Technology
Rigorous physics. Transparent validation.
Every answer traces back to an open-source reference solver, a documented validation study, and a logged uncertainty estimate. No black boxes.
Open-source reference solvers
We build on community-audited, peer-reviewed solvers — not proprietary engines. This means our models can be independently verified, and our validation studies are reproducible.
OpenFOAM
Finite-volume CFD for fluid dynamics, mass transport, and conjugate heat transfer. Used for device-scale flow modeling, drug elution under physiological conditions, and injection-site fluid mechanics.
FEniCSx
Finite-element framework for structural mechanics, diffusion-reaction systems, and coupled multi-physics. Used for implant stress analysis, polymer degradation kinetics, and PDE-constrained optimization.
Neural operators
Neural operators learn mappings between function spaces — not just input-output pairs. Trained on validated solver data, they generalize across geometry and parameter variations while respecting the physics they were trained on.
DeepONet
Deep operator networks approximate solution operators for parametric PDEs. We use DeepONet surrogates for release-kinetics prediction and sensitivity analysis across formulation design spaces.
FNO
Fourier neural operators learn in frequency space, enabling fast inference over high-dimensional spatial fields. Used for flow-field prediction in device geometries and thermal dose mapping.
Validity-envelope checks
Every query is evaluated against the validated operating envelope before a response is returned. Queries inside the envelope are answered by the neural-operator surrogate. Queries outside it trigger a full reference-solver run. Queries involving unsupported physics are flagged for expert review.
Inside envelope
Surrogate inference — seconds to minutes. Uncertainty bounds included.
Outside envelope
Reference-solver run triggered automatically. Results logged and used to extend the envelope.
Unsupported physics
Expert review initiated. Scoping estimate provided within 48 hours.
Credibility ledger
Every simulation run is logged with its context of use, verification evidence, validation evidence, and uncertainty quantification. The ledger is structured to support ASME V&V 40 credibility assessments and regulatory submissions.
The specific question the model was used to answer, including scope and intended decision.
Evidence that the model solves the equations correctly — mesh convergence, solver residuals, benchmark comparisons.
Evidence that the model represents the physical system — comparison to experimental or clinical data with quantified discrepancy.
Propagated input uncertainty and model-form uncertainty, reported as confidence intervals on all outputs.
Data privacy
Customer data — CAD files, formulation parameters, experimental results — is never used to train shared models without explicit opt-in. Each customer's digital twin is isolated, versioned, and accessible only to that customer.
Customer data is never used to train shared or public models without explicit written opt-in.
Each digital twin is isolated per customer — no cross-customer data leakage.
All simulation inputs and outputs are versioned and auditable.
Data handling practices are documented and available on request for regulatory due diligence.