{"id":"943fd115-a0f4-4959-a803-b4a88a78632e","arxiv_id":"2507.19316","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Raising the cold reactor temperature in a two-reactor continuous crystallizer reduces magnesium impurity in the final lithium carbonate, and the authors used a human-in-the-loop active learning framework to find this relationship.","lead":"This paper describes a human-in-the-loop active learning workflow used to optimize a continuous crystallization process for battery-grade lithium carbonate. The authors report that raising the cold reactor temperature lowers magnesium impurity in the product, which could allow cheaper, lower-grade lithium feedstocks to be used.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 6000 ppm Mg-tolerance claim is GPC extrapolation: no run above 3000 ppm initial Mg succeeded, and bracketing runs at 4028-6225 ppm at high T_cold all failed, so the abstract's number is unsupported.","rationale":"I read the paper as making two distinct claims: (1) the HITL-AL framework efficiently identified that elevated cold-reactor temperature improves Mg rejection in this dual-reactor continuous crystallizer, and (2) this raises initial-Mg tolerance from roughly 80 ppm to several thousand ppm, with the abstract specifying 6000 ppm. Claim (1) is genuinely well supported by the tabulated data: matched-pair comparisons (Exp. 38/40 vs 39/42 at 743 ppm initial Mg; Exp. 62/63/66 vs 61/65 at 1300 ppm) show final Mg falling from 100-684 ppm at T_cold ~ 10-15°C to 3-60 ppm at T_cold = 63-73°C, and the trend reproduces across different feedstocks. Claim (2) is the load-bearing weak point. My independent read of Table S4 confirms the reader's weakest assumption and sharpens it: the extrapolation region is not merely empty of positive evidence, it is bracketed by direct failures at 4028-6225 ppm initial Mg run at the very high T_cold values the claim relies on (63-68°C). The most defensible statement of the result is 'tolerance up to about 3000 ppm was demonstrated,' not 'as high as 6000 ppm.' The main caveat is that the bracketing runs differ in co-contaminant levels, so the boundary failure is demonstrated only for the feeds actually tested; a controlled sweep holding other conditions fixed is the decisive test. Additional concerns reinforce conditionality rather than change the verdict: the post-hoc exclusion of 'failed' experiments noted in Sec. 2.1, with Exps. 31, 73, and 77 absent from Table S4 without explanation; the in-silico comparison in Fig. 9 is partly circular because the surrogate GPC is trained on HITL-derived data and the 'informed' dataset encodes the HITL result that it is then credited with discovering; and the absence of error bars or reproducibility statistics for the ICP-OES measurements. Credit where due: the forced-trials, matched-pair temperature comparisons, and per-run ICP results give the directional finding real independent weight, and the GPR-derived discovery of the temperature correlation is plausible. The reader's CONDITIONAL verdict is the right calibration: the directional core of the paper likely survives, but the abstract's 6000 ppm claim must either be supported by controlled experiments in the 4000-6000 ppm range or revised down to the demonstrated ~3000 ppm.","tokens_in":21214,"tokens_out":12558,"duration_ms":103982,"concrete_test":"Controlled Mg sweep in the claimed success region: replicate Exp. 45's other conditions (T_hot ~ 89°C, flow 5 mL/min, slurry 6 g/100 mL, Ca ~ 196,366 ppm, K ~ 764 ppm, Li ~ 84,126 ppm, Na ~ 9,651 ppm) and run at least three replicates at initial Mg = 4000, 5000, and 6000 ppm, with T_cold at the values where Figure 7's boundary predicts battery-grade (roughly 66-75°C). If the majority of replicates at any Mg >= 4000 ppm reach final Mg < 80 ppm, the 6000 ppm claim stands; if all fail, the claim must be revised to the experimentally demonstrated ~3000 ppm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim (abstract, Fig. 7, Conclusions) is that raising the cold-reactor temperature beyond 60°C lifts initial-Mg tolerance 'to as high as 6000 ppm.' Table S4 shows the highest initial Mg ever yielding a battery-grade product (final Mg < 80 ppm) is Exp. 45: 3000 ppm initial Mg, T_cold = 66°C, final Mg = 73.9 ppm, only 6 ppm below the threshold. Every experiment above 3000 ppm failed, including the high-T_cold runs closest to the claimed boundary: Exp. 78 (4028 ppm, T_cold 63°C -> 122 ppm), Exp. 48 (5000 ppm, T_cold 68°C -> 189.8 ppm), Exp. 68 (6225 ppm, T_cold 65°C -> 464 ppm), Exp. 55 (7441 ppm, T_cold 79°C -> 933 