Experiments searching for light dark matter must make many analysis choices before they look at their data. These choices include the lowest signal size they accept, how far from the electrodes an event must sit to count, and how they veto bursts of spurious electrons. Each choice helps some candidate dark-matter masses and hurts others. A tighter cut against detector noise sharpens the search at 5 GeV but removes the tiny signals expected at 1 GeV. Today these choices are tuned by hand into one analysis, judged by a single loss or a single mass point. We propose to treat the expected exclusion limit at each mass as its own objective. We then use sample-efficient multi-objective optimization (search methods that return the whole set of best compromises rather than one answer) to find the small set of analyses that together come close to the best achievable sensitivity at every mass.
This matters because the most sensitive light dark matter searches now work at the detector's noise floor. In Search for Light Dark Matter with 259 Days of Data in PandaX-4T, the reach comes from "ionization-only" data: events where only the delayed charge signal is seen, extending nuclear-recoil sensitivity to roughly 0.33–3.0 keV. In that regime, non-radiogenic backgrounds such as electrode micro-discharges compete with signal, and the sub-keV charge yield is poorly calibrated. Every candidate analysis is costly to score, because its expected limit comes from thousands of simulated pseudo-experiments. At the same time, every pseudo-experiment produces limits at all masses at once. That output is naturally a vector, and that is exactly the setting where multi-objective Bayesian optimization has been shown to beat repeated single-objective tuning.
Core Claim
The claim has three parts. First, the cut configurations of a low-threshold ionization-only search form a genuinely multi-objective problem whose objectives are expected limits across dark-matter mass. Second, the trade-off curve between masses is non-convex in parts. Third, a handful of analyses chosen from this curve, specified in advance before looking at the data, comes within a stated tolerance of the per-mass best limit at every mass, at a fraction of the simulation cost of tuning each mass separately.
Concretely, let $\theta$ be a vector of analysis settings, and let $m_1,\dots,m_K$ be a grid of dark-matter masses. The optimization targets the objective vector
$$\mathbf{f}(\theta) = \big[\log_{10}\tilde{\sigma}_{90}(m_1;\theta),\ \dots,\ \log_{10}\tilde{\sigma}_{90}(m_K;\theta),\ \Delta_{Q}(\theta)\big].$$
Here $\tilde{\sigma}_{90}(m_k;\theta)$ is the median 90% confidence-level upper limit on the spin-independent cross-section across background-only pseudo-experiments. $\Delta_{Q}(\theta)$ is the worst-case degradation of that limit when the sub-keV charge yield is moved across its calibration band. All entries are to be minimized.
To our knowledge, nothing in the literature we reviewed applies Pareto-front optimization to dark matter detection analyses. The closest boundary work treats the problem differently. Deep learning based pulse shape discrimination for germanium detectors tunes one classifier cut at high energy. Determination of responses of liquid xenon to low energy electron and nuclear recoils using a PandaX-II detector fits detector yields with one combined likelihood. Neither treats sensitivity across mass as a vector of objectives.
Research Program
The first component is a simulated replica detector. We build an open, simulated stand-in for a PandaX-4T-like ionization-only search, using the NEST2.0 response model. NEST2.0 is the standard simulation of light and charge production in liquid xenon. We parametrize its yields as in the PandaX-II work, where Legendre-polynomial light and charge yields were fit to tritium, radon and americium-beryllium calibration data. We set the charge-yield uncertainty band below about 4.6 keVnr using the reach of neutron calibrations reported in The design, implementation, and performance of the LZ calibration systems.
The background model has four parts:
- electron recoils;
- the solar ⁸B coherent neutrino-nucleus scattering floor;
- cathode events;
- an empirical micro-discharge and delayed-electron component.
For the last component we keep several plausible variants, so that we can test how much the optimized cuts depend on modeling choices we cannot check against real data. We do not claim to reproduce PandaX-4T's observed limit. We claim only that a replica of its published cut configuration provides a fair reference point within the simulation.
