Particle detectors built from scintillator, a material that emits light when a charged particle passes through it, do not produce light in proportion to deposited energy. Dense ionization quenches light emission, Cherenkov light adds a component with its own energy threshold, and the readout electronics lose charge when photon pulses pile up. Experiments such as Daya Bay and JUNO-TAO describe this curve with a small physical model. They fit that model to about ten radioactive calibration sources plus continuous beta-decay spectra by minimizing one summed chi-square. When the model is right, every source can be fit at once. When the model is slightly wrong, the sources pull the parameters in different directions. The summed fit hides that disagreement by averaging over it, and experiments then cover it with hand-assigned, often conservative systematic errors.
This note proposes treating each calibration source as a separate objective and computing the Pareto front of the non-linearity model: the set of parameter settings where no source can be fit better without fitting another worse. The spread of this front is a direct, data-driven measure of how much the sources disagree. The family of non-linearity curves along the front becomes the systematic band on the energy scale. The front is traced with a power-mean aggregation of per-source misfits, a single dial that runs from the usual summed chi-square to a minimax fit in which every source must be matched. A subset-level analysis borrowed from attribute-interaction models then identifies which groups of sources jointly conflict, and so points to the part of the physics model that is misspecified.
Core Claim
The claim is statistical, not a better optimizer. If the non-linearity model is correctly specified, the Pareto front across calibration sources collapses to a region no larger than statistical noise. Any excess width therefore measures model misspecification, and the non-linearity curves along that front give an uncertainty band on the energy scale that covers the true curve under misspecification. The claim is that this band keeps near-nominal coverage where the summed-chi-square band undercovers, and is narrower than ad hoc envelopes that add alternative model shapes by hand. It is falsifiable in simulation, where the true curve is known.
Where This Sits
The calibration practice being questioned comes from A high precision calibration of the nonlinear energy response at Daya Bay. That work separates Birks-law quenching plus Cherenkov light from electronics non-linearity using a dedicated waveform digitizer. It fits gamma sources and the ¹²B beta spectrum jointly in one chi-square and reports positron energy uncertainty below 0.5% above 1 MeV.
Calibration strategy of the JUNO-TAO experiment follows the same single-objective approach. It reports non-linearity control within 0.6% from sources, improving to 0.4% with ¹²B. Both papers motivate this proposal: they show the payoff of combining heterogeneous sources, and JUNO-TAO shows that source-enclosure and shadowing effects are handled as separately assigned biases rather than inferred from source disagreement.
The closest work linking detector inversion and multi-objective optimization is An application of genetic multi-objective optimization algorithm to neutron spectrum unfolding problem. It trades response error against spectral entropy with a genetic algorithm. That trade-off is between data fit and a regularizer, and the front is used to pick one answer. Here the objectives are different data sets, and the front is itself the uncertainty estimate. That paper also does not model non-linearity.
Unfolding methods address a downstream problem and are left aside:
- iterative unfolding on EJ-301 (High-fidelity unfolding of 252Cf neutron spectra);
- hierarchical Bayesian unfolding (A Hierarchical Bayesian Approach to Neutron Spectrum Unfolding);
- neural forward unfolding (Simulation-Prior Independent Neural Unfolding Procedure).
All three take the detector response as given. This proposal targets how much to trust that response.
Two papers supply the optimization method.
- Optimal Scalarizations for Sublinear Hypervolume Regret shows that weighted sums cannot reach concave parts of a front. It also shows that random-direction Chebyshev-style scalarizations explore such fronts at the optimal rate. That rate degrades with the number of objectives, which constrains the design to about ten sources, not hundreds. - Power mean pyramid scores supplies the power mean as a one-parameter continuum between averaging and taking the maximum. The parameterized t-norms of Reducing t-norms and augmenting t-conorms offer an alternative continuum, but they are defined on [0,1] scores and would need an extra normalization step. The power mean is used instead.
Multiobjective Optimization using GAI Models constrains the scope. It shows that exact Pareto computation over interacting components blows up with interaction width, and that ε-dominance compression works well: more than 10⁶ solutions reduced to 109 at ε = 0.1. The subset analysis below is kept to a few sources for the same reason, and ε-compression is reused to report the front.
Research Program
Let $\theta$ be the non-linearity model parameters: Birks coefficient, Cherenkov fraction, electronics terms, and absolute scale. For calibration source $s$ with $n_s$ degrees of freedom, let $q_s(\theta)=\chi^2_s(\theta)/n_s$ be its normalized misfit. For weights $w$ on the simplex and exponent $\alpha$, the aggregate misfit is
$$L_{\alpha,w}(\theta)=\Big(\sum_s w_s\, q_s(\theta)^{\alpha}\Big)^{1/\alpha}.$$
Setting $\alpha=1$ with $w_s \propto n_s$ recovers the standard summed chi-square, up to scale. As $\alpha\to\infty$, the aggregate becomes the weighted minimax fit, which reaches concave parts of the front.
Minimizing $L_{\alpha,w}$ over random $w$ and a grid of $\alpha$ traces the front, following the random-weight recipe of Zhang. Each front point $\theta$ maps to a curve $f_\theta(E)$, the visible-to-true energy ratio. The proposed band at energy $E$ is the envelope of $f_\theta(E)$ over front points whose worst per-source misfit stays within a statistical tolerance. That tolerance is set from pseudo-experiments generated under a correctly specified model.