ppm). The Fig. 7 GPC boundary must therefore cross the 3000-6000 ppm interval through a region with zero positive observations, bracketed by negative ones. With a Matern-kernel GPC (Table S6, length_scale ~0.3, ~80 curated points), the boundary there is set by the kernel's inductive bias and the nearest negative training points, not by evidence; the 6000 ppm number is prior-driven. These bracketing runs differ from Exp. 45 in Ca/K/Na co-contaminants and flow/slurry settings, so they do not formally falsify a narrowly conditioned boundary; but no feedstock at >= 4000 ppm has produced battery-grade material under any tested condition, and no variant of the 6000 ppm claim has direct support. The directional mechanism (higher T_cold improves Mg rejection) is separately and convincingly supported by matched pairs (Exps. 38/40 vs 39/42; 62/63/66 vs 61/65), so the fix is to the magnitude claim, not the mechanism. Compounding factors: 'significantly divergent' experiments were post-hoc excluded (Sec. 2.1), and Table S4 skips Exps. 31/73/77 without comment, so the fitted boundary rests on curated data; Exp. 45's 73.9 ppm margin is also within typical ICP-OES uncertainty, weakening even the 3000 ppm anchor.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a human-in-the-loop active learning (HITL-AL) framework for optimizing continuous crystallization of lithium carbonate from high-impurity brines, with a focus on magnesium rejection. The central process claim is that raising the cold-reactor temperature beyond the conventional 60°C limit substantially improves Mg removal, increasing tolerable initial Mg concentrations from roughly 80 ppm (or a few hundred ppm) to as high as 6000 ppm. The experimental program comprises 80 runs, the first 38 of which were used to identify the temperature effect; the authors also compare HITL-AL against random and Bayesian active learning baselines using informed and uninformed simulated datasets. The paper includes statistical analyses (Pearson correlation, SHAP, sensitivity), a GPC-based decision boundary, and makes code and data available on GitHub.","tokens_in":21634,"tokens_out":4613,"duration_ms":43149,"significance":"If the temperature effect and the claimed tolerance levels hold, the work provides a practically valuable process insight for lithium purification from lower-grade brines and a useful demonstration of HITL-AL in a low-data chemical optimization setting. The matched-pair experiments (e.g., Exps. 38/40 vs. 39; 62 vs. 63/65) directly support the directional claim that higher cold-reactor temperature reduces final Mg concentration, and the provision of code and data on GitHub is a concrete reproducibility strength. However, the headline quantitative claim of 'as high as 6000 ppm' initial Mg tolerance is not supported by the experimental record and rests on extrapolation from a Gaussian process classifier trained on data with no successful run above 3000 ppm initial Mg.","major_comments":[{"comment":"The claim that the process tolerates initial Mg 'as high as 6000 ppm' is unsupported by the experimental data. In Table S4, the highest initial Mg that yields a battery-grade product (final Mg < 80 ppm) is Exp. 45 at 3000 ppm initial Mg, with a final Mg of 73.9 ppm, only 6 ppm below the threshold. Every run above 3000 ppm failed, including Exp. 78 (4028 ppm, T_cold 63°C -> 122 ppm), Exp. 48 (5000 ppm, T_cold 68°C -> 189.8 ppm), Exp. 68 (6225 ppm, T_cold 65°C -> 464 ppm), and Exp. 55 (7441 ppm, T_cold 79°C -> 933 ppm). The GPC boundary in Figure 7 therefore crosses the 3000–6000 ppm interval through a region with zero positive observations and negative bracketing points; with the Matern kernel length_scale of 0.3 listed in Table S6, the boundary location in that interval is set by the kernel's inductive bias rather than by evidence. The quantitative claim should be revised to reflect the demonstrated 3000 ppm limit, with any 'several thousand ppm' statement presented explicitly as an unverified model extrapolation, or the claim should be supported by new experiments in the 3000–6000 ppm range.","section":"Section 3 (Figure 7) and Abstract"},{"comment":"The comparison against 'human-independent' active learning is not a fair test of the HITL interaction. The 'informed' dataset used for the Bayesian and random baselines is 'explicitly constrained by parameter ranges informed by the HITL-identified optimal temperature settings and impurity conditions.' This means the baselines are given the central process insight (elevated cold-reactor temperature) without any human-in-the-loop