{"id":"b3d60968-433c-467c-900b-0eb1c04c4afb","arxiv_id":"2509.10705","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Metastable phase-separated states in multicomponent liquids can store and retrieve compositional information, as shown in a Hopfield-liquid model with matching simulations.","lead":"This paper develops a thermodynamic theory for when phase-separated mixtures of many liquids can be metastable, and shows that such states can act as associative memories by retrieving stored compositions from partial cues. It applies the theory to a liquid version of a Hopfield neural network, with spatial simulations matching analytical predictions.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sec. IVB 'iff' theorem fails for the bulk functional (7): without surface tension, non-global common tangents admit infinitesimal droplets with negative bulk energy, so positive phase Hessians are not sufficient for metastability.","rationale":"The reader's weakest assumption identified the omission of surface tension in Eq. (7) as the critical premise. I agree with that identification but sharpen it: surface tension is not merely needed to stabilize soft modes; it is what provides the nucleation barrier that makes non-global common-tangent states locally stable at all. Without surface tension, a phase-separated state whose tangent plane is not the global convex hull is unstable to infinitesimal nucleation of a composition lying below that plane. Positive Hessians at the P phases only ensure convexity in a neighborhood of each tangent point, not that the tangent plane is globally supporting. The paper's proof in Appendix E restricts attention to perturbations of the existing compartments; the necessity argument splits an unstable phase, but the sufficiency argument never analyzes adding a compartment with an entirely new composition. The binary quartic example in Sec. VI provides a concrete instance: families II and III are classified as metastable from Hessian checks, but in the bulk functional (7) they are saddle-like with respect to nucleating the global family I composition. The simulations work because the Cahn-Hilliard gradient term supplies a surface-energy barrier, which is absent from the analytical criterion. This affects the Hopfield application directly: the green regions in Fig. 5 and the retrieval claim rest on the theorem as stated, so the analytical basis for 'metastable phase separation' is incomplete. The issue is significant but likely fixable by restating the theorem for local stability within a fixed phase inventory and treating nucleation barriers with surface tension explicitly, or by including surface energy in the stability functional. Therefore the reader's CONDITIONAL verdict remains appropriate, but the condition should explicitly require amendment of the theorem's sufficiency claim.","tokens_in":43293,"tokens_out":21032,"duration_ms":283568,"concrete_test":"Use the binary quartic model of Sec. VI at c=135. Solve stationarity conditions (I10)-(I14) for the family II tangent endpoints (φ_A, φ_B) and common slope μ; also solve for the family I high binodal φ_high. Compute g(φ_high) = f(φ_high) - [f(φ_A) + μ(φ_high - φ_A)]. If g(φ_high) < 0, then an infinitesimal droplet of the global high phase lowers the bulk free energy (7), contradicting the Sec. IVB theorem in the absence of surface tension. A complementary check: repeat the Cahn-Hilliard simulations of Fig. 4B with the surface-tension coefficient k reduced by a factor of 10 (or equivalently, a larger box at fixed dimensionless k) and measure whether the family II state decays to family I on the simulation time scale; if it does, the reported metastability is an artifact of surface tension, not of positive Hessians.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central theorem (Sec. IVB) states that a phase-separated state is metastable iff each phase Hessian h^(c) is positive definite, with the only exception soft volume modes. The proof establishes stability with respect to perturbations of the existing compartments and to splitting one unstable phase, but it does not establish stability with respect to adding an infinitesimal compartment with a new composition. For the bulk functional (7), the first-order change from nucleating an infinitesimal volume ε of composition φ* is ε [f(φ*) - μ·φ* - (f^(c)-μ·φ^(c))], the signed distance to the common tangent plane. Positive definiteness of the h^(c) does not imply this distance is nonnegative for all φ*; it only implies local convexity at the P tangent points. A non-global common tangent—exactly the families II and III of the binary quartic model in Sec. VI—has compositions below its extrapolated tangent plane (the family I binodal), so in the bulk free energy (7) an infinitesimal droplet of those compositions lowers F linearly. The states are metastable in the Cahn-Hilliard simulations only because the k∇²φ term adds a surface-energy barrier, which is absent from Eq. (7). Thus the 'iff' theorem is false for the model it is derived from, and the analytical metastability regions in Fig. 5B-D (computed from phase Hessians only) do not by themselves establish that the retrieval states are metastable. The fix is to state the criterion as local stability within the fixed phase set and treat nucleation barriers/surface tension separately, or to include surface energy in the functional.