KINOS is an interpretable computational engine for kinetic geochemistry. Built on physics-informed neural networks (PINN) and neural operators (PINO, DeepONet, FNO), it solves the equations of reactive transport, isotope fractionation and molecular degradation under explicit conservation laws. The engine runs entirely on the operator's hardware, with no cloud dependency, no third-party inference API, and no black-box reasoning.
Where a standard neural network learns an input-output mapping from data, a physics-informed neural operator learns the solution of a family of partial differential equations. KINOS extends this framework to chemical and isotopic kinetics, explicitly embedding reaction, fractionation and degradation into the loss.
PINN (Physics-Informed Neural Networks) solve a given partial differential equation with fixed parameters. Neural operators (DeepONet, FNO) learn the solution operator itself, enabling generalisation to new parameter sets without retraining.
KINOS combines both. The K.01 core is an ADE PINO that learns the solution family C(x, t ; K, α, R) of the advection-dispersion-reaction equation on heterogeneous fields. Once trained, it evaluates a contaminant plume in milliseconds, where an explicit solver requires seconds to minutes depending on grid resolution.
Kinetics are treated as a first-order constraint in the loss. Fractionation coefficients ε, degradation rate constants k, retardation factors R become learnable variables conditioned by geochemical context (pH, Eh, fOC, lithology).
Off-the-shelf reactive transport solvers and general-purpose machine learning frameworks treat kinetics as an afterthought. Reaction rate laws are bolted on, isotope fractionation is post-processed, and conservation constraints are enforced loosely if at all. The result is software that simulates geochemistry but does not reason with it.
KINOS was built to close this gap. The kinetic structure is not an addition but the core architectural choice: the neural operators are constructed so that mass conservation, reaction stoichiometry, isotope fractionation and thermodynamic constraints are part of the model's hypothesis space, not learned approximations of them.
This matters because the use cases that require kinetic geochemistry, forensic provenance, contaminant attribution, defence-grade traceability, demand interpretable and falsifiable inference, not statistical correlation. A black-box model that happens to fit the data is not admissible evidence. KINOS is designed from the outset to produce reasoning that holds up under scientific and legal scrutiny.
KINOS emerged from years of operational use of IsoFind, a desktop platform for isotopic traceability deployed in research and analytical contexts. IsoFind worked, and continues to work, without any machine learning component.
But operational use revealed structural limits of classical approaches: solver runtimes incompatible with iterative scenario exploration, parameter calibration that did not generalise, and the inability of standard tooling to fuse heterogeneous analytical evidence into a single interpretable model.
KINOS is the answer to those limits. Not a replacement for classical methods, but a complement, an inference engine where the existing toolchain reaches the edge of what it can deliver.
Each operator solves a family of partial or ordinary differential equations specific to one type of geochemical problem. All share the same base architecture (parameter encoder, neural operator, solution decoder) and the same composite loss: data + physics + boundary conditions.
PINO solving the advection-dispersion-reaction equation on heterogeneous fields. Dirichlet and Neumann conditions supported, parametric source term, learned longitudinal and transverse dispersivity.
Mechanism identification through δ¹³C / δ³⁷Cl, δ²H / δ¹⁸O or arbitrary isotope pair cross-plots. Spatialised ε system, accounting for remaining fraction and non-linear Rayleigh curves.
Network of degradation pathways for PFAS, chlorinated pesticides and halogenated solvents. Rate constants k conditioned by pH, Eh and matrix. Native coupling with routine analytical data (GC-MS, LC-MS/MS).
Random Forest trained on the FOREGS European continental geochemistry dataset. ONNX output. Per-lithology derivation of hydraulic parameters K, longitudinal dispersivity α_L and organic carbon fraction f_oc.
Origin reconstruction via Bayesian inversion of the direct system K.01–K.04. Process identification, multi-scenario comparison, uncertainty propagation. Output directly usable as evidentiary material.