The second component is the search space and the search. The settings vector $\theta$ has roughly 10–15 dimensions. It mixes continuous cuts (charge threshold in electrons, drift-time and radial fiducial bounds, signal-width cut, veto window after large signals, upper edge of the energy window) with categorical choices (veto scheme, likelihood binning). Each evaluation runs on the order of a thousand pseudo-experiments and returns the full objective vector. Its cost is minutes to hours, so evaluation budgets are in the hundreds, not tens of thousands.
We use two search methods. The first is tree-ensemble surrogate optimization with mixed-integer acquisition from Multi-objective constrained optimization for energy applications via tree ensembles, because it handles categorical variables and hard constraints. One such constraint is a minimum signal acceptance at the lowest mass. The second is the trust-region method from Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces, because it handles larger continuous spaces cheaply.
With $K$ around 6–8 masses plus the robustness objective, we enter the many-objective regime. There, nearly all candidates become mutually non-dominated, a difficulty analyzed in A Crowding Distance That Provably Solves the Difficulties of the NSGA-II in Many-Objective Optimization. We exploit the fact that limits vary smoothly with mass: the surrogate is fit with mass as an input, so neighbouring objectives share statistical strength rather than being modeled independently.
The third component turns the front into a pre-registered analysis set. From the front we select the smallest set of $k$ analyses whose lower envelope stays within a tolerance, for example 10% in cross-section, of a per-mass best envelope at every mass. We then fix the trials penalty incurred by reporting $k$ analyses. This step is deliberately deterministic: the experiment commits to specific analyses, never to a random mixture. This echoes the warning in A practical guide to multi-objective reinforcement learning and planning that randomizing between policies is unacceptable in high-stakes settings.
The fourth component is a transfer check on germanium. We repeat a smaller version on a simulated point-contact germanium array configured after Projected WIMP sensitivity of the CDEX-50 dark matter experiment. There, the settings are the energy threshold near 160 eVee and the cut separating surface from bulk events. This tests whether the mass-wise structure of the front is a general feature of near-threshold searches.
Baselines
The naive reference point is the replica of the published PandaX-4T ionization-only configuration, evaluated in our simulation. The paper's observed result is a spin-independent limit of 1.1 × 10⁻⁴³ cm² at 3 GeV/c² at 90% confidence. That number anchors the scale of our replica, but it is a single observed realization on real data. We will compare only against the replica's median expected limit, computed with the same pseudo-experiments as every other candidate. Beating this baseline means a front point that lowers the median expected limit at some masses without raising it anywhere beyond Monte Carlo error. More importantly, it means a pre-registered set of two or three analyses that improves the envelope at the mass extremes, below 2 and above 6 GeV/c².
The strongest competitor is independent single-objective Bayesian optimization run separately at each mass. Pooling its per-mass optima also gives us the per-mass best envelope we use as a target. It gets the same total number of simulator calls, split across masses, and the same surrogate family and tuning budget. To be fair, each per-mass run may read the full vector from its own evaluations, but it optimizes only its own mass. The proposal beats it if, at equal budget, the multi-objective front's best $k$-analysis set lies closer to the envelope than per-mass optimization reaches at that budget. Separately, we will note whether the front recovers that envelope with roughly $K$-fold fewer calls.
The second competitor is a weighted-sum sweep over masses, with weights on a grid and the same budget. The comparison of vector-valued search against such sweeps in Multi-Objective Optimization of a Path-following MPC for Vehicle Guidance: A Bayesian Optimization Approach reported a front with hypervolume 0.8480 after 2,500 evaluations, against nearly 20,000 for the weighted sum. That gap is the effect size we test for in our domain; it is not assumed.