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a thermodynamic formalism for metastable phase-separated states in multicomponent incompressible mixtures. The authors derive stationarity conditions (common tangent plane) and argue that a phase-separated state is metastable if and only if each constituent phase's free-energy Hessian is positive definite, with an exception for soft volume modes. They illustrate the criterion on a binary quartic model, where multiple common-tangent families arise, and on the liquid Hopfield model in the canonical ensemble, where stationary target/anti-target phases have an overlap a* determined by a self-consistent equation. The analytical predictions are compared with Cahn-Hilliard simulations that include surface tension. The paper claims that Hopfield liquids can retrieve information from partial cues via metastable phase separation and that the number of coexisting retrieved phases grows with the number of components.","tokens_in":43727,"tokens_out":10970,"duration_ms":122697,"significance":"If the central equivalence held, it would reduce the difficult question of metastability of multiphase states to the local stability of individual phases, which would be a strong and useful result. The geometric common-tangent picture, the application of associative-memory ideas to liquids, and the quantitative simulation strategy are all original and timely. The stationarity derivations in Sec. III and Appendix D are clean; the overlap a* is obtained by solving Eq. (33), not fitted to simulation output; and the spatial simulations with N up to 24 and multiple encoded targets provide falsifiable, quantitative predictions. However, as detailed below, the central theorem as stated is not correct for the bulk free-energy functional used in its proof, so the analytical metastability regions and the abstract's 'iff' claim require substantial revision.","major_comments":[{"comment":"The sufficiency half of the 'iff' theorem does not prove metastability with respect to nucleation of new phases. The perturbation expansion in Eq. (14) is restricted to perturbations of the C existing compartments (with the C-th eliminated by the constraints). For a stationary state, the first-order free-energy change from adding an infinitesimal compartment of volume ε and composition φ* is ε[f(φ*) − μ·φ* − (f^(c) − μ·φ^(c))], i.e., the signed distance to the common tangent plane. Positive definiteness of h^(c) at the P tangent points implies local convexity at those points, but not that the tangent plane lies below f(φ*) for every φ*. The binary quartic model provides a concrete counterexample: in Sec. VI, families II and III are non-global common tangents (Fig. 3A-B); for c=120, the family-I binodal lies below the family-II tangent plane, so an infinitesimal droplet of family-I compos","section":"Sec. IVB, Eq. (14) and Appendix E"},{"comment":"The proof also conflicts with the paper's own treatment of homogeneous metastability. Appendix B states that nucleation of phases with entirely different composition is discarded because such perturbations are not small and require a finite nucleation barrier. Exactly the same caveat applies to nucleation of a new phase inside a phase-separated state. The quadratic form (14) analyzes only infinitesimal perturbations in δ and ε and cannot capture the finite-composition, infinitesimal-volume nucleation channel. Therefore the claimed equivalence between metastability of the phase-separated state and positive Hessians of its phases conflates local stability within a fixed phase set with true metastability. The theorem should be restated as a local-stability condition for the fixed phase set, with an additional condition (supporting tangent plane, or sufficiently strong surface tension) for m","section":"Sec. IVB and Appendix B"},{"comment":"The analytical Hopfield stability regions are computed from the individual phase Hessians through inequalities (35)-(36), relying on the theorem of Sec. IVB. Because that theorem is unproven for the bulk functional, the green regions in Fig. 5B-D are not established as metastable retrieval regions: they do not rule out the existence of a composition φ* with f(φ*) below the common tangent plane, which would make the retrieval state unstable to infinitesimal bulk nucleation. The simulations in Fig. 6 are encouraging and match the predicted overlap a* (Eq. 33), but they use surface tension and specific finite domains; they do not validate the analytical boundaries. Please provide a supporting-tangent-plane check for the Hopfield free energy (or an explicit surface-tension criterion) before claiming that 'Hopfield liquids can retrieve information from partial cues via metastable phase separa","section":"Sec. VII.C and Fig. 5B-D"}],"minor_comments":[{"comment":"The text says 'In Fig. 7 we present the final snapshots...' but the varying-N data appear to be in Fig. 8. Please fix the cross-reference.","section":"Sec. VII.F"},{"comment":"There are unresolved references: 'See SI Fig.??' in the Fig. 8 caption and '...based on Eq.??' in SI Fig. 13. These must be completed.","section":"Fig. 8 caption and SI Fig. 13 caption"},{"comment":"Typographical issues: 'naif look' should be 'naive look'; 'hessiansh^(c)' should be 'hessians h^(c)' or similar; 'homogenenous' appears in Sec. V.B.","section":"Sec. IVB and throughout"},{"comment":"The statement that the number of metastable retrieval phases 'scales linearly with the number of components' is presented as an expectation, and the text later admits that Fig. 8 is not a detailed scaling analysis. Please label this explicitly as a conjecture, since no scaling law is derived.","section":"Sec. VII.F"},{"comment":"The sentence 'For simple two-component mixtures phase-separated states are global free energy minima' is too broad; the quartic binary example in Sec. VI itself shows binary mixtures can have metastable phase-separated states. Please qualify this statement (e.g., 'for standard quadratic interactions').","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper builds substantially on the authors' own prior PNAS work [11], but the new canonical-ensemble treatment of phase-separated metastability is a legitimate extension. The central-theorem issue is serious and load-bearing: the sufficiency proof does not handle nucleation of new phases