Hydrometallurgy papers report leaching efficiency as a single percentage. Leaching efficiency is the fraction of a target element moved from a solid into solution. Process conditions are then ranked by these numbers. Alkaline pressure oxidative leaching (APOL) is one such process: waste is heated in sodium hydroxide under pressurized oxygen so that one element dissolves and the rest stay solid. In the standard example, arsenic is pulled out of lead smelter flue dust at 95.6% efficiency while antimony, cadmium, indium and lead stay below 4.0% (Selective Separation of Arsenic from Lead Smelter Flue Dust by Alkaline Pressure Oxidative Leaching). Each such number is the ratio of two chemical assays. The first assay measures the dissolved element. The second measures the total in the starting material. Each assay carries biases that repeating the experiment will not reveal. A digestion can leave part of the sample undissolved, a precipitate can trap the target element, and liquor volume can shift during autoclave quenching. Particle physicists call these leftover, unmodeled biases residual systematic uncertainties. They handle them in two ways. The first is to measure the same quantity through two channels whose biases are independent and check that the channels agree. The second is to form ratios in which shared unknowns cancel.
This note proposes carrying that practice into leaching. Every APOL run would be measured twice. One channel is the solution side: dissolved arsenic times liquor volume, divided by the arsenic in the starting dust. The other is the solid side: one minus the arsenic left in the residue, normalized by an element that provably stays in the solid. The disagreement between the two channels gives a direct, per-run estimate of the systematic error that replicate standard deviations cannot see. The central claim is testable. For the small co-leaching fractions that the field currently optimizes against, such as antimony at a few percent, systematic error may exceed both replicate scatter and the differences between competing process conditions. If so, the ranking of conditions, and part of the "unresolved antimony removal" problem, rest on measurement rather than chemistry.
Core Claim
The two channels have complementary strengths, much like the two neutron-capture channels at Daya Bay. There, neutron capture on hydrogen and neutron capture on gadolinium gave independent measurements of the same oscillation parameter: $\sin^2 2\theta_{13} = 0.083 \pm 0.018$ from hydrogen alone and $0.089 \pm 0.008$ combined. The agreement bounded the detector systematics (Independent measurement of the neutrino mixing angle $\theta_{13}$ via neutron capture on hydrogen at Daya Bay).
In leaching, the solution channel measures a major element directly, which suits arsenic at 95.6%. For a minor co-leached element it also measures the small dissolved amount directly, but divides by a head assay (the measurement of the total in the starting material) whose completeness is uncertain. The solid channel measures the 4.4% of arsenic left behind directly, so it is precise for high recoveries. For antimony at under 4% dissolution, however, it subtracts two nearly equal large numbers and becomes imprecise. Combining the channels therefore does more than average them. It lets each element be estimated where the better channel applies. It also exposes the specific failure modes that damage each channel: incomplete head digestion, antimony trapped in precipitated sodium antimonate hydrate, and carbonate contamination from atmospheric CO₂.
The second ingredient is a ratio-cancellation step. It is modeled on reactor antineutrino analyses, which divided spectra from detectors at different distances. That division cancelled the unknown reactor flux and exposed a roughly 5% bias in the theoretical ²³⁵U flux prediction (Model-independent $\bar{\nu}_e$ short-baseline oscillations from reactor spectral ratios).
In APOL the residue mass is itself a nuisance quantity: it changes with filter losses, washing and drying, and with new solids that precipitate during the run. Instead of weighing it, the solid channel is normalized by the ratio of arsenic to an immobile tracer, measured in both residue and feed. For element $j$ with tracer $t$, the solid-channel extraction is
$$ E_j^{\text{solid}} = 1 - \frac{(c_j/c_t)_{\text{residue}}}{(c_j/c_t)_{\text{feed}}}\,(1 - E_t), $$
where $c$ denotes a measured mass concentration and $E_t$ is the small, separately bounded dissolution of the tracer. Lead is the natural tracer candidate, since the APOL paper reports it below 4% dissolution. Residue mass, moisture and sampling dilution cancel. The price is that the tracer's own dissolution must be bounded, which the solution channel can do.
