Arkaic

Computing Built for Discovery.

The future of high-performance computing software is here. Arkaic is building a next-generation, open platform for discovery science and engineering.

01Evolve

Don’t just optimize, Evolve.

Evolve drives long-running simulations. It compresses design cycles by an order of magnitude and removes computational barriers to discovery. Run parameter exploration and sensitivity analysis at scale. Perform model verification and validation. Build multi-fidelity surrogates and find optimal designs using Bayesian optimization. Create loops for cyclical workflows. Use agents to help craft, scale, and mend your workflows. Evolve turns computing into discovery.

02Design Databases

Accurate. Adaptable. Portable.

A design database captures how a system performs across its whole design space, and engineers use it to explore trade-offs before committing to a design. With Evolve, building one is nearly effortless. Its data structures and modules were designed from the outset with interoperability in mind. Build a pipeline once, then add runs, add a fidelity, or swap a solver. Frictionless and auditable.

  • Surrogates provide fast inference throughout the design space.
  • Uncertainty-aware sampling improves a surrogate with one command.
  • Reports on each surrogate’s quality are generated automatically.
  • Trained surrogates are portable: download one as a .surrogate.h5 file.
RAE 2822 airfoil simulated using SU2, over Mach number from 0.65 to 0.78, across the drag rise, and angle of attack from 0 to 3.5 degrees. Evolve ran 100 RANS simulations, 97 of which converged, and fit a Gaussian process surrogate of drag to them automatically: a prediction anywhere in the space, with its uncertainty.

03Efficient Optimization

High-fidelity results. Low-fidelity bill.

Optimization using high-fidelity computational fluid dynamics (CFD) or finite-element analysis (FEA) often takes hundreds of runs and millions of core-hours. Evolve uses multi-fidelity surrogates that blend coarse-grid or reduced-order results, reducing the time-to-solution and computational cost.

  • Multi-fidelity surrogates learn how your solver outputs relate.
  • Handles constraints and multiple objectives.
  • Batched evaluations parallelize the workflow.
  • Bayesian optimization that scales to dozens of parameters.
A destroyer hull from the L2-Sea benchmark, reshaped by Evolve over 14 shape variables on a coarse and a fine flow simulation. The design it found has 11% less drag than the original: 36.8 N against 41.5 N.

04Cycles

New data arrives. The workflow runs again.

Some workflows are never finished: a forecast runs again with every new analysis, a digital twin recalibrates as sensors report. A cycle repeats a graph of tasks at each step in time, each from fresh initial conditions. Describe the tasks once, and Evolve runs the sequence.

  • Each cycle runs the same tasks on the newest data.
  • A task starts when its conditions are met: after another task, or after it fails.
  • Failed tasks retry after a delay, or a backup task takes over.
  • Every cycle keeps its own files and logs.
Hurricane Matthew, forecast with WRF on a 27 km grid. Five cycles, six hours apart, from 00Z on October 6 to 00Z on October 7. Each cycle starts from its own NCEP FNL analysis (archived data, played back, which also supplies the boundary conditions) and runs a 24-hour forecast in about five minutes. The track and the central pressure update with every cycle.

05Community

Don’t work in a vacuum. Publish the whole study.

With Evolve, your published article gets a digital companion. Publish the data, the pipeline, and the trained models all in one place. Evolve's community page showcases your work and allows community members to fork the study and run it again: check your results, extend them, or start from them. It's also great for small teams working together in private on big problems.

  • Data, pipeline and trained models publish as one study.
  • Anyone can fork a published study and run it again.
  • A fork reads the published data in place: no download.
  • Every fork and reproduction is counted on the study.
A study from our CALEBO paper, published to the Evolve community with its runs, pipeline and trained models. Another member forks it, and the fork rebuilds the results table from the published runs, matching the paper. The study page counts the fork.

06Portable Workflows

Your work doesn’t stay in the cloud.

Run campaigns across many machines, then take the results to your workstation — or to a collaborator.

  1. 1

    Train in the cloud

    Campaigns scale across machines that start and stop as needed.

  2. 2

    Download your results

    The trained model as .surrogate.h5. Field datasets as CGNS.

  3. 3

    Refine locally

    Query the model from Python and run more trials on your own hardware.

Evolve's model page for a transport wing model, its Download button clicked, then a terminal running evolve inspect on the downloaded wing.surrogate.h5 file and listing its inputs, outputs, surrogates and provenance

Evolve with us.

We’re looking for researchers to try Evolve.

We use your email only to tell you when Evolve is open. See our privacy page.

07Team

  • Brandon Gusto

    Brandon Gusto, PhD

    Founder & CEO

    Brandon is a former Navy civilian scientist at NUWC Newport and DoW SMART Scholar. He led projects in multi-fidelity Bayesian optimization and GPU-accelerated adaptive beamforming for sonar. He trained at FSU as a mechanical engineer, applied mathematician, and computational scientist, and his doctoral work advanced wavelet-based adaptive mesh refinement approaches in the context of compressible and reacting flows, as well as ML sub-grid scale modeling.

  • Toan Tran

    Toan Tran

    Founding Engineer

    Toan is a research software engineer building scientific tools from statistical packages to full-stack data systems. At FSU, he analyzed large-scale in-vivo neural recordings and built a computer-vision labeling pipeline to extract neural clusters of motor activity in mice. At the University of Michigan, he collaborated with Stan developers on a Bayesian data analysis package, streamlining MRP workflows with visual analytics and a domain-specific model compiler.

Curious?

inquiries@arkaic.inc