Magnetic solid-phase extraction is one of the few practical ways to pull trace pollutants out of water without liters of organic solvent. A sorbent shell is grown on an iron-oxide core, the particles soak up the target molecules, and a magnet collects them in seconds. This only works for particles that do both jobs: they must be magnetic, and they must absorb. When such sorbents are scaled from a vial to a reactor, poor mixing produces three kinds of failed particles. Some are bare magnetic cores that absorb nothing. Some are free, non-magnetic lumps of sorbent that absorb analyte but stay in the water when the magnet is applied. Some are clumped aggregates with blocked pores. Standard quality control measures the batch in bulk, using X-ray diffraction and surface area. Those methods report the average crystallinity and the total surface, so they cannot tell how many particles are both magnetic and absorbing.
We propose to measure that joint fraction directly. Each particle passes through a flow cell with several cheap sensors: magnetic drift speed, light scattering, and fluorescence from a probe dye the particle has taken up. A statistical template fit, borrowed from how particle physicists estimate the yields of each particle species from incomplete multi-detector data, then converts the sensor readings into the fractions of good, bare-core, free-sorbent and aggregated particles. The claim is that this fraction vector, measurable in minutes, predicts downstream extraction recovery far better than bulk characterization does. It should also identify which mixing failure caused a bad batch.
Core Claim
In magnetic extraction, the recovered amount of analyte depends on the correlation between two particle properties, magnetic capture and uptake, not on either property averaged over the batch. Let $f_k$ be the fraction of particles in class $k$, with classes for good core-shell, bare core, free amorphous sorbent, and aggregate. Let $m_k$ be the probability that a class-$k$ particle is captured by the magnet, and $u_k$ its mean analyte uptake. Then the expected recovery scales as
$$R \propto \sum_k f_k\, m_k\, u_k ,$$
whereas bulk surface area tracks $\sum_k f_k u_k$ and bulk magnetization tracks $\sum_k f_k m_k$. The two failure modes trade off against each other in the bulk numbers. A batch with many free-sorbent particles shows high surface area and lower magnetization. A batch with many bare cores shows the reverse. Two batches can therefore look identical in bulk and differ sharply in recovery. Our contribution is to estimate the $f_k$ in-line and to test whether they explain recovery where bulk metrics fail.
The estimation step is where particle identification comes in. It supplies the method, not an analogy. Collider experiments separate pions, kaons and protons by combining per-detector likelihoods, as in the RICH and energy-loss combination of Danko et al. 2004. They extract species yields by fitting templates to the observed distributions rather than by counting hard classifications, the kind of signal-yield fit used in De Mori et al. 2023. ALICE has shown that most tracks lack at least one detector signal: 62.8% in Deja et al. and Karwowska et al. 2024. Both papers handle this with attention models that read only the measurements present. Flow-cell particle sensing has the same structure. A non-magnetic particle has no drift-speed reading, and a dark particle has no fluorescence reading.
Positioning Against Prior Work
The sorbents come from two groups of work that motivate the gap.
- Room-temperature route. Bao et al. 2022 grow a Schiff-base framework shell (TbBd) on amine-functionalized Fe₃O₄ at room temperature. They extract parabens with limits of detection of 0.2–0.4 µg/L and spiked recoveries of 86.1–110.8%. This was done only at vial scale, with morphology judged qualitatively by microscopy.
- Solvothermal route. Zhang et al. 2023 make Fe₃O₄@COF-300 solvothermally. They report Langmuir capacities of 173.31 mg/g for bisphenol A and 303.03 mg/g for ibuprofen. Again there is no per-particle characterization.
The synthesis-scaling literature supplies the reason to expect heterogeneity at scale.
- The Scalability Bottleneck in Nanostructured Material Synthesis argues that mixing time grows with reactor volume faster than nucleation time.
- Yan et al. 2024 describe how sensitive framework crystallization is to concentration and temperature.
Neither group measures how the defects are distributed across particles. We do, and we choose the room-temperature Bao route because it avoids sealed solvothermal vessels. That choice keeps the scaled process itself consistent with the solvent-reduction goal described in Green Alternatives to Solvent Intensive Chemical Workflows.
The particle-physics work also constrains the design.
- Classifiers trained on simulation. Graczykowski et al. 2022 and Temirkhanov et al. 2022 document that such classifiers degrade on real data. They also learn spurious links between kinematics and species. Temirkhanov et al. report that their decomposition method cut error on unseen muon decays by 97–98% relative to the standard LHCb network.