The ablation that isolates the key ingredient keeps the multi-objective optimizer but drops mass as a shared surrogate input, modeling each objective independently. A second ablation drops the robustness objective $\Delta_Q$. This checks whether fronts optimized without it choose cuts whose gains vanish under yield uncertainty. Metrics throughout are hypervolume in log-limit space, maximum gap to the per-mass envelope, the number of analyses $k$ needed to meet the tolerance, and total simulator CPU-hours.
Why not just…?
Why not just tune one analysis per mass and report them all? This is the real threat. With unlimited compute, it defines the per-mass best envelope. It falls short on cost, since $K$ separate searches each discard the free vector output. It also falls short on statistics: reporting $K$ separately tuned analyses inflates trials and multiplies the work of keeping each analysis blind until its cuts are fixed. Our main experiment measures both points directly. If per-mass optimization matches the front at equal budget, and the front needs nearly $K$ analyses to meet tolerance anyway, the multi-objective formulation adds nothing and we will say so.
Why not just sweep weights over a weighted sum of log-limits? A weighted sum can only reach the convex part of the trade-off curve, as both the practical multi-objective guide and the tree-ensemble paper show. Whether our front is non-convex is an empirical question. We test it by checking whether front points selected for the pre-registered set fall inside the convex hull, where no weighting can reach them. If the front turns out convex, the weighted sum stays a legitimate competitor and is judged only on budget.
Why not just use the fast "Asimov" approximation and run NSGA-II? The Asimov approximation estimates median sensitivity from a single idealized dataset in seconds, which would make cheap evolutionary search feasible. This alternative could plausibly win, so we test it head-on. We run NSGA-II on Asimov sensitivities with tens of thousands of evaluations, then re-score its front with full pseudo-experiments. Our hypothesis is that the approximation is biased exactly where expected counts are a few events near threshold, which is the regime that matters. If the re-scored Asimov front matches ours, the recommendation becomes "use Asimov plus NSGA-II," and the expensive surrogate machinery is unnecessary.
Why not just keep the hand-tuned analysis, since cuts cannot create signal below threshold? That is true: optimization does not lower the kinematic limit. But the analyst's choices decide how much of the deposited ionization survives micro-discharge and fiducial cuts. Those cuts are what separate an ionization-only search from a paired one. The replica baseline measures exactly how much is left on the table.
Proposed Experiments
The decisive experiment has four parts. All share one pseudo-experiment generator and seed policy, and run at budgets of 200, 500 and 1,000 simulator calls.
First, we run the main front search against per-mass optimization, weighted-sum sweeps and the published-configuration replica.
Second, we run the Asimov-plus-NSGA-II alternative, with the re-scoring described above.
Third, we stress-test against background misspecification. We optimize the front under one micro-discharge model and evaluate it under the held-out variants. A front whose advantage over the replica disappears under a held-out model is reported as a failure, not a result.
Fourth, we run the germanium transfer test against the CDEX-50 projected configuration, whose reported projected sensitivity is 5.1 × 10⁻⁴⁵ cm² at 5 GeV/c². Again, the comparison is against our replica's median expected limit, not the projected number.
The proposal succeeds if three conditions all hold under the held-out background models: two or three pre-registered analyses fall within 10% of the per-mass best envelope at every mass; the front is reached with substantially fewer calls than per-mass search; and the Asimov shortcut fails to reproduce the front near threshold.
References
- Search for Light Dark Matter with 259 Days of Data in PandaX-4T
- Deep learning based pulse shape discrimination for germanium detectors
- Determination of responses of liquid xenon to low energy electron and nuclear recoils using a PandaX-II detector *
- The design, implementation, and performance of the LZ calibration systems
- Multi-Objective Constrained Optimization for Energy Applications via Tree Ensembles
- Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces
- A Crowding Distance That Provably Solves the Difficulties of the NSGA-II in Many-Objective Optimization
- A practical guide to multi-objective reinforcement learning and planning
- Projected WIMP sensitivity of the CDEX-50 dark matter experiment
- Multi-Objective Optimization of a Path-following MPC for Vehicle Guidance: A Bayesian Optimization Approach