SPARK
Trace Recognition and Amortized Conditional Estimation — Bayesian inference for reaction mechanisms and kinetic parameters from cyclic voltammetry (CV) or temperature-programmed desorption (TPD) data, in one forward pass.
One CSV per scan rate. Expected columns: Potential (V), Current (A/mA/µA), optional Time (s) — if present, the scan rate is inferred as the median of |dE/dt| over the forward sweep (no need to specify it manually).
Results
One image per scan rate (potential on x-axis, current on y-axis). Axis bounds auto-detected via OCR — override in Advanced if needed. Note: digitized curves are inherently noisier than the simulated data the model was trained on, so the OOD detector may flag image inputs and reconstruction quality is typically lower than the CSV path; for production use, prefer CSV.
Select the unit shown on the y-axis of your plot
Results
One CSV per heating rate. Expected columns: Temperature (K), Signal, optional Time (s) — if present, the heating rate β is inferred as dT/dt from the time stamps (no need to specify it manually).
Results
One image per heating rate (temperature on x-axis, signal on y-axis). Axis bounds auto-detected via OCR — override in Advanced if needed. Note: digitized curves are inherently noisier than the simulated data the model was trained on, so the OOD detector may flag image inputs and reconstruction quality is typically lower than the CSV path; for production use, prefer CSV.
Results
How it works
SPARK (Simulation-based Posterior Amortization for Reaction Kinetics) uses conditional normalizing flows with a Set Transformer encoder to perform amortized Bayesian inference. Given one or more experimental curves, it simultaneously classifies the reaction mechanism and produces full posterior distributions over kinetic parameters — in a single forward pass.
The deployed checkpoints are the noise-augmented headline models
(CV: v14_9mech, TPD: tpd_11mech_v2) that retain 92.4 % (CV) and
95.6 % (TPD) classification accuracy under realistic measurement noise.
Inference runs in ~50 ms on CPU; mean 90 %-credible-interval coverage
is 92.2 % (CV) and 92.9 % (TPD).
Training data is generated from physics-based simulators (Crank–Nicolson for CV, ODE integrators for TPD). Posteriors are calibrated via a coverage-aware loss with per-parameter inverse-spread weighting.
Citation
Yan, B. (2026). SPARK: Amortized Bayesian Inference for
Mechanism Identification and Parameter Estimation in
Electrochemistry and Catalysis via Conditional
Normalizing Flows. [Preprint]
Supported mechanisms
Electrochemistry (CV, 9): Nernst, Butler–Volmer, Marcus–Hush–Chidsey, Adsorption, EC, Langmuir–Hinshelwood, EE, EC′, CE.
Catalysis (TPD, 11): First-order, Second-order, Zeroth-order, FirstOrderCovDep, DiffLimited, Precursor, Dissociative, ActivatedAdsorption, LH Surface, Mars–van Krevelen, TwoSite.