in the bulk functional (7), and the apparent counterexamples in the paper's own binary model confirm this. In my view the problem is fixable by reformulating the theorem as local stability within a fixed phase set and adding the appropriate supporting-tangent-plane or surface-tension condition, and by rechecking the Hopfield stability regions accordingly. If the authors are unwilling to qualify the theorem, rejection would be warranted; with the correction, the paper's numerical and conceptual contributions are publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the Hopfield liquid half of this paper is real and worth your time; the general metastability theorem is stated too strongly. The canonical-ensemble setting is a genuine extension of the authors' grand-canonical work [11]: target/anti-target pairs are stationary by construction, the overlap a* is solved from the stationarity condition rather than fitted, and the Cahn-Hilliard simulations (up to N=24) match the predicted a* and show retrieval from partial cues. The binary quartic toy model with three common-tangent families is also a good illustration of why metastable phase separation is generic in nonlinear mixtures. Credit where due: the simulations and the geometric framing are solid.\n\nThe soft spot is the Sec. IVB theorem: 'metastable iff all phase Hessians are positive definite.' The necessity direction is fine. The sufficiency argument is fine for perturbations of the existing phases - the reader's worry about Eq. (14) dropping volume terms doesn't land, because delta^(C) in Eq. (15) carries the epsilon-dependence and the weighted sum of positive definite quadratic forms is positive. What the proof never addresses is nucleating a composition outside the existing phase set. For the bulk functional (7), an infinitesimal droplet of a composition lying below the common tangent plane lowers F linearly: no barrier, so the state is not a local minimum. The paper knows this for homogeneous states (Appendix B discards exactly such perturbations, citing the surface-tension nucleation barrier) but does not carry the caveat into the phase-separated theorem. In the binary quartic model, families II and III sit above the family I binodal, so a family II state is unstable in the bulk functional to nucleating high-concentration droplets; the Cahn-Hilliard simulations only survive because the k grad^2 phi term supplies the barrier the theorem assumes away. The paper even gestures at this for C>2 ('soft modes can be stabilised by surface tension') without integrating it into the theorem. Consequently, the analytical metastability regions in Fig. 5B-D are not established by the stated criterion. They are probably right for the spatial model - the simulations do show retrieval - but the paper overclaims.\n\nFix: state the criterion as local stability within the fixed phase set, and treat nucleation barriers/surface energy explicitly. Minor: unresolved 'SI Fig.??' (Fig. 8 caption) and 'Eq.??' (SI Fig. 13). Self-citation of [11] is legitimate here.\n\nThis deserves a serious referee: the idea is worth developing properly, the simulations are evidence, and the theorem needs restating, not discarding.","headline":"Real, well-simulated Hopfield-liquid retrieval, but the central 'iff' metastability theorem is overstated: for the surface-tension-free bulk functional, positive phase Hessians don't stop infinitesimal droplets of lower-lying compositions from nucleating.","tokens_in":44146,"tokens_out":10816,"would_cite":true,"duration_ms":105377,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A phase-separated mixture is metastable exactly when every one of its phases is locally stable on its own.","keywords":["metastable phase separation","multicomponent mixtures","Hopfield liquids","Hessian stability criterion","associative memory","liquid-liquid phase separation","Cahn-Hilliard dynamics","pattern retrieval"],"falsifier":"Simulate a Hopfield liquid in the predicted metastable region with a droplet whose composition matches a stored target but whose size is small enough that interfacial energy matters: if the droplet shrinks or changes composition even though the bulk Hessian is positive definite, the bulk-only criterion misses finite-size effects. Conversely, if a phase whose Hessian has a negative eigenvalue is observed to persist stably in a strong-surface-tension simulation, the necessity direction of the theorem fails.","tokens_in":43173,"feed_emoji":"🧠","tokens_out":3769,"duration_ms":50883,"temperature":0.7,"pith_summary":"The paper establishes a general condition for when a multicomponent liquid mixture can sit in a long-lived demixed state that is not the global free-energy minimum. Its central claim is that a phase-separated state is metastable precisely when each of its constituent phases has a positive-definite free-energy Hessian; if any single phase is locally unstable, the whole demixed state is unstable, because a new phase can nucleate inside it. This matters because liquids with many components, such as biological cytoplasm, can host many alternative arrangements of coexisting condensates, and metastability is what would let them act as associative memories. The paper applies the criterion to a toy binary mixture with higher-order interactions and to Hopfield liquids, where it shows analytically and in spatial simulations that stored target phases are retrieved from partial composition cues. If correct, the result gives a practical stability test for demixed states and explains how complex biological mixtures could perform pattern completion without being at equilibrium.","feed_headline":"Phase-separated liquids retrieve stored information from partial cues","feed_subtitle":"New criterion: a demixed state is metastable exactly when each of its phases is locally stable, with simulations confirming