Cryptographically signed exchange of isotope data. Three-tier PKI (Root CA · Issuing CA · Lab), optional asymmetric encryption of the scientific payload (X25519 + AES-256-GCM), ECDSA P-256 signature.
KINOS is a standalone native application, designed for isolated workstations. The PINN/PINO core, the orchestrator, the simulation engine and the scientific database are packaged into a single binary, with no outbound calls during execution.
Model weights are versioned, exportable in ONNX and cryptographically signed. Data, parameters and results remain entirely on the operator's infrastructure. No telemetry is emitted by the runtime, no licence validation occurs online.
A scientific inference engine is defined as much by what it refuses as by what it provides. The following are not implementation limitations awaiting future fixes, they are deliberate architectural commitments. They define the boundary between KINOS and general-purpose machine learning platforms.
KINOS does not contact any external service to compute, infer, or verify. The full pipeline runs on the operator's machine. No API key, no remote model call, no licence ping. An isolated network is a supported deployment mode, not a degraded one.
KINOS does not produce results that cannot be inspected, replayed, or contested. Every output carries its model identifier, parameter set, residual physics check, and uncertainty bound. A claim that cannot be defended scientifically is not a claim KINOS will make.
KINOS does not produce results that depend on hidden state, undisclosed randomness, or moving model versions. Every inference is bound to a versioned model weight, a deterministic seed, and a signed input dataset. The same inputs produce the same outputs, today and in five years.
KINOS is not a wrapper around someone else's model. It does not require an OpenAI, Anthropic, or Google API key to function, nor does it fall back on commercial generative AI for any computational task. The inference is done by physics-constrained neural operators trained for the task, and shipped with the suite.
KINOS does not phone home. It collects no usage statistics, no error reports, no model performance metrics, and emits no analytics. What you do with the suite remains entirely within your infrastructure. Anomalies are logged locally and never transmitted.
KINOS stores its data in open formats (SQLite, ISOF, ONNX) that can be read and reused without the suite itself. Trained model weights are exportable. Scientific outputs can be exchanged with any conforming tool. The operator owns the data, the models, and the workflows.
Sovereignty is not reduced to European hosting. It requires that code, weights, data and jurisdiction be simultaneously controlled by the operator. KINOS was designed along this principle from the first line of code.
The runtime contacts no external service. Database encrypted at rest (SQLCipher · AES-256), signed exports (ECDSA P-256), cryptographically chained audit trail (HMAC-SHA256).
All neural operators are exported in ONNX, accompanied by their composite loss function, their referenced training dataset and their signature certificate. Scientific reproducibility guaranteed.
KINOS is developed and operated by Colin Ferrari (sole trader) (Foix, Ariège, France). Intellectual property, hosting and infrastructure within the European Union. Outside the reach of the Cloud Act, FISA and extra-territorial jurisdictions.
KINOS is built as a long-term piece of European scientific infrastructure, not as a venture-funded race to product-market fit. The goal is not to ride the current wave of generative AI, but to provide laboratories, agencies and institutions with a stable, interpretable, locally owned computational engine that will still be running, auditable and useful a decade from now.
Stabilise the six operators, publish the underlying methods in peer-reviewed venues, build a community of qualified users across French and European research institutions. Make ISOF a de facto exchange format in isotope geochemistry.
Achieve the qualifications required for evidentiary use of KINOS outputs in forensic, regulatory and defence contexts. Engage with CEA-DAM, DGA and European equivalents on traceability of critical materials and environmental forensics.
Extend the physics-constrained inference framework to adjacent scientific domains where kinetics and conservation laws structure the problem: hydrogeology, atmospheric chemistry, materials degradation. KINOS as the reference engine for interpretable scientific inference.
KINOS is available in early access for research laboratories, environmental agencies, forensic institutions and defence bodies. Deployment includes installation of the suite, access to forthcoming modules, and support for the scientific qualification of the models.
The first partner institutions receive direct follow-up from the scientific team and contribute to defining the next families of operators.