{"id":"6d4bf5dd-0634-49a0-833c-9cec92d320a5","arxiv_id":"1908.02519","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Tracking control of a two-phase Stefan problem via collocated and flatness-based output feedback is developed and demonstrated in simulation.","lead":"This paper designs and simulates feedback controllers that track the solidification front in a model of the Vertical Gradient Freeze crystal growth process, using collocated and flatness-based methods with an observer for output feedback. It matters because it brings modern control-theoretic tools to a moving-boundary problem that is central to industrial semiconductor crystal growth.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop design rests on an unvalidated N=5 truncation of the power-series model; no error bound connects it to the full Stefan PDE, and the general stability proof depends on an unproven total-variation bound.","rationale":"The reader's weakest assumption names exactly the load-bearing issue: the controller and observer are designed on a low-order power-series truncation with no error bounds, and the general stability argument is conditional on a bound that is acknowledged as hard to establish. My stress-test agrees with that assessment. The paper is honest about these gaps, and the simulations are plausible, so they do not warrant rejection; they do warrant a conditional acceptance. No verdict change from the reader's CONDITIONAL is needed, hence UNCHANGED.","tokens_in":16118,"tokens_out":8112,"duration_ms":94697,"concrete_test":"Re-run the complete output-feedback simulation of Sec. 7.4 with controller and observer approximation order increased from N=5 to N=7 and N=9 (FEM plant, gains, noises, and reference unchanged); if the phase-boundary trajectory or growth rate shifts by more than about 1 mm or 0.1 mm/h, the N=5 design is not faithful to the full PDE. Complement this by computing the L2 defect of the truncated power series inserted into (7a) along the closed-loop trajectory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—output-feedback tracking for the two-phase Stefan problem—stands or falls on the fidelity of the order-N=5 flat-state model used by both the distributed feedback (Sec. 5) and the observer (Sec. 6). Equation (31) closes the truncated model by forcing the boundary condition onto coefficient c_{N+1}, but no estimate is given for the neglected tail; the closed loop is simulated on a FEM plant, not on the model used for design. Consequently the separation of observer and controller, and the stability of the actual infinite-dimensional loop, are assumed rather than established. The paper itself concedes the adjacent gaps: Appendix B needs the total variation of Δγ to grow at most linearly (B.10), stated to be 'hard to show' for general references, and Sec. 8 admits the series may not converge for smaller transition times. Without a truncation-error bound or a higher-order check, the simulation results in Figs. 5–8 could reflect the reduced model rather than the Stefan plant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers one-dimensional, two-phase Stefan problem models for the Vertical Gradient Freeze crystal growth process. It derives a flatness-based feedforward control, a collocated state feedback with Lyapunov stability analysis, a distributed flatness-based feedback on a truncated power-series state, and an observer based on the same finite-dimensional flat model. All components are validated in simulations against an independent finite-element (FEM) model of the Stefan problem, and the paper claims to introduce the first output-feedback tracking control methods for the two-phase Stefan problem with multiple inputs.","tokens_in":16366,"tokens_out":8600,"duration_ms":87625,"significance":"If fully established, the paper would fill a genuine gap in moving-boundary system control: output-feedback tracking for a two-phase Stefan problem with multiple boundary inputs. The flatness-based parameterization is used consistently for feedforward, feedback, and observer design, and the use of an independent FEM plant in Section 7 is a strong point in favor of the practical relevance of the approach. The explicit comparison between fixed and shifted error definitions, motivated by avoiding remelting of the already solidified crystal, is also valuable. However, the theoretical support for the central claim is incomplete: the main feedback and observer designs rely on an unvalidated low-order truncation of the power series, and the Lyapunov proof for the collocated controller is only completed for a constant reference profile.","major_comments":[{"comment":"The distributed feedback (Section 5) and the observer (Section 6) are designed on the truncated flat-state model of order N=5 (Table 1). Equation (31) determines c_{N+1} from the boundary condition, but no estimate is given for the neglected tail coefficients of the power series (8). Consequently, the linear error dynamics (35) are proven stable only for the truncated state χ^N, and there is no theoretical link to convergence of the full infinite-dimensional Stefan state. The independent FEM simulations in Section 7 are encouraging numerical evidence, but they do not quantify the approximation error. Please add a truncation-error bound, or at least a systematic convergence study with increasing N (e.g., N=3, 5, 7, 10) for the closed-loop trajectories.","section":"Section 5.1, Eq. (31)"},{"comment":"The Lyapunov analysis for the collocated feedback is carried out only for the constant reference profile T_r^0 ≡ T_m (Appendix B, 'Simplified variant'). For general reference profiles the derivation stops at condition (B.10), which requires the total variation Ψ_t^0(Δγ) to grow at most linearly in t; the authors state this is 'hard to show' (Section 4.3). Since the VGF benchmark