REVIEW 4 major objections 6 minor 30 references
Lee-Yang Theory Guided Force Field Refinement Based on Phase Diagrams
T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Minimizing the partition function modulus on a target Lee-Yang circle moves a force field's predicted phase boundary to the target.
desk verdict A genuinely new differentiable loss that fits phase diagrams by zeroing the partition function modulus on the Lee-Yang circle; the empirical checks are solid, but the imported circle theorem needs sharper quantitative backing before the universality claim is trusted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the normalized complex-temperature partition function $\bar{Z}(\tilde{\beta}, p)$, built from the reweighted enthalpy histogram $G_p(H)$ obtained from enhanced-sampling molecular dynamics combined with a multistate reweighting estimator. A Lee-Yang circle is the geometric locus of the transition-associated zeros: a circle in the complex temperature plane with diameter $(0, T_{\mathrm{pt}})$, so that one target transition temperature fixes one set of evaluation points. Evaluating $|\bar{Z}|$ at points on the upper arc of that circle yields a differentiable loss whose minimum corresponds to the predicted zeros coinciding with the target circle. The normalization step removes the artifact coming from the total number of microstates so that the loss responds to the zero locations, not to changes in the overall density of states.
What would settle it
Compute the actual Lee-Yang zeros of the unrefined Cu EAM potential at pressures like 2.97 GPa and compare their locus with the predicted circle of diameter (0,Tpt); if the edge zeros deviate systematically from that circle beyond sampling error, the loss minimum will be biased away from the target melting temperature.
Extended reading notes
Core claim
The central claim is that minimizing the normalized partition function modulus $|\bar{Z}|$ at points on the target Lee-Yang circle drives the predicted Lee-Yang zeros onto that circle, thereby moving the predicted transition temperature to the target value. The paper derives a polynomial representation of the NpT partition function whose coefficients are the discretized enthalpy probability distribution from simulation, then evaluates this polynomial at points $\tilde{T}_j^r$ on the circle with diameter $(0, T_r)$ for each target pressure $p_r$. The loss aggregates the modulus over all target conditions and evaluation points; no explicit zero-finding is needed. The numerical experiments show that this single loss improves gas-liquid coexistence, solid-liquid melting, and Cu melting simultaneously, and that the refined Cu potential also matches enthalpy and heat capacity data that were not part of the loss.
Load-bearing premise
The method assumes that every discontinuous transition with a double-peaked enthalpy histogram has its transition-related partition-function zeros on a circle whose diameter runs from zero to the transition temperature; if real potentials deviate from that circle, the loss minimum will no longer sit at the target transition.
Editorial extensions
If this is right
- A single set of refined parameters can improve several phase boundaries at once: the Lennard-Jones refinement simultaneously reduces the gas-liquid and solid-liquid mean absolute errors in transition temperature.
- Because the loss uses only the p-T phase diagram as input, the same formulation applies to any discontinuous transition in which a target boundary is known, without system-specific order parameters or response functions.
- For the copper EAM potential, refinement against experimental melting data brings the predicted enthalpy and isobaric heat capacity closer to experiment in both solid and liquid phases, even though those observables are not in the loss.
- Avoiding explicit polynomial root-finding lets the optimization proceed smoothly under automatic differentiation and automatically weights the evaluation points where physical zeros concentrate.
- The framework scales naturally to more complex potentials, with extension to machine-learning force fields identified in the paper as the next target.
Reading between the lines
- If the circle theorem holds only approximately for real materials, the loss minimum may be systematically biased for transitions with strong non-circulatory zero patterns; a direct test would be to compute the zero set of a known potential and compare its edge locus to the predicted circle.
- The same loss could be applied to multi-component or structural transitions where no good order parameter exists, provided the enthalpy histogram has a double-peak structure near coexistence.
- Because the loss is built from a reweighted histogram, its gradient noise grows as parameters drift from the sampling reference; an adaptive resampling schedule based on effective sample size would likely make the framework more robust for larger parameter spaces.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes a force-field refinement framework based on Lee-Yang theory. In the NpT ensemble, the analytically continued partition function is written as a polynomial in y = exp(-(β̃−β)ΔH), with coefficients taken from MBAR-reweighted OPES enthalpy histograms. The loss function sums |Z̄_θ(β̃_j^r, p_r)|^m over target thermodynamic conditions r and P evaluation points on the target Lee-Yang circle, i.e., the circle in the complex temperature plane with diameter (0, T_r). The authors argue that minimizing this loss drives the predicted Lee-Yang zeros toward the target circle and therefore matches the target phase diagram, without requiring order parameters or response functions. They validate the framework by refining a Lennard-Jones potential against both gas-liquid and solid-liquid reference data, and a Cu EAM potential against simulated and experimental melting curves. The refined force fields substantially reduce the mean absolute error in the target phase boundaries, and the Cu refinement against experiment also improves enthalpy and heat capacity predictions at 1 bar, which are observables outside the loss.
