A generator built around gradients
The central result is a software demonstration. ADoNIS is presented as a fully differentiable neutrino-interaction event generator that reproduces the prediction shapes of ACHILLES in the reported validation comparisons and exposes parameter gradients. Those gradients run through the nuclear-ground-state, hard-scattering-amplitude and stochastic intranuclear-cascade components. They can be evaluated for individual events and after events are combined into predictions.
The construction generates a fixed stochastic event bank, a stored collection of simulated events, and stores auxiliary quantities for each event. It then recomputes each event's weight as a function of the physics parameters. The demonstration considers 28 differentiable parameters: 11 electroweak-vertex parameters, 11 cascade parameters, four spectral-function parameters and two channel-normalisation parameters.
The comparison has a clear target
Validation ranged across different probes, interaction channels, target nuclei and beam energies. At MINERvA energies, quasi-elastic and resonant contributions were both substantial, and agreement with ACHILLES was reported for both.
That comparison has a defined boundary. It tests agreement with ACHILLES rather than independent validation of the underlying physics across every relevant data set, and the work does not present ADoNIS as more physically accurate than ACHILLES.
Gradients map which probes matter
The study used Jacobian and Fisher analyses to examine how different probes respond to different parameters. In plain terms, the analysis tracked how predicted observables changed as the parameters changed. It found complementary sensitivity: hadron beams constrained only cascade effects, neutrino and electron samples together constrained nuclear parameters, and neutrino samples were most informative for form-factor parameters. This was a model-based sensitivity demonstration, not an independent experimental parameter estimate.
Inference tests stayed inside simulation
Parameter inference was tested with a closure fit, in which the data are generated from the same prediction used in the fit. The exercise used 21 observables and 17 parameters. The injected truth was defined by randomly displacing parameters within their physical ranges. Each bin was assigned a 5% uncertainty, no informative priors were imposed, and the fit used a Gaussian chi-squared with exact-Jacobian Gauss-Newton optimisation.
The fit was repeated for 2,000 statistically fluctuated toys using the same injected prediction and per-bin uncertainties. Coverage, the proportion of intervals containing the value used to generate the data, was 67.8% ± 1.0%, 89.4% ± 0.7% and 94.9% ± 0.5%, compared with nominal levels of 68.3%, 90% and 95%, respectively. The median chi-squared per degree of freedom was 0.992, and the reported coverage values were close to those nominal levels.
Timing and sampling comparisons
The timing comparison was reported for a representative fit. With 1.25 × 10^5 events per sample and 17 fit parameters, the median time was 1.6 seconds with Gauss-Newton, 4.8 seconds with gradient-supplied MIGRAD and 12.3 seconds with MIGRAD without derivatives. The timings were medians under the stated hardware and benchmark conditions and excluded JAX compilation.
After convergence checks, the compared gradient-based NUTS method produced more effective samples per unit of wall-clock time than MH. NUTS yielded 2.0 times the minimum effective sample size and 1.4 times the median effective sample size of MH. The ratio was 3.4 for tail effective sample size at the 5% and 95% quantiles bounding a 90% credible interval. No uncertainty interval for these ratios was reported.
Unfolding recovered simulated spectra
The final demonstration used template-fit unfolding, a calculation intended to recover an underlying spectrum after detector effects. It mirrored a T2K muon-neutrino CC0π setup, with 10% momentum smearing, 5 degrees of angular smearing and 50% efficiency for charged pions below 400 MeV/c. Signal purity was 0.915, the selected sample contained approximately 170,000 events and the event bank had about 24 times larger Monte Carlo statistics.
The unfolding fit contained 196 parameters: 58 unconstrained signal templates, 10 flux parameters, 12 retained interaction-model parameters and 116 independently constrained detector parameters.
The fit recovered the injected spectrum in a closure dataset and in a second dataset whose true signal was uniformly increased by 20%. Both were closure-style simulations using a simple stochastic detector model. The result concerns recovery under those assumptions, not a measurement from real detector data.
The boundaries are part of the result
ADoNIS inherits the physics choices used by ACHILLES, and the work does not present those choices as preferred. Meson-exchange currents and deep-inelastic scattering are absent from both models. The demonstrations do not establish improved oscillation-parameter precision or performance on real detector data. They also do not establish native differentiability for omitted hadronisation or interaction models, or show that every model parameter is constrained in every analysis configuration.
The paper states that ADoNIS is released as an open-source implementation for practical use and as a reference construction. The document is identified as arXiv version 1 in hep-ex, dated 25 August 2026. The work was supported by compute credits from Anthropic's AI for Science program.
Paper data and sources
Original title: ADoNIS: A Differentiable generatOr of Neutrino Interaction Samples
Authors: César Jesús-Valls
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text