Research Program
The work builds a joint statistical model for each run. It has one true extraction $E_j$ per element and separate nuisance parameters for the head-assay completeness, residue entrapment, liquor-volume error and tracer dissolution. Two measurement equations tie these to the two channels.
Nuisance priors come from dedicated side experiments rather than guesswork. Head digestion is repeated with and without a secondary complexing step. This mirrors the boric-acid fix for fluoride entrapment in Determination of Pb, Cd, Cr and Hg in Coal Gangue and Fly Ash Product with Microwave Digestion and Atom Absorption Spectrometry, a paper that motivates this work by showing that digestion completeness is a real, matrix-dependent bias. Liquor volume is tracked gravimetrically before and after quench. A subset of runs is carried out under inert gas so that carbonate contamination can be separated out.
Across the grid of process conditions, a response surface is fitted: a low-order polynomial in NaOH concentration, oxygen pressure, temperature and time. Its uncertainties use a sandwich (robust) covariance, which estimates parameter uncertainty from the curvature and the spread of per-run contributions together. This estimator stays valid when the polynomial is the wrong shape. Frequentist Uncertainties on Neural Density Ratios with wifi Ensembles supplies the method. It showed that sandwich covariances computed on a fixed basis reach nominal 68% coverage without retraining the model thousands of times. It also constrains the design, because it states that the estimate stays biased when the truth lies outside the basis. No neural basis is used here; the classical estimator is used on a small basis. The misspecification caveat is tested explicitly by comparing quadratic and cubic surfaces.
Generality is tested on a second feedstock with a different chemistry: sodium-roasted boron tailings, for which Effects of sodium roasting on the leaching rate of boron-bearing tailings and its mechanism analysis reports boron leaching rising from 58.78% to 86.78%. The question is whether a 28-point improvement is robust while smaller step-to-step gains are not.
Baselines
The naive reference point is the current reporting practice. Efficiency comes from the solution channel only, with the head assay taken as exact, and uncertainty is the standard deviation of three to five replicate autoclave runs. It is run on exactly the same runs and assays as the proposal, so the comparison costs nothing extra. The proposal beats it if the naive intervals under-cover the known truth on synthetic feeds (described below) while the two-channel intervals cover at the nominal rate.
The strongest existing method is the APOL protocol and analysis of Selective Separation of Arsenic from Lead Smelter Flue Dust by Alkaline Pressure Oxidative Leaching. It uses solution-channel efficiencies, one-factor-at-a-time optimization, and reports 95.6% arsenic leaching efficiency with under 4.0% for Sb, Cd, In and Pb at 80 g/L NaOH, 1.0 MPa O₂, L/S 8 mL/g, 120 °C and 2.0 h. No reported uncertainty on those figures was found on Lacuna. The optimum is reproduced with the identical condition set. The same number of autoclave runs is spent on the grid, so the proposal's added value comes from measurement design and not from more experiments.
The strongest calibration competitor anchors the head assay to a certified reference material of similar matrix, or to a gravimetrically spiked blank, and keeps single-channel reporting. It gets the same assay budget. This baseline directly attacks the largest bias on the solution side, so the proposal must beat it on the terms it cannot address: residue entrapment and volume errors.
For the response surface, a bootstrap is the comparison. The surface is refitted on 1000 resamples of runs, and its coverage and interval widths are compared with the sandwich estimator's. The goal here is agreement at lower cost, not a win.
Two ablations isolate the key ingredients. The first uses both channels but weighs the residue directly instead of using the tracer ratio. The second applies the tracer ratio but averages the channels without the joint nuisance model. Together they show whether the gain comes from the second channel, the cancellation, or the modeling.
Why not just…?