- What we take from this. We do not rely on a classifier trained once and frozen. We build templates from reference particle populations measured in the same flow cell on the same day. This is the materials analogue of the data-driven calibration samples physicists prefer. It addresses the shift between sample matrices, river water versus wastewater, more directly than adversarial training would.
We adopt the ALICE missing-data attention model as one way to produce per-particle likelihoods. We compare it against simpler options rather than assuming it is needed.
Research Program
Reference populations. The first component is a set of pure particle populations that define the templates:
- bare amine-Fe₃O₄ cores;
- TbBd framework synthesized with no cores (free sorbent);
- deliberately aggregated material, made by drying and redispersing without sonication;
- the best-performing vial-scale Fe₃O₄@TbBd, enriched by repeated magnetic washing.
Electron microscopy on several hundred particles per population checks that each is in fact dominated by its intended class.
Flow cell. The second component is a microfluidic cell with three stations:
- a magnetic gradient section, where drift speed perpendicular to flow reports magnetic moment over drag;
- a two-angle light-scattering station, reporting size and refractive contrast;
- a fluorescence station.
Before the cell, particles are briefly incubated with a fluorescent probe molecule chosen to adsorb into the framework pores the same way parabens do. Its per-particle signal is therefore a proxy for uptake $u$. Missing readings are expected and recorded as missing, not imputed. Non-magnetic particles leave the magnetic channel undefined, and low-uptake particles fall below fluorescence threshold.
Population fit. The third component is the fit. For each reference class we estimate a density $p_k(s)$ over the available sensor readings $s$, marginalizing over absent channels. We fit an unknown batch as the mixture $p(s)=\sum_k f_k\,p_k(s)$ by maximum likelihood, with bootstrap uncertainty on $f_k$.
The per-class densities come from one of three options:
- kernel density estimates per channel combined under an independence assumption, which is the classical product-likelihood approach;
- a gradient-boosted classifier converted to likelihood ratios;
- a set-attention network in the style of Karwowska et al. 2024.
Scale-up campaign. The fourth component synthesizes Fe₃O₄@TbBd across conditions that produce varied defect mixtures:
- vial scale;
- a 1 L stirred tank at three stirring rates and two monomer addition rates;
- a 10 L tank;
- a continuous coaxial-flow reactor.
The target is roughly forty batches. Each batch gets bulk X-ray diffraction, nitrogen surface area, bulk magnetization, the flow-cell fit, and the full Bao paraben extraction on spiked river water with HPLC-UV.
Baselines
Bulk characterization (do-nothing reference). This is what the field currently does. We regress paraben recovery on bulk crystallinity, surface area and saturation magnetization across the same batches, using the same cross-validation folds as our method. Our method beats it if the fitted fractions explain recovery with clearly higher held-out R², and specifically on the subset of batch pairs that bulk metrics cannot distinguish.
Single-sensor cut. A particle is called good if its magnetic drift speed and its fluorescence each exceed thresholds tuned on the reference populations. This is the $n_\sigma$-style cut that ALICE used before machine learning. Particles missing a channel are discarded or counted as failures. The cut gets the same reference data and threshold search budget as our method. Beating it means lower error on known-fraction mixtures, especially when free-sorbent particles, which lack a magnetic reading, are abundant.
Gradient-boosted trees with native missing-value handling. This is the strongest simple competitor, and it may match the attention model. The trees are trained on the same labelled reference particles, with a hyperparameter budget equal to the attention model's. We compare them under classify-and-count and under template fitting.
Attention model (strongest existing method). This is the per-particle classifier of Karwowska et al. 2024 and Deja et al.. The file reports that it beat $n_\sigma$ cuts and imputation baselines in F1, precision and recall on the ALICE Run 2 simulation. No reported scalar F1 value was found on Lacuna for that model.
Ablations. The key ingredient is the joint per-particle measurement of capture and uptake, so the main ablation drops the fluorescence channel. A second ablation replaces template fitting with classify-and-count. This tests whether the yield-fitting idea from physics, rather than the classifier, carries the accuracy under matrix shift.
Why not just…?
Why not just run a small extraction test on every batch? A mini-extraction directly measures the quantity we care about, so it is a real threat. If the only goal were pass or fail, it would likely do as well as our method. It falls short in two ways that we will measure.