experimentation, so the reported 67% vs. 14% success rates demonstrate the value of that prior information rather than the value of the HITL workflow itself. The authors should either compare against baselines that receive the same experimental data stream without human intervention, or reframe the result as an ablation of informed priors rather than a demonstration of 'human-AI synergy.'","section":"Section 3 (Figure 9) and active learning comparison"},{"comment":"The baseline Mg tolerance is stated inconsistently, which changes the claimed improvement factor. The results section states that with initial Mg above approximately 80 ppm 'none of the preliminary experiments successfully reduced Mg below this desired limit,' and later that above 200 ppm success was 'consistently unattainable'; the abstract, however, describes the improvement as being from 'industry practices at a few hundred ppm.' These baselines differ by roughly an order of magnitude and should be harmonized, with explicit distinction between the conventional initial-Mg tolerance and the battery-grade final-Mg specification.","section":"Section 3 (Mg challenge) and Abstract"}],"minor_comments":[{"comment":"Experiment numbers 31, 73, and 77 are missing from Table S4; please renumber the entries or explicitly state that those runs were excluded, and if so, why.","section":"Table S4"},{"comment":"Figure 7 projects a GPC decision boundary onto the initial-Mg vs. T_cold plane, but the model was trained on additional features (Ca, K, Li, Na, flow rate, slurry concentration, temperature differential). Please clarify how the other features are fixed for this projection and whether the boundary location is sensitive to those choices.","section":"Figure 7"},{"comment":"The criteria for excluding 'failed' experiments after reproducibility tests are not quantified; providing the threshold or the number of excluded runs would improve transparency and reproducibility of the training dataset.","section":"Section 2.1"},{"comment":"The phrase 'about more than approximately 1,024 experiments' is awkward and the factorial count depends on the number of levels per variable, which is not stated; please rephrase with a clear calculation.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"I would ask the editor to ensure the revision directly addresses the 6000 ppm claim; as written, the abstract is likely to be reported in a way that overstates the experimental evidence. The GitHub repository availability is a genuine strength that should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one for the gap between its abstract and its own data, but there is a real result underneath.\n\nThe real result: in their hot-cold continuous crystallizer, raising the cold reactor temperature into the 63–70°C range — past the conventional ≤60°C limit — substantially improved magnesium rejection. The matched pairs with identical feed composition are the cleanest evidence: at ~1300 ppm initial Mg, cold reactor at 15°C gave 100–155 ppm final Mg, while 53–70°C gave 34–60 ppm. That counterintuitive, experimentally validated effect is new relative to the cited literature and could matter for low-grade brine processing. The HITL-AL workflow (GPR with Pareto fronts, random walkers, SHAP screens, GPC boundary exploitation) is competently assembled; the authors are transparent that model suggestions were filtered through human judgment; and the GitHub repo with data and scripts makes the experimental record checkable.\n\nThe soft spots, in proportion.\n\nFirst, the abstract's headline — 'as high as 6000 ppm' Mg tolerance — is an extrapolation. The highest initial Mg that ever produced battery-grade product is Exp. 45: 3000 ppm initial, finishing at 73.9 ppm final, six ppm inside the 80 ppm cutoff and close to typical ICP-OES uncertainty. Every run above 3000 ppm failed, including the high-cold-temperature runs at 4028, 5000, and 6225 ppm initial Mg (final 122, 190, 464 ppm). The GPC boundary in Fig. 7 that would sustain a 6000 ppm claim crosses a region with zero positive observations, bracketed by failures; that boundary is kernel inductive bias, not evidence. The body's softer 'several thousand ppm' phrasing still leans on the same unverified extrapolation. The honest claim is 'up to roughly 3000 ppm,' or the authors run the bracketing experiments.\n\nSecond, the HITL-versus-autonomous comparison at the end is partly circular. The 'informed' surrogate space already encodes the human's temperature discovery, so the 67%-vs-14% success rates show the value of prior information, not of the human-in-the-loop mechanism itself. Fine as a simulation; the framing overclaims.