uses a nonconstant reference trajectory (Figures 3 and 4), this is a load-bearing gap in the tracking claim. The statement that simulations show convergence for non-trivial reference profiles is empirical evidence, not a proof. Please either prove (B.10), provide a bound on Δγ̇ that implies it, or state the stability theorem with the constant-reference caveat.","section":"Section 4.3 and Appendix B, Eq. (B.10)"},{"comment":"The observer is designed from a linearisation (41) of the finite-dimensional error dynamics, and the complete observer-based output-feedback loop is validated only by simulation (Figure 8). No separation theorem or stability proof is given for the interconnection of the observer with either feedback law on the infinite-dimensional two-phase Stefan plant. Since output feedback is part of the stated contribution (Section 1.1), the paper should either provide such a proof or explicitly frame Sections 5-7 as a design procedure whose stability is demonstrated numerically for the reported parameter set.","section":"Section 6 and Section 7.4"},{"comment":"The paper acknowledges that the power-series parameterisation used for the feedforward and feedback designs may not converge for smaller transition times. This restricts the class of admissible reference trajectories and therefore the scope of the claimed tracking result. Please formulate the main claim with this restriction and, if possible, quantify the admissible range of the transition time ϑ in (14) for the parameters in Table 1.","section":"Section 8"}],"minor_comments":[{"comment":"The sentence 'in order to meat the technological requirements' contains a typo ('meat' should be 'meet').","section":"Section 4.1"},{"comment":"The sentence 'by using (36) and (16) with ˜v2(t) instead of ˜v2(t)' contains the same symbol twice; it should distinguish ˜v2(t) from v2(t).","section":"Section 5.2"},{"comment":"In equation (31), the summation should be evaluated at the boundary coordinate ˜Γ◦(t), with the time dependence made explicit; as written, ˜z^i without evaluation at the boundary does not reproduce the boundary condition (7b).","section":"Section 5.1, Eq. (31)"},{"comment":"The maps ¯ψ^N and ¯ψ^{N+1} appearing in (28), (30), (33), and (36) are not defined; please define them explicitly.","section":"Section 5.1"},{"comment":"The entries of the scaling matrix T^5 mix dimensionless factors and entries with units (m, m^2), but the units are not specified per entry; please state the units of each diagonal entry.","section":"Appendix D"},{"comment":"The caption states that the dashed orange and dashed green curves are nearly equal; if that is the intended message, separate panels or a close-up view would make the comparison clearer.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is more of an engineering design-and-simulation study than a rigorous mathematical control-theory contribution. The computational demonstrations are solid and the process motivation is genuine, but the missing truncation-error bounds, the conditional stability result in Appendix B, and the absence of a stability statement for the observer-based closed loop are significant for the claimed contribution. I would encourage the authors to add a convergence study with respect to N and to sharpen the claims accordingly; after that, the paper may become acceptable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper does something new: output-feedback tracking control for the one-dimensional two-phase Stefan problem with multiple boundary inputs. Second, the formal backing is narrower than the headline suggests: the Lyapunov proof in Section 4 and Appendix B is complete only for a spatially constant reference profile, and the general case depends on a bound the authors admit is hard to show. Neither gap kills the paper, but they define its ceiling.\n\nThe design chain is coherent. They build on prior flatness-based feedforward, add collocated and flatness-based feedback variants, design a flat-state observer, and then test the full loop on a finite-element discretization of the PDE, not on the same reduced model used for design. That is a genuinely useful feature. The shifted-error variants that tolerate an initial phase-boundary error instead of remelting grown crystal are a sensible engineering touch for VGF. The simulations look plausible and use realistic GaAs parameters. The authors are also candid about known limitations, including series convergence for smaller transition times and the conditional stability bound in Appendix B.\n\nThe soft spots are real but proportionate. The largest is the order-N=5 truncation used for both the distributed controller and the observer. There is no a priori error bound connecting that finite-dimensional model to the full Stefan PDE. A referee should ask for an N-sensitivity or convergence study. That said, the closed-loop plant is an independent 41-node FEM model, so the simulations are not circular; they are a consistency check, just not a proof. The stability analysis is likewise only fully proven for a constant reference; for actual VGF trajectories the argument needs the total-variation bound in (B.10), which is left unproven. The authors do not overclaim this point.\n\nThe citation pattern is fine. The self-citations to earlier flatness work are the relevant prior art, and the novelty claim is consistent with the surveyed literature.