Why not just replicate runs and report the standard deviation? Replicates capture run-to-run scatter. A bias shared by every run is invisible to them, for example a head digestion that always misses 3% of the arsenic locked in silicate. The reactor-flux case in Model-independent $\bar{\nu}_e$ short-baseline oscillations from reactor spectral ratios is exactly a well-replicated measurement sitting on a biased normalization. The objection is still partly correct. A shared head-assay bias cancels when conditions are ranked on the same feed, so replicates may suffice for ranking arsenic conditions. The proposal checks this rather than assuming otherwise. If the closure residuals are no larger than replicate scatter, that component of the program is not worth its cost, and that is a publishable negative result.
Why not just calibrate against certified reference materials? Reference materials fix the head assay. They do not certify a fresh autoclave residue full of newly precipitated antimonate and carbonate, and they do nothing about liquor volume. This is the most serious competitor for the arsenic headline number, and it is included as a baseline. The proposal's advantage is predicted to appear mainly for minor co-leached elements and for residue-side entrapment. If the calibrated single-channel method matches two-channel coverage on the spiked feeds, the core claim fails for APOL.
Why not just check mass-balance closure, as many leaching papers already do? Reporting that inputs and outputs balance to within some percent is a pass/fail sanity check, and it relies on weighed residue mass. The proposal differs in three ways. It removes residue mass through the tracer ratio. It assigns the closure residual to named nuisance sources using side experiments. It propagates the result into the condition ranking. The first ablation, two channels without the tracer ratio, is essentially the conventional closure check made quantitative, so the experiment measures directly how much the extra structure adds.
Why not just fit a shrinking-core kinetic model or a black-box regressor to all the runs? A shrinking-core model, as used in Lithium Extraction from Lithium-Bearing Clay Minerals by Calcination-Leaching Method, idealizes each particle as a sphere dissolving from the outside in. Both it and a neural regressor fit whatever efficiency numbers they are given, so a biased input produces a biased model. Neither replaces measuring the inputs correctly, and either could be fitted afterward to the corrected two-channel estimates.
Proposed Experiments
The first experiment establishes ground truth. Synthetic feeds are prepared by mixing gravimetrically known amounts of an arsenic compound and antimony compounds into a lead-oxide/silicate matrix free of arsenic, so that true total and true extractable content are known. This plays the role of the one-dimensional Gaussian toy in the wifi-ensembles paper. Coverage of the naive, calibrated and two-channel intervals is measured over about 30 runs at the reference APOL conditions and at deliberately perturbed ones. The perturbations are a shorter time, lower oxygen pressure, and an air-exposed quench designed to induce carbonate formation.
The second experiment uses real flue dust. It runs a grid of roughly 25 conditions around the reported optimum, with two channels on every run for As, Sb, Pb, Zn, Cd and In.
The decisive checks are three. The first asks whether the distribution of closure residuals is wider than replicate scatter, especially for antimony. The second asks whether the arsenic-optimal condition and the antimony-minimizing condition stay distinguishable from their neighbors once systematic intervals are included. The third asks whether the channel disagreement for antimony tracks the amount of sodium antimonate hydrate identified by diffraction in the residue. That link would show that part of the "deep antimony removal" problem is residue entrapment measured inconsistently, rather than true dissolution.
The third experiment repeats the audit on the boron tailings series to test whether the pattern generalizes. The expected result is that the large roasting-temperature effect stays robust. The effects of intermediate steps, such as the soda-ash multiple, are the ones expected to lose significance under the full systematic budget.
References
- Selective Separation of Arsenic from Lead Smelter Flue Dust by Alkaline Pressure Oxidative Leaching
- Independent measurement of the neutrino mixing angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>θ</mml:mi></mml:mrow><mml:mrow><mml:mn>13</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>via neutron capture on hydrogen at Daya Bay
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- Determination of Pb, Cd, Cr and Hg in Coal Gangue and Fly Ash Product with Microwave Digestion and Atom Absorption Spectrometry
- Frequentist Uncertainties on Neural Density Ratios with wifi Ensembles
- Effects of sodium roasting on the leaching rate of boron-bearing tailings and its mechanism analysis
- Lithium Extraction from Lithium-Bearing Clay Minerals by Calcination-Leaching Method