retrieval.","key_machinery":"The load-bearing object is the free-energy Hessian h(c), with elements h_ij = ∂²f/∂ϕ_i∂ϕ_j evaluated at the composition of phase c. The paper shows that a phase-separated state is stationary when exchange chemical potentials and osmotic pressures balance across phases, and that it is metastable if and only if all h(c) are positive definite. This reduces a many-variable stability problem for the whole demixed state to a local check on each phase. Around this criterion, the paper builds a geometric common-tangent construction for metastable states and, in the Hopfield liquid, a self-consistent tanh equation for the retrieval overlap a* together with explicit inequalities on the interaction str","core_discovery":"The paper proves that a phase-separated state with P phases is metastable if and only if the Hessian of the free energy density evaluated in each phase is positive definite. The only exception is a family of soft modes in which compartment volumes change without changing phase compositions; these leave the bulk free energy unchanged and require surface tension for stabilization. The necessity direction uses a perturbation that nucleates a new phase inside an unstable phase, so a single concave direction anywhere destroys metastability of the whole state. Applied to Hopfield liquids, the paper shows that a liquid whose interaction matrix is a projector onto stored target compositions can retr","pith_inferences":["Because the analytical criterion omits surface tension from the bulk stability calculation, a natural extension is that finite droplets may be stabilized or destabilized by interfacial energy at small length scales; this predicts a droplet-size-dependent boundary for metastability that could be tested in simulations by varying the gradient coefficient k.","The criterion suggests a practical experimental assay for real condensates: measure composition fluctuations within each phase, fit a free-energy model, and check whether the Hessian is positive definite; a negative direction should predict droplet splitting or compositional drift, which could be tested with existing condensate data.","The liquid-Hopfield mapping points toward an evolutionary or learning interpretation of cytoplasmic organization: if protein interaction networks are shaped by selection, condensate composition storage could be a form of learned associative memory, though the paper presents this as an analogy rather than a demonstrated biological mechanism.","A concrete engineering step follows from the paper: encode the Hebbian interaction matrix in designed DNA or synthetic sequences to build programmable multi-phase liquids whose stored phases share components, a task that is conceptually straightforward but technically demanding."],"forward_implications":["Stability analysis of any multicomponent phase-separated state reduces to checking the Hessian of each constituent phase, rather than analysing all coupled volume-composition perturbations.","Metastability is not a rarity in many-component mixtures: high-dimensional free-energy landscapes naturally have many local minima, and the criterion gives a direct way to identify them.","Hopfield liquids provide associative memory in the liquid state: a partial composition cue drives the mixture into a target/anti-target demixed pair whose final overlap matches the stored pattern.","Multiple stored phases can coexist in one mixture, and the number of phases that coexist grows with the number of components, as shown by simulations.","A repulsive cubic interaction stabilizes retrieval phases in the canonical model, but the supplementary analysis shows that a programmable surface-energy tensor can also stabilize them even without cubic nonlinearities."],"fun_headline_variants":["Metastable phase separation gives multicomponent liquids a memory","Metastable demixing retrieves stored patterns in biological liquids","Metastable phase separation acts as information retrieval in mixtures","Liquid mixtures can recall stored phases via metastable demixing","Multicomponent liquids use metastable phase separation to retrieve info"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The stability criterion treats a phase-separated state's bulk free energy as a sum of homogeneous-phase free energies with no surface-tension term, so if interfacial energy materially changes the stability of real phases or of volume-perturbation modes, the equivalence between phase stability and whole-state metastability could fail.","fun_headline_variants_meta":{"raw":{"variants":["Metastable phase separation gives multicomponent liquids a memory","Metastable demixing retrieves stored patterns in biological liquids","Metastable phase separation acts as information retrieval in mixtures","Liquid mixtures can recall stored phases via metastable demixing","Multicomponent liquids use metastable phase separation to retrieve info"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000559,"raw_usage":{"total_tokens":2489,"prompt_tokens":734,"completion_tokens":1755,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":1670}},"tokens_in":478,"tokens_out":1755,"duration_ms":14222,"temperature":1.0,"reasoning_tokens":1670,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:37:42.347661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a Hopfield liquid in the predicted metastable region with a droplet whose composition matches a stored target but whose size is small enough that interfacial energy matters: if the droplet shrinks or changes composition even though the bulk Hessian is positive definite, the bulk-only criterion misses finite-size effects. Conversely, if a phase whose Hessian has a negative eigenvalue is observed to persist stably in a strong-surface-tension simulation, the necessity direction of the theorem fails.","supporting_citations":[],"review_version":1}