\n\nBottom line: this is an engineering-oriented methods paper, not a fully rigorous PDE control theorem. For a control journal or an applied mathematics outlet, it deserves a serious referee. I would send it out and ask for a higher-order check plus a cleaner statement of what is proven versus simulated. The gaps are fixable and the core contribution is worth engaging with.","headline":"A solid, honest extension of flatness/collocated control to the two-phase Stefan problem with observer-based output feedback; the main formal gaps are a general-reference stability proof and truncation error bounds, but the paper deserves peer review.","tokens_in":16912,"tokens_out":2340,"would_cite":true,"duration_ms":29249,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C20","35R35","80A22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper presents what it identifies as the first output-feedback tracking controllers for a one-dimensional two-phase Stefan problem, combining collocated and flatness-based state feedback with a Riccati-based observer, and shows them…","keywords":["Vertical Gradient Freeze","Stefan problem","Differential flatness","Observer design","Tracking control","Output feedback","Distributed parameter systems","Crystal growth"],"falsifier":"Run the open-loop flat parametrisation at truncation orders N=5, 10 and 20 and compare the resulting temperature profiles and interface trajectories with the FEM boundary-immobilisation solution over the full 25-hour benchmark; if the error does not decrease with N, or if the N=5 and N=10 closed-loop responses diverge, the low-order model used for control is not a faithful representation of the plant.","tokens_in":15895,"feed_emoji":"💎","tokens_out":8585,"duration_ms":87964,"temperature":0.7,"pith_summary":"The Vertical Gradient Freeze production of compound semiconductor crystals is modeled here as a one-dimensional two-phase Stefan problem: heat conduction in solid and melt, separated by a moving solid-liquid interface, with the interface velocity tied to the latent-heat balance. The paper sets out to provide tracking controllers for this moving-boundary system that work from boundary temperature measurements only, which is what an industrial furnace can actually measure. It develops two state-feedback designs, a collocated boundary controller that tracks a planned temperature profile and a flatness-based controller that works directly on a finite-dimensional model of the interface dynamics, and completes them with a Riccati-based observer that reconstructs the interface state from noisy boundary outputs. The closed loop is tested in a simulated 25-hour gallium arsenide growth run with initial errors in interface position and growth rate, and the paper claims this is the first output-feedback tracking treatment of the two-phase Stefan problem with multiple inputs.","feed_headline":"Two-phase Stefan problem tracked with boundary-only feedback","feed_subtitle":"Flatness-based controllers keep the growth rate on target in a 25-hour Vertical Gradient Freeze simulation.","key_machinery":"The machinery is the flat parametrisation of the two-phase Stefan problem. After moving to a coordinate frame attached to the interface, the temperature in each phase is expanded as a power series in the shifted spatial coordinate, and the Stefan condition fixes all series coefficients from just two quantities: the interface position $\\gamma(t)$ and the solid-side gradient at the interface $\\partial_{\\tilde z} T_s(0,t)$. That pair is the flat output. Truncating the series defines a finite-dimensional state in flat coordinates on which the distributed controller and observer are built, while the original infinite-dimensional PDE is retained as the simulated plant. The collocated controller instead relies on a shifted temperature error and a Lyapunov function over the spatial temperature error, with stability certified for a spatially constant reference.","core_discovery":"The central claim is that the two-phase Stefan problem can be tracked with output feedback using two boundary heat inputs, and that the key to the design is differential flatness: everything needed for control can be generated from the interface position and the solid-side temperature gradient at the interface, together with a finite number of their time derivatives. Truncating the associated power-series parametrisation at order N=5 yields a low-dimensional state in flat coordinates, on which the feedback law imposes decoupled linear error dynamics and on which the observer is constructed. A separate collocated design uses a spatially shifted error that compares each phase with the same phase of the reference, so that tracking the interface velocity does not force remelting of already grown crystal; for a simplified reference profile the authors prove convergence with a Lyapunov argument. The claimed novelty is that these are the first tracking designs for the two-phase Stefan problem via output feedback with multiple inputs, with performance demonstrated in simulation rather than by a full closed-loop stability proof.","pith_inferences":["If future work supplies rigorous error bounds between the N-order power-series model and the full Stefan PDE, the same flatness-based controller could be extended to axisymmetric two-dimensional VGF models where the interface is not forced to be flat; the parametrisation is spatial rather than intrinsically one-dimensional.","The 'track interface velocity, accept position offset' principle embodied in the shifted-error designs is a transferable control objective for other seeded melt-growth processes, where remelting is equally damaging.","The current observer linearizes the error dynamics along the reference trajectory; a gain-scheduled or moving-horizon observer would be a direct testable upgrade if disturbances drive the state far from that trajectory.","The stability proof covers only a spatially constant reference profile, and the paper itself flags that the total-variation bound needed for general reference profiles is hard to show; a rigorous general proof is the most pressing open step."],"forward_implications":["Furnace control can in principle be built from the two boundary temperatures plus the observer, removing the