- Speed. Equilibrium takes tens of minutes, and Zhang et al. 2023 report that bisphenol A reached only 69.4% of capacity within 60 minutes. The flow-cell fit should take a few minutes.
- Diagnosis. A low recovery does not say whether to slow monomer addition (too many free-sorbent particles) or increase core coverage time (too many bare cores).
We will report time to decision for both methods. We will also run a diagnosis test: on held-out failed batches, does the dominant fitted defect class predict which process change restores recovery? If the mini-extraction is equally informative, our contribution reduces to a faster proxy, and we will say so.
Why not just use a gradient-boosted classifier and skip the attention model? We expect this may suffice. Trees already handle missing values natively, and the flow cell has only three stations, not the dozens of features in ALICE. The proposal's claim does not rest on the network architecture. If trees match attention on fraction error, we adopt trees.
Why not just add bulk magnetization to bulk surface area? Two bulk numbers still give only the two marginal sums, not the joint sum that governs recovery. The decisive test is the matched-bulk-pair analysis below.
Why not switch to deep eutectic solvents or water-phase probes and avoid solid sorbents?
- Deep eutectic solvents are viscous and need prolonged ultrasound or centrifugation (Zhang et al. 2025).
- Water-phase optical probes such as Zhou et al. 2025 target specific analyte classes, long-chain cationic surfactants, and do not cover neutral pollutants like parabens.
Magnetic sorbents remain worth scaling, which is why their manufacturing control matters.
Proposed Experiments
Calibration on known mixtures. We mix reference populations in about twenty known ratios and measure them in two water matrices. We report the mean absolute error of the estimated $f_k$ against the true fractions, for every method and ablation. The product-likelihood fit with same-day templates should keep error low even when matrix shift distorts scattering. Classify-and-count with a frozen classifier should degrade.
Recovery prediction. Across the roughly forty scaled batches, we compare held-out R² for predicting paraben recovery from three inputs: fitted fractions, bulk metrics, and the single-sensor cut. As a sanity anchor, our best vial batch should reproduce Bao's 86.1–110.8% recovery range.
Decisive check: matched bulk pairs. We identify batch pairs whose crystallinity, surface area and magnetization agree within measurement error. The proposal succeeds if the fitted $\sum_k f_k m_k u_k$ correctly orders their recoveries significantly more often than chance. It fails if bulk-matched batches turn out to have matched recoveries, which would mean defect sub-populations do not matter at the scales tested.
Process attribution. For the stirred-tank series, we test whether the free-sorbent fraction rises with addition rate and the bare-core fraction rises with stirring intensity. Each defect class should map to a distinct, correctable process cause. This is the information a closed-loop synthesis platform of the kind proposed in the materials acceleration platform study would need as feedback.
References
- Measurement of the decay rate of Ξc0→pK-K-π+ relative to Ξc0→Ξ-π+
- Measurements of the branching fractions of the inclusive decays <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo stretchy="false">(</mml:mo><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo stretchy="false">)</mml:mo><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mi>X</mml:mi></mml:math>
- Machine-learning-based particle identification with missing data
- Particle identification with machine learning from incomplete data in the ALICE experiment
- Analysis of parabens in environmental water samples by covalent organic framework-based magnetic solid-phase extraction-high performance liquid chromatography
- Adsorption Performance of Magnetic Covalent Organic Framework Composites for Bisphenol A and Ibuprofen
- COx hydrogenation to methanol and other hydrocarbons under mild conditions with Mo3S4@ZSM-5
- Bioinspired design of Na-ion conduction channels in covalent organic frameworks for quasi-solid-state sodium batteries
- Separation and Detection of Catechins and Epicatechins in Shanxi Aged Vinegar Using Solid-Phase Extraction and Hydrophobic Deep Eutectic Solvents Combined with HPLC
- Using machine learning for particle identification in ALICE
- Robust Neural Particle Identification Models
- Ultrasound-Assisted Determination of Selenium in Organic Rice Using Deep Eutectic Solvents Coupled with Inductively Coupled Plasma Mass Spectrometry
- A Phthalocyanine Optical Probe Responding to Cationic Surfactants Containing Long Carbon Chains with High Selectivity in Total Water Phase and Its Applications
- Designing materials acceleration platforms for heterogeneous CO2 photo(thermal)catalysis