Pith. sign in

REVIEW 2 major objections 3 minor 41 references

Fast phase prediction of charged polymer blends by white-box machine learning surrogates

T0 review · 2 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read With only 50 training samples, a surrogate placed inside the random phase approximation reproduces charged-polymer-blend phase labels with over 99% accuracy and cuts phase-map computation by about 100 times.

desk verdict A clean, useful white-box ML paper: PPGP emulation of the RPA form factor gives >99% phase accuracy and large speedups, though the continuous-to-integer mapping of bead counts is under-specified. read the letter →

arxiv 2509.07164 v2 pith:XIXLGJGV submitted 2025-09-08 cond-mat.soft stat.AP

Clayton Ellis , Xinyi Fang , Christopher Balzer , Timothy Quah , M. Scott Shell , Glenn H. Fredrickson , Mengyang Gu This is my paper · ORCID
classification cond-mat.softstat.AP
keywords white-boxmachinelearningrandomphaseapproximationGaussianprocesssurrogatechargedpolymerblendspredictionformfactormatrixcompatibilizationfieldtheory
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a 'white-box' machine-learning surrogate—one that replaces a single expensive step inside a physics calculation, not the whole calculation—can make phase prediction for charged polymer blends nearly free. The authors show that the RPA phase classifier is bottlenecked by the computation of the form-factor matrix entries G11, G12, G22 over 200 wavevectors, and that these entries depend only on the two architecture variables (N_c,N_u). A parallel partial Gaussian process models their logarithms from just 50 training architectures; when plugged back into the RPA stability condition, the surrogate reproduces held-out phase labels (disordered, macrophase-separated, microphase-separated) at better than 99% accuracy, speeds up the form-factor inversion by roughly 50,000–70,000 times, and full phase determination by about 100 times. The broad claim is that opening the simulation's box and emulating its most expensive inner component is a better path to fast, accurate prediction than learning the final label from the full 13-parameter design space.

What carries the argument

The load-bearing object is the block-diagonal form factor matrix $G(k)$, whose diagonal blocks $\Gamma(x,k)$ encode intrachain correlations between charged and uncharged beads. Each block's entries are double sums over bead indices weighted by powers of the Gaussian linker $\Phi(k)=\exp(-k^2 b^2/6)$, and they are expensive because the sums scale as $O(m(N_c^2+N_u N_c+N_u^2))$. The PPGP surrogate models the natural logarithm of the three distinct entries of $\Gamma$ as a vector-valued function of $x=(N_c,N_u)$, using a Matérn 5/2 correlation kernel shared across the $3m$ output coordinates; its predictions are exponentiated and substituted back into the RPA equation $S^{-1}(k)=G^{-1}(k)+U(k)$, with the phase determined by whether $\det S^{-1}(k)$ vanishes at $k=0$ (MACRO) or $k>0$ (MICRO) or never (DIS). The positive-definiteness filter plus majority vote over predictive samples is what turns a raw function fit into physically coherent phase predictions.

What would settle it

Evaluate the trained 50-point surrogate on a dense fine grid concentrated in the boundary strips near $N_c+N_u=50$ and 200 (and near $N_c=1$ or $N_u=1$, where the paper itself notes $G_{12}$ is less smooth) and compare NRMSE for $G_{12}$ against the reported 0.068; if the boundary-region NRMSE is substantially larger, or if the sample-corrected phase accuracy on that region falls below 99%, the 50-point generalization claim is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that the phase output of RPA is mediated by a low-dimensional, smooth intermediate: the species-specific 2x2 intrachain correlation blocks $\Gamma(x,k)$, whose entries are non-negative and whose logarithm varies gently across the $(N_c,N_u)$ plane for each wavevector. Rather than classify phases directly, the authors fit a parallel partial Gaussian process to the logarithms of $G_{11}$, $G_{12}$, and $G_{22}$ at $m=200$ wavevectors, then re-enter the predicted values into the existing RPA linear-stability calculation. A sampling step enforces physical validity by drawing from the predictive distribution and keeping only draws that give positive-definite $\Gamma$ matrices, with a majority vote resolving the phase. On held-out architectures this pipeline exceeds 99% phase-classification accuracy with 50 training points, reduces the $\Gamma$-inversion time by four orders of magnitude, and reduces full phase-determination time by two orders of magnitude.

Load-bearing premise

The load-bearing premise is that the logarithms of the form-factor entries change smoothly and with a consistent character across the whole $(N_c,N_u)$ architecture domain, so that 50 boundary-driven training points can represent every possible chain architecture; a sharp or locally erratic feature in that surface, missed by the 50 points, would break the near-100% accuracy for unsampled architectures.