need for in-situ measurements of the phase boundary.","The flatness-based feedback turns the nonlinear moving-boundary tracking problem into two decoupled linear error equations, so engineers can tune the transient by choosing pole locations.","The shifted-error collocated variant offers a no-remelting mode: it corrects growth-rate error quickly while tolerating a stationary offset in interface position, which matches the process goal of not destroying the grown crystal.","The observer gain is computed offline from the reference trajectory, so the online output-feedback loop is cheap to implement once the reference profile is known.","The same feedforward parametrisation supplies reference profiles, reference inputs, and the model used by the observer, so all components of the control system share one consistent representation."],"supporting_citations":[{"why":"Supplies the flatness-based feedforward motion planning for the one-phase Stefan problem that this work extends to the two-phase case.","marker":"[13]"},{"why":"Extends the flatness-based trajectory planning to the two-phase Stefan problem and supports the convergence of the power-series parametrisation.","marker":"[18]"},{"why":"Provides the two-phase flatness-based feedforward design that Section 3 recaps for generating reference profiles and inputs.","marker":"[23]"},{"why":"States a Lyapunov-based feedback law for the two-phase Stefan problem with single-boundary actuation, the earlier feedback result the collocated design builds on.","marker":"[20]"},{"why":"Establishes observer-based output feedback for the one-phase Stefan problem, the closest prior output-feedback result this contribution targets.","marker":"[9]"},{"why":"Supplies the non-coercive Lyapunov framework for stability of distributed parameter systems used in the collocated stability analysis.","marker":"[26]"},{"why":"Gives the power-series coefficient state-space representation for diffusion equations that underlies the flat state used by the distributed feedback.","marker":"[27]"},{"why":"Presents the summability-based observer design for diffusion-convection-reaction systems that the flatness-based observer extends.","marker":"[28]"},{"why":"Provides the filtering Riccati differential equation result used to compute the optimal observer gain.","marker":"[29]"},{"why":"Source of the modified Poincaré inequality used to bound the Lyapunov derivative in the stability proof.","marker":"[32]"}],"fun_headline_variants":["Flatness-based tracking control for Vertical Gradient Freeze","Output feedback tames two-phase Stefan problem in crystal growth","Observer design enables boundary-only tracking of crystal growth","First tracking control with output feedback for two-phase Stefan model","Flatness simplifies VGF control: interface and gradient suffice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Correctness rests on the assumption that the fifth-order power-series truncation of the two-phase Stefan problem faithfully represents the infinite-dimensional plant over the whole simulation horizon, because neither the flatness-based feedback nor the observer carries an error bound to the full PDE, and the general-case stability argument needs a total-variation bound on the phase-boundary error that the authors state is 'hard to show'.","fun_headline_variants_meta":{"raw":{"variants":["Flatness-based tracking control for Vertical Gradient Freeze","Output feedback tames two-phase Stefan problem in crystal growth","Observer design enables boundary-only tracking of crystal growth","First tracking control with output feedback for two-phase Stefan model","Flatness simplifies VGF control: interface and gradient suffice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1217,"prompt_tokens":850,"completion_tokens":367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":466,"tokens_out":367,"duration_ms":4760,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:41:00.598345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the open-loop flat parametrisation at truncation orders N=5, 10 and 20 and compare the resulting temperature profiles and interface trajectories with the FEM boundary-immobilisation solution over the full 25-hour benchmark; if the error does not decrease with N, or if the N=5 and N=10 closed-loop responses diverge, the low-order model used for control is not a faithful representation of the plant.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the flatness-based feedforward motion planning for the one-phase Stefan problem that this work extends to the two-phase case."},{"cited_title":"Rudolph, J","cited_arxiv_id":null,"evidence_quote":"Extends the flatness-based trajectory planning to the two-phase Stefan problem and supports the convergence of the power-series parametrisation."},{"cited_title":"Rudolph, J","cited_arxiv_id":null,"evidence_quote":"Provides the two-phase flatness-based feedforward design that Section 3 recaps for generating reference profiles and inputs."},{"cited_title":"Meurer, Feedforward and feedback tracking control of diffusion-convection-reaction systems using summability meth- ods, Ph.D","cited_arxiv_id":null,"evidence_quote":"Gives the power-series coefficient state-space representation for diffusion equations that underlies the flat state used by the distributed feedback."},{"cited_title":"Meurer, M","cited_arxiv_id":null,"evidence_quote":"Presents the summability-based observer design for diffusion-convection-reaction systems that the flatness-based observer extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the filtering Riccati differential equation result used to compute the optimal observer gain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the modified Poincaré inequality used to bound the Lyapunov derivative in the stability proof."}],"review_version":1}