Interactive Scientific-ML Playgrounds
Physics-informed and Neural-ODE demos that train live in your browser — no install. Inverse kinetics with maximum-likelihood uncertainty (KINNs / rKINNs), Neural-ODE structure discovery from noisy data, and pharmacokinetics with real theophylline measurements.
Kinetics-Informed Neural Networks
Recover the rate constants of a known reaction from noisy, sparse transients — a physics-informed inverse problem with maximum-likelihood uncertainty, no tuned weights.
Neural ODE: discover the Lotka–Volterra rate law 🐇 🐺
Learn the unknown Lotka–Volterra predator–prey vector field through the integrator, then recover its structure: black-box with physics priors, a hybrid power-law network, and sparse discovery with learnable exponents (SINDy inside the Neural ODE).
Pharmacokinetics: a non-negative mean-variance Neural ODE 💊 🩸
Learn the unknown absorption-then-elimination rate law of a drug from sparse, noisy blood samples through the integrator (discrete adjoint), kept non-negative by construction. As a mean-variance Neural ODE it predicts its own plus/minus sigma error bar. Real theophylline data, runs in the browser.
Which playground should I use?
Start with KINNs to estimate parameters in a known reaction model, Lotka-Volterra to explore learning a dynamical law, or pharmacokinetics to fit a structured model to real measurements.
| Playground | What it learns | Structure and data |
|---|---|---|
| KINNs: inverse kinetics | Concentration trajectories and, in inverse mode, rate constants. The neural network represents the trajectories directly. | A chosen chemical reaction mechanism constrains the fit. Explore synthetic observations, noise, physics residuals, and uncertainty. |
| Lotka-Volterra: Neural ODE discovery | A rate field whose numerical integration predicts predator-prey trajectories. Compare neural, hybrid, and sparse discovery models. | A synthetic system supplies a known reference law. Vary the data split, noise, and structural assumptions to inspect generalization. |
| Pharmacokinetics: structured Neural ODE | Drug absorption and elimination rates, with a learned noise model in the mean-variance formulation. | Real theophylline measurements, a compartment structure, and non-negative rates. Compare fits across subjects and against an analytical baseline. |
The demos connect to the KINNs paper and the maximum-likelihood formulation. Each playground explains its equations, assumptions, controls, and references.
Try a worked example on the existing pages: recover four reaction rates, check a learned predator-prey law, or hold out the late concentration tail.