Significance. If the central geometric premise is correct, the method is a genuinely useful contribution: phase-diagram-guided refinement is reduced to minimizing a zeroth-order partition-function modulus, avoiding order-parameter choices, response-function targets, and the numerical ill-conditioning of explicit zero-finding. The construction is coherent and end-to-end differentiable via DMFF/JAX, and the simultaneous refinement of gas-liquid and solid-liquid coexistence with a single set of LJ parameters is a nice demonstration. The independent Cu enthalpy and heat-capacity checks provide evidence that optimizing the phase boundary does not merely overfit the phase boundary. However, the load-bearing assumption that the transition-associated zeros lie exactly on the ideal circle for realistic enthalpy distributions is imported from Refs. [18,21] and not quantitatively verified for the systems studied here, and the reported MAE improvements are given without statistical uncertainties. The empirical results are promising, but the theoretical grounding and the quantitative validation need strengthening before the generality claim is fully supported.
major comments (4)
- [§2, after Eq. (3) and Eq. (6)] The inference that minimizing L(θ) in Eq. (7) drives the predicted Lee-Yang zeros onto the target circle relies on the NpT Lee-Yang circle theorem being quantitatively accurate for the enthalpy distributions produced by LJ and Cu EAM. For the polynomial in Eq. (3), zeros lie on the unit circle |y|=1 only for special coefficient patterns. For a two-peak enthalpy distribution with unequal weights a and b separated by ΔH, the roots satisfy |y|=(b/a)^{1/ΔH}, which is not 1, and unequal peak widths also move the zeros off the circle. Thus the loss minimum does not automatically coincide with the target transition temperature for a general realistic distribution. Figure 1(b) is presented as a confirmation for Cu, but it is qualitative, and no quantitative comparison of the computed zero locations to the ideal circle is reported. I request either a direct numerical test that computes the actual zeros from the target enthalpy distributions and measures their distance from the ideal circle, or a synthetic test with asymmetric double-peak distributions showing that the loss minimum still tracks T_c for realistic asymmetry.
- [§5, quantitative results after Figs. 3 and 4] The central quantitative claims are reported without statistical uncertainties: for LJ gas-liquid, the MAE decreases from 0.0148 to 0.0007 a.u. (95.1% improvement); for LJ solid-liquid, from 0.0246 to 0.0038 a.u. (84.4%); for Cu simulated target, from 36.7 K to 1.4 K (96.1%); for Cu experimental target, from 25.4 K to 2.4 K (90.6%). Since the enthalpy histograms come from stochastic OPES/MBAR sampling, these numbers must be accompanied by error bars, confidence intervals, or at least a statement of the number of independent runs. Without this, the claim that the improvements are significant is not statistically grounded.
- [§5, simulated-reference benchmarks] The simulated-reference tests in Figs. 3 and 4(a) recover a phase diagram that is itself generated by a well-converged force field. Because L(θ) is built directly from the target transition temperatures, the reduction in MAE after refinement is, to a large extent, a self-consistency check rather than an independent validation of the physical content of the loss. The paper should state this explicitly. The genuinely independent evidence is the Cu experimental-target refinement and the out-of-target enthalpy and heat-capacity predictions in Fig. 5, and the discussion should make clear that those carry the weight of the transferability claim.
- [Eq. (7) and the limitations paragraph] The loss function has several tunable ingredients: the exponent m=2, chosen after preliminary numerical testing; the number of evaluation points P; their angular range; the number of enthalpy bins N; and the target enthalpy interval [Hmin, Hmax]. The statement that the circle geometry introduces no extraneous parameters is therefore only partially accurate. The limitations paragraph acknowledges these choices, but the accompanying assertion that the results are robust to reasonable variations is not supported by any displayed sensitivity analysis. A table reporting the dependence of the final MAE on m, P, N, and the enthalpy interval would substantially strengthen the universality claim.
minor comments (6)
- [Eq. (7)] The indexing is inconsistent: Eq. (6) defines j = 0, ..., P−1, while Eq. (7) sums j = 1 to P. Please align the two ranges.
- [Fig. 1] The caption of Fig. 1(b) describes a realistic case for Cu but does not say how the zeros were computed or from which simulation. Please provide the computational details in the main text or Supplemental Material.
- [Sampling details] The parameter γ in the well-tempered OPES target p_tg(s) ∝ [P(s)]^{1/γ} is not defined in the main text; please define it in the paragraph introducing Eq. (8).