\n\nThird, data curation: Section 2.1 says 'significantly divergent' experiments were excluded post hoc, and Table S4 skips Exp. 31, 73, and 77 without comment.\n\nFourth, Table S4 carries no error bars anywhere; the single 3000 ppm success is a one-run anchor.\n\nThe process chemistry reads sound and the central mechanism claim holds up; the fixes are to the magnitude claim, the baseline framing, and the curation disclosure. This paper deserves a serious referee, with heavy revision expected. For hydrometallurgy and autonomous-discovery readers, the temperature finding is worth engaging with, and the claims-versus-evidence gap is itself a useful cautionary case.","headline":"The cold-reactor temperature effect is real and experimentally grounded; the '6000 ppm Mg tolerance' headline is GPC extrapolation with zero experimental support above 3000 ppm.","tokens_in":22266,"tokens_out":8706,"would_cite":true,"duration_ms":71626,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a human-in-the-loop active learning framework can raise the magnesium tolerance of continuous lithium carbonate crystallization from roughly 80 ppm to several thousand ppm by running the cold reactor hotter than…","keywords":["human-in-the-loop active learning","continuous crystallization","lithium carbonate","magnesium impurity tolerance","Gaussian process classifier","battery-grade lithium","low-grade brine","process optimization"],"falsifier":"Run the continuous crystallization with initial Mg at 4000–6000 ppm, cold reactor at 70–80 °C, and otherwise matched conditions, then measure final Mg in the product; if final Mg does not stay below 80 ppm, or the product purity drops below battery grade, the extrapolated tolerance claim is false.","tokens_in":1387,"feed_emoji":"🔋","tokens_out":1369,"duration_ms":75469,"temperature":0.7,"pith_summary":"The paper reports that a human-in-the-loop active learning framework, combining Gaussian process models with expert judgment, found a counterintuitive lever for producing battery-grade lithium carbonate: running the cold reactor of a continuous crystallizer well above the conventional 60 °C limit greatly improves rejection of magnesium impurities. The authors claim this raises tolerable initial magnesium contamination from the usual ~80 ppm battery-grade threshold to several thousand ppm, potentially 6000 ppm. If correct, lower-grade brines with high magnesium loads could be processed directly into battery-grade lithium carbonate with reduced pre-refinement. The paper also argues that human guidance, not automation alone, was what uncovered the effect: 38 experiments sufficed to establish the key temperature relationship, while simulated uninformed active learning on broad parameter ranges found battery-grade conditions far less often.","feed_headline":"Hotter cold reactor lifts magnesium tolerance to 6000 ppm","feed_subtitle":"If confirmed, lower-grade brines would need far less magnesium pre-refinement before crystallization.","key_machinery":"The quantitative core of the work is a Gaussian process classifier (GPC), a probabilistic model that assigns each combination of initial impurity levels and reactor settings a probability of yielding battery-grade lithium carbonate. Its 0.5-probability contour is the decision boundary used to define how much initial magnesium can be tolerated. A ray-tracing algorithm selects experimental candidates nearest that boundary, and human experts adjust the surrogate space and feature ranges based on statistical diagnostics. The experimentally discovered inverse relation between cold reactor temperature and final magnesium concentration is the physical signal that the GPC encodes and extrapolates.","core_discovery":"The paper's central claim is that continuous crystallization can produce battery-grade lithium carbonate from feedstocks carrying far more magnesium than the conventional ~80 ppm limit, provided the cold reactor is run hotter than previously recommended. The authors experimentally validated a counterintuitive inverse correlation: raising the cold reactor temperature (to roughly 68–80 °C in successful runs, versus the old ≤60 °C rule) lowers final Mg below the 80 ppm battery-grade threshold even when initial Mg is in the hundreds to several thousand ppm range. They formalize the boundary with a Gaussian process classifier whose decision frontier, at class probability 0.5, predicts battery-grade outcomes for initial Mg up to about 6000 ppm when cold