The interaction component comes from decision theory, where capacities score subsets of criteria rather than single ones. For any subset of sources $A$, define
$$v(A)=\min_\theta \max_{s\in A} q_s(\theta),$$
the best achievable worst-case misfit when only the sources in $A$ must agree. This set function grows as sources are added. Its excess over pairwise values is a synergy of tension in the Choquet sense of Bi-capacities - II: the Choquet integral: some triples of sources conflict even though every pair within them can be reconciled. With about ten sources, all 1,024 subset fits are cheap. The resulting map of tension shows which physical component the model gets wrong. For example, low-energy gammas conflicting with the ¹²B spectrum would implicate the electronics term. This role is diagnostic, not the main result.
The testbed is a simulated liquid-scintillator detector with the Daya Bay source list and a ¹²B spectrum at realistic statistics. The truth models include quenching, Cherenkov, and electronics shapes. Misspecification is injected in four ways:
- an electronics shape outside the fit family;
- unmodeled energy loss in source enclosures;
- a quenching law with a second-order term;
- a correctly specified control.
Baselines
Naive reference. The summed chi-square fit with only statistical errors, whose one-sigma curve band is the error estimate. It shows how badly undercoverage appears under misspecification.
Strongest competitor. Current practice: summed chi-square with nuisance parameters, plus an envelope over alternative model shapes, as in Daya Bay and JUNO-TAO. The reported numbers to match are below 0.5% positron energy uncertainty above 1 MeV (Daya Bay) and 0.4% non-linearity control with ¹²B (JUNO-TAO). Those are real-detector numbers, so on the simulated detector this baseline is re-implemented with the same parameters, data, and a hand-built set of alternative shapes of comparable effort. The Pareto band beats it if it reaches nominal coverage of the true positron non-linearity curve in all misspecification scenarios with equal or smaller mean width. It also beats it if it matches that coverage without the analyst having to guess the alternative shapes.
Scale-factor inflation. Rescale all errors by the square root of the global chi-square per degree of freedom when that value exceeds one. This is the cheapest way to absorb tension.
Leave-one-source-out envelope. Refit with each source dropped and take the spread of the resulting curves.
Ablations isolating the key ingredient.
- Linear-weight sweep: $\alpha=1$ only. This isolates whether non-linear aggregation matters.
- Pure random-direction hypervolume scalarization from Zhang: tests whether the power-mean dial adds anything over it.
All methods receive the same pseudo-data, the same model family, and the same budget of minimizer calls, set to 2,000 per pseudo-experiment. No numeric result for Pareto-based calibration bands was found on Lacuna.
Why not just…?
Why not just use the summed chi-square with nuisance parameters, which already reaches sub-percent precision? Sub-percent precision is not the same as coverage. A summed fit reports small errors exactly when the model is wrong in a way the sources average over. JUNO-TAO's separately assigned enclosure and shadowing biases show that practice already recognizes the gap and fills it by hand. The correctly specified control scenario is the test. If the summed-fit band covers the truth there, and also covers it under the injected misspecifications, the proposal adds nothing, and that result will be reported.
Why not just inflate errors by a global scale factor? This is the real threat for mild misspecification. The scale factor spreads tension uniformly over energy. Tension from electronics non-linearity should instead concentrate at low energy. The per-energy band widths of the two methods will be compared against the true error curve. If scale-factor inflation achieves comparable coverage and width everywhere, the recommendation becomes to use it.
Why not just run leave-one-source-out refits or a standard dataset-consistency test? Dropping one source probes only the corners of the front. A consistency test returns one p-value with no curve band. When tension involves three or more sources jointly, which the subset analysis detects, leave-one-out can miss it. The leave-one-out envelope is run as a baseline. If it matches the front band's coverage in all scenarios, the front computation is unnecessary and the contribution shrinks to the subset diagnostic.
Why not just give the model more freedom, such as a spline non-linearity, so the tension disappears? A flexible curve removes the disagreement by absorbing it. Its error band then widens, most at low energy, where the EJ-301 study notes the response is hardest to constrain. A spline fit is run as a further baseline, and bands are compared at matched coverage.
Decisive Checks
Coverage. Over 1,000 pseudo-experiments per scenario, measure the fraction whose true positron non-linearity curve lies inside each band at every energy from 1 to 8 MeV, and separately below 1 MeV.
Mean band width. Measured in percent of energy scale.
Downstream effect. Propagate each band into the bias of a fitted reconstructed-positron spectrum shape.
Is the front concave? Compare fronts from $\alpha=1$ and $\alpha\gg1$ using hypervolume. If the linear sweep recovers the same front, the non-linear dial is unneeded and the method reduces to weighted-sum sweeps.
Does front width scale with the injected misspecification? In the control, front width should match the width from statistical noise alone. If it does not, the tolerance calibration is wrong and the band is not trustworthy.
Does the subset analysis name the right culprit? Check whether the most tense source subsets correspond to the injected fault in each scenario. A random-culprit null sets the bar.
A positive outcome is nominal coverage across scenarios with width no larger than the hand-built envelope, together with correct culprit identification in at least three of the four fault types.
References
- A high precision calibration of the nonlinear energy response at Daya Bay
- Calibration strategy of the JUNO-TAO experiment
- An application of genetic multi-objective optimization algorithm to neutron spectrum unfolding problem
- High-fidelity unfolding of 252Cf neutron spectra in EJ-301 scintillators via calibrated response matrices and Monte Carlo simulations
- A Hierarchical Bayesian Approach to Neutron Spectrum Unfolding With Organic Scintillators
- Simulation-Prior Independent Neural Unfolding Procedure
- Optimal Scalarizations for Sublinear Hypervolume Regret
- Power mean pyramid scores for summarization evaluation
- Reducing t-norms and augmenting t-conorms
- Multiobjective Optimization using GAI Models
- Bi-capacities - II: the Choquet integral