- [Notation] The reduced Lennard-Jones units are denoted 'a.u.' in the text without definition. Please define these reduced units in Section 4.
- [Eq. (4)] The normalization denominator is written as PN−1 k=0 Gp(Hk), which is clear in context, but the notation in the displayed equation is cramped. Please rewrite the summation limits and indices cleanly.
- [Section 5, Cu validation] The ambient-pressure melting point in Fig. 5(a) is obtained by extrapolating the experimental optimization targets to 1 bar via the Simon equation. Please state the extrapolation uncertainty, since this point is used to judge the refined force field's accuracy.
Circularity Check
The Lee-Yang circle geometry that fixes the loss is imported from overlapping-author prior work, making the phase-diagram match partly built into the surrogate; out-of-sample enthalpy and heat capacity validation remains independent.
-
uniqueness imported from authors
[Theoretical foundation, paragraph after Eq. (3); used to define Eq. (6) and Eq. (7).]
"The Lee-Yang circle theorem provides the geometric foundation for inverting this chain. Originally proved for the grand canonical ensemble [14, 15], it has been extended to the N V T temperature [19, 20] and the N p T temperature plane [18]: for a system possessing an equilibrium enthalpy distribution like that depicted in the ideal case of Fig. 1(a) at pressure p with a phase transition at T pt, the partition function zeros most directly associated with the phase transition lie on a circle in T-tilde-space with diameter (0, Tpt) [21] ..."
The loss evaluation points in Eq. (6) are placed on this circle, whose diameter is set by the input target temperature T_r. Equation (7) then minimizes |Zbar| at those points, so small |Zbar| is, by the imported circle theorem, equivalent to the predicted Lee-Yang zeros lying on the target circle. The reported convergence of the Lee-Yang edge to T_r and the phase-diagram MAE reductions are therefore consequences of the loss construction rather than independent predictions. The theorem is cited to Refs. [18] and [21], whose author lists include the present authors Q.-J. Ye and X.-Z. Li; Ref. [21] is an arXiv preprint, and Fig. 1(b) offers only a qualitative confirmation for Cu.
full rationale
The paper does not commit the crudest forms of circularity: the target phase diagram is input data, the loss is an explicit function of that input, and the optimization is a standard inverse problem. The phase-diagram MAE improvements after refinement are fitting outcomes, but the paper does not mislabel them as predictions; its genuine out-of-sample check is the Cu enthalpy and heat-capacity comparison, which is independent of the fitted melting points. The one load-bearing circular element is the geometric foundation: the Lee-Yang circle with diameter (0, T_pt) is imported from Refs. [18,21], which share authors with this paper, and Ref. [21] is an arXiv preprint. The evaluation points of the loss (Eq. 6) and the claim that minimizing L matches the phase diagram (Eq. 7 and the following paragraph) reduce, by construction, to forcing zeros onto that self-cited circle. Fig. 1(b) is only a qualitative check for Cu, not a quantitative verification of the theorem. Because the central loss design is thus contingent on a load-bearing self-citation, while the out-of-sample enthalpy/C_p validation provides independent content, the circularity score is 4.
Assumptions & free parameters
free parameters (4)
- Loss exponent m =
2
- Number of enthalpy bins N
- Lee-Yang circle evaluation points (count P and angular range)
- Target enthalpy interval [Hmin, Hmax]
assumptions (4)
- domain assumption Lee-Yang circle theorem in the NpT complex-temperature plane: transition-associated zeros lie on a circle with diameter (0,Tpt).
- domain assumption A finite discrete enthalpy histogram with N bins faithfully represents the Lee-Yang zeros of the continuum partition function.
- domain assumption OPES/MBAR reweighting recovers the unbiased NpT enthalpy distribution and remains accurate as force field parameters drift between resampling steps.
- domain assumption Finite-size systems have enough Lee-Yang zero density near the real axis for the modulus signal to drive optimization.
