triton GPU inference endpoints, sub-50ms latency REST/gRPC APIs, managed domain fine-tuning pipelines, and private benchmark CI/CD.

🚀 Quickstart

Installation

pip install kontinua

For benchmark evaluation and hardware acceleration:

pip install kontinua[benchmark,gpu]

1. Run Local Neural Surrogate Inference

import torch
from kontinua.core.models import CNextUNet
from kontinua.core.data import WellDataset

# Load pre-trained surrogate checkpoint
model = CNextUNet.from_pretrained("kontinua/cnext-unet-turbulence")
model.eval()

# Load initial condition / boundary states
dataset = WellDataset(well_base_path="./data", well_dataset_name="turbulence", well_split_name="test")
sample = dataset[0]["input"].unsqueeze(0)

# Generate 100-step spatiotemporal prediction in milliseconds
with torch.no_grad():
    prediction = model(sample)

print(f"Prediction shape: {prediction.shape}")

2. Connect to Kontinua Cloud API

import kontinua

# Authenticate with your Kontinua Cloud API key
client = kontinua.Client(api_key="ko_live_...")

# Execute cloud inference on managed Triton GPU cluster
result = client.predict(
    domain="turbulence",
    input_path="./boundary_condition.hdf5",
    model="cnext-unet-v2",
    rollout_steps=100
)

print(f"VRMSE: {result.vrmse:.4f} | Latency: {result.latency_ms}ms")
result.export("./output_rollout.hdf5")

🏗️ Model Zoo & Benchmark Baselines

🌐 Open-Core vs. Kontinua Cloud

🔬 Dataset & Scientific Citation

Kontinua is built on the 15TB The Well benchmark published at NeurIPS 2024 by Polymathic AI in collaboration with the Flatiron Institute, Cambridge, NYU, Princeton, UC Berkeley, Los Alamos, and Cornell.

@article{ohana2024well,
  title={The Well: A Large-Scale Collection of Diverse Physics Simulations for Machine Learning},
  author={Ohana, Ruben and McCabe, Michael and Meyer, Lucas and Morel, Rudy and Agocs, Fruzsina and Beneitez, Miguel and Berger, Marsha and Burkhart, Blakesly and Dalziel, Stuart and Fielding, Drummond and others},
  journal={Advances in Neural Information Processing Systems (NeurIPS)},
  volume={37},
  pages={44989--45037},
  year={2024}
}

📄 License & Community

• Open-Source Core: Licensed under the Apache-2.0 License.
• Documentation: docs.kontinua.ai
• Cloud Platform: kontinua.ai
• Community: Join our Discord or GitHub Discussions.