reactor temperature is high enough. The claim is supported by 80 total experiments; 38 were needed to establish the temperature effect. No experiment with initial Mg above 3000 ppm appears in the tables as a battery-grade success, so the 6000 ppm tolerance rests on extrapolation of the fitted boundary rather than direct demonstration.","pith_inferences":["The paper does not establish the physical mechanism behind the temperature effect; a solubility or dissolution-kinetics study at cold-reactor temperatures between 60 and 80 °C would test whether faster re-dissolution of impure solids is the cause.","A direct confirmatory run at 4000–6000 ppm initial Mg with a hot cold reactor would convert the extrapolated 6000 ppm tolerance from a model boundary into a demonstrated result.","The simulated comparison against uninformed Bayesian and random search uses a surrogate trained on the HITL data; a fully closed-loop comparison on fresh physical experiments would be a stronger test of the human-in-the-loop advantage.","The same decision-boundary active-learning recipe could generalize to other impurities and other continuous crystallizations, but that generalization is an extrapolation beyond the paper's evidence."],"forward_implications":["Feedstocks with hundreds to low-thousands of ppm Mg can be fed to the crystallizer and still yield battery-grade lithium carbonate when the cold reactor is hot enough.","The old rule of thumb (cold reactor ≤60 °C and at least 20 °C differential between reactors) is not a universal constraint; operating beyond it improves magnesium rejection.","Pre-refinement steps targeting magnesium can be reduced or skipped for many lower-grade brines, cutting water, reagent, and energy use.","Expert-guided active learning found the key parameter in 38 experiments, far fewer than the roughly 1,024 runs a full factorial design would require, and more reliably than uninformed computational search in the paper's simulated comparison.","The Gaussian process decision-boundary method gives an explicit operating map from initial Mg and cold-reactor temperature to battery-grade outcome, enabling process control decisions."],"supporting_citations":[{"why":"Supplies the continuous crystallization technique that the optimization targets.","marker":"[16]"},{"why":"Defines the mixed-suspension mixed-product-removal reactor concept used in the two-reactor setup.","marker":"[35]"},{"why":"Provides the Gaussian process formalism underlying both the regressor and classifier.","marker":"[36]"},{"why":"Supports using Gaussian process models for active learning of physical behavior in low-data settings.","marker":"[37]"},{"why":"Establishes the earlier AI-based optimization of battery-grade lithium carbonate that this framework extends.","marker":"[25]"},{"why":"Sets the battery-grade magnesium threshold used to classify outcomes and define the decision boundary.","marker":"[41]"}],"fun_headline_variants":["AI finds hotter cold side purifies lithium carbonate","Turn up the cold reactor: AI boosts Mg tolerance","AI-guided heat boost lifts magnesium limit to 6000 ppm","Hotter cold reactor: key to 6000 ppm Mg tolerance","Crank heat on cold reactor to hit 6000 ppm Mg"],"cache_read_input_tokens":24192,"weakest_assumption_plain":"The main load-bearing premise is that the model's boundary, learned from runs with up to 3000 ppm of initial magnesium, correctly predicts success near 6000 ppm even though no run at that level actually succeeded in the reported experiments.","fun_headline_variants_meta":{"raw":{"variants":["AI finds hotter cold side purifies lithium carbonate","Turn up the cold reactor: AI boosts Mg tolerance","AI-guided heat boost lifts magnesium limit to 6000 ppm","Hotter cold reactor: key to 6000 ppm Mg tolerance","Crank heat on cold reactor to hit 6000 ppm Mg"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001188,"raw_usage":{"total_tokens":4923,"prompt_tokens":985,"completion_tokens":3938,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":3855}},"tokens_in":601,"tokens_out":3938,"duration_ms":30387,"temperature":1.0,"reasoning_tokens":3855,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:54:50.759874+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the continuous crystallization with initial Mg at 4000–6000 ppm, cold reactor at 70–80 °C, and otherwise matched conditions, then measure final Mg in the product; if final Mg does not stay below 80 ppm, or the product purity drops below battery grade, the extrapolated tolerance claim is false.","supporting_citations":[],"review_version":1}