Cite this review
Pith. "Pith review of Lee-Yang Theory Guided Force Field Refinement Based on Phase Diagrams." pith.science (2026). https://pith.science/paper/TYERAJOH
@misc{pith2026260808546,
author = {Pith},
title = {Pith review of: Lee-Yang Theory Guided Force Field Refinement Based on Phase Diagrams},
year = {2026},
howpublished = {\url{https://pith.science/paper/TYERAJOH}},
note = {Machine review of arXiv:2608.08546}
}
read the original abstract
We propose a general framework for automatic force field refinement guided by phase diagrams, grounded in Lee-Yang phase transition theory. The central idea is to directly use the partition function modulus as a phase-diagram-guided optimization target. Evaluating the modulus at points close to the real axis, where the Lee-Yang zeros are mostly associated with the phase transition, is more physically meaningful and avoids the numerical difficulty of explicitly solving for the zeros. This approach requires no system-specific order parameters or response properties for characterizing phase transition points, making it universal across various discontinuous phase transitions and material systems. We validate the method on refining parameters of a Lennard-Jones potential covering both gas-liquid and solid-liquid transitions, and a Cu embedded-atom method potential based on experimental melting curves. The refined force fields reproduce the target phase diagrams with significant improvement across all systems. For Cu, the refinement simultaneously improves predictions of enthalpy and heat capacity, which are observables beyond the optimization target. These results establish Lee-Yang theory as a practical tool for contemporary force field development.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
P. W. Anderson,Basic Notions of Condensed Matter Physics(Addison-Wesley, 1984)
work page 1984
-
[2]
P. M. Chaikin and T. C. Lubensky,Principles of Con- densed Matter Physics(Cambridge University Press, Cambridge, 1995)
1995
-
[3]
H. E. Stanley,Introduction to Phase Transitions and Critical Phenomena, International Series of Monographs on Physics (Oxford University Press, Oxford, 1987)
work page 1987
-
[4]
D. Frenkel and B. Smit,Understanding Molecular Sim- ulation: From Algorithms to Applications, 2nd ed. (Aca- demic Press, San Diego, 2002)
work page 2002
-
[5]
M. E. Tuckerman,Statistical Mechanics: Theory and Molecular Simulation(Oxford University Press, Oxford, 2010)
work page 2010
- [6]
- [7]
- [8]
Show all 30 references
-
[9]
B. Jin, B. Han, W. Feng, K. Yu, and S.-Z. Xu, J. Chem. Theory Comput.22, 3872 (2026)
2026
-
[10]
J. E. Jones, Proc. R. Soc. Lond. A106, 441 (1924)
1924
-
[11]
J. E. Jones, Proc. R. Soc. Lond. A106, 463 (1924)
1924
-
[12]
J. E. Jones, Proc. R. Soc. A43, 461 (1931)
1931
-
[13]
J. J. Potoff and J. I. Siepmann, AIChE J.47, 1676 (2001)
2001
-
[14]
C. N. Yang and T. D. Lee, Phys. Rev.87, 404 (1952)
1952
-
[15]
T. D. Lee and C. N. Yang, Phys. Rev.87, 410 (1952)
1952
-
[16]
M. E. Fisher, The nature of critical points, inLectures in Theoretical Physics, Vol. 7C (University of Colorado Press, Boulder, CO, 1965)
1965
-
[17]
Ouyang, Q.-J
X.-Y. Ouyang, Q.-J. Ye, and X.-Z. Li, Phys. Rev. E109, 024118 (2024)
2024
-
[18]
L. Liu, Y. Dong, Q.-J. Ye, and X.-Z. Li, Phys. Rev. B 112, 104102 (2025)
2025
-
[19]
J. C. S. Rocha, S. Schnabel, D. P. Landau, and M. Bach- mann, Phys. Rev. E90, 022601 (2014)
2014
-
[20]
Costa, L
B. Costa, L. M´ ol, and J. Rocha, Comput. Phys. Commun. 216, 77 (2017)
2017
-
[21]
Y.-H. Dong, L. Liu, F.-C. Wang, Q.-J. Ye, and X.-Z. Li, arXiv preprint (2026), arXiv:2606.08004
2026 arXiv
-
[22]
Invernizzi and M
M. Invernizzi and M. Parrinello, J. Phys. Chem. Lett.11, 2731 (2020)
2020
-
[23]
M. R. Shirts and J. D. Chodera, J. Chem. Phys.129, 124105 (2008)
2008
-
[24]
Bradbury, R
J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. Van- derPlas, S. Wanderman-Milne, and Q. Zhang, JAX: com- posable transformations of Python+NumPy programs (2024),https://github.com/jax-ml/jax
2024
-
[25]
M. S. Daw and M. I. Baskes, Phys. Rev. B29, 6443 (1984)
1984
-
[26]
Mishin, M
Y. Mishin, M. J. Mehl, D. A. Papaconstantopoulos, A. F. Voter, and J. D. Kress, Phys. Rev. B63, 224106 (2001)
2001
-
[27]
Errandonea, J
D. Errandonea, J. Appl. Phys.108, 033517 (2010)
2010
-
[28]
N. V. Kozyrev, Int. J. Thermophys.44, 31 (2023)
2023
-
[29]
Simon and G
F. Simon and G. Glatzel, Z. Anorg. Allg. Chem.178, 309 (1929)
1929
-
[30]
Salter, Philos
L. Salter, Philos. Mag.45, 369 (1954)
1954
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.