Editorial extensions

If this is right

  • Phase diagrams in the $(\alpha_A,\alpha_B)$ plane can be produced in about 3 seconds per 25x25 grid instead of 5–6 minutes by direct RPA, making high-throughput screening of charge fractions routine.
  • The 13-dimensional design space can be screened without the combinatorial explosion that a uniform grid would require: a full 10-point grid over all 13 parameters would need roughly $10^{13}$ RPA calls, while the surrogate covers the space with 50 training architectures.
  • Because the surrogate's prediction cost is independent of polymer length $N_A$ and $N_B$, the speed advantage grows for longer chains, where the direct double-sum cost is largest.
  • Training sizes as small as 20 points reach over 99% accuracy once the GP's predictive distribution is used to flag the rare uncertain inputs and run full RPA on only those inputs; by 32 training points no added simulations are needed.
  • The same white-box pattern—emulate the expensive low-dimensional inner kernel, then feed predictions into the unchanged outer calculation—is proposed as a way to accelerate other field-theoretic solvers such as self-consistent field theory.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 50-sample result depends on the even spacing of charges along the chains: the double sums in $\Gamma$ are built from Kronecker-delta charge patterns, and blocky, tapered, or otherwise clustered charge placements would change the smoothness of $\log \Gamma$ over $(N_c,N_u)$, so the maximin design and kernel would likely need re-tuning rather than reuse.
  • A natural active-learning loop is left implicit: because the GP returns a full predictive distribution, a screening campaign could request full RPA simulations only where sampled $\Gamma$ matrices disagree about the phase, potentially holding 99% accuracy at even fewer than 50 training points.
  • The same surrogate logic could transfer to other simulation methods whose bottleneck is a smooth low-dimensional subcomputation—SCFT's self-consistent iteration being the obvious next target—though the identity of the 'expensive inner kernel' would differ and the smoothness assumption would have to be revalidated.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 3 minor

Summary. The paper proposes a "white-box" machine-learning workflow for predicting the phase of charged polymer blends from random-phase-approximation (RPA) calculations. Instead of training a classifier directly on the 13-dimensional input space, the authors train a parallel partial Gaussian process (PPGP) surrogate on the two architecture variables (N_c, N_u) to predict the logarithm of the entries of the single-chain form-factor matrix G_11, G_12, G_22 at m=200 wavevectors. The predicted form factors are then inserted into the RPA machinery to obtain phase labels. With 50 training points, the surrogate achieves reported NRMSE values of 0.01-0.068 for the form-factor entries and over 99% out-of-sample phase-prediction accuracy on two test sets; the reported speedups are roughly 50,000-70,000 for the form-factor computation and about 100 for the full phase-determination procedure. The paper includes three phase-diagram case studies, an uncertainty-driven corrective sampling step, and a comparison with neural-network, random-forest, and gradient-boosting classifiers.

Significance. If the claims hold, the paper makes a useful and practical contribution: it shows that replacing only the expensive inner component of an RPA calculation can yield a fast and accurate surrogate for phase prediction, and the idea is plausibly transferable to other field-theoretic calculations. The authors deserve credit for training the surrogate on true form-factor values rather than on phase labels, for comparing against direct RPA ground truth, and for making code and data publicly available. The decomposition S^{-1}(k)=G^{-1}(k)+U(k) is exact by construction, and the surrogate evaluation does not appear circular. The main significance is as a demonstration of component-level emulation for polymer field theory, with a credible quantitative speedup claim.

major comments (2)
  1. [Methods, paragraph after Eq. (5); "Comparison with Black-box Models"] The mapping from the continuous architecture inputs r, alpha_A, alpha_B to the integer bead counts used in Equations (3)-(5) is never specified. The Methods state N_B = r N_A and N_A,c = alpha_A N_A, but Equations (3)-(5) are double sums over integer bead indices. The random 2,000-point and grid-based 10,000-point test sets sample r, alpha_A, and alpha_B continuously, so for generic test points quantities such as N_B = r N_A and N_A,c = alpha_A N_A are non-integer. The paper does not state whether rounding to the nearest integer, floor/ceiling, or a continuous extension of the discrete-chain model is used. If rounding is used, the true form factor is piecewise constant in the continuous inputs, and a smooth Matern 5/2 GP trained on the integer triangle may have uncontrolled errors near rounding thresholds; if a continuous extension is used, it must be defined because Equations (3)-(5) are not well-defined for fractional bead counts. This missing definition directly affects the validity of the reported >99% accuracy on the continuous test sets and the "full 13-dimensional design space" claim. Please specify the exact discretization rule and, ideally, report a sensitivity check of the phase-accuracy results to that rule.
  2. [Figure 3 and Figure 5] The central accuracy claim is reported only as overall accuracy, without the marginal phase frequencies in the test sets or a per-class breakdown. If the test sets are dominated by one phase, such as DIS, a trivial majority-class predictor can achieve high overall accuracy, which would make the headline "near 100%" number hard to interpret. The "at most one misclassification" statement in the case studies is likewise an aggregate count. Please add confusion matrices or per-class recall/precision (with particular attention to MICRO, which is typically the rarest and most design-relevant class), and report the accuracy of a majority-class baseline in Figure 3. This is needed to substantiate the "significantly more accurate" claim beyond the comparison with the black-box classifiers.
minor comments (3)
  1. [Figure 4, caption] The caption states that prediction cost is invariant to N_A and N_B, but the surrogate is trained on the domain corresponding to N_A=100, i.e., 50<=N_c+N_u<=200. For other N_A values, new training inputs could fall outside this domain, so the invariance statement should be qualified as applying only to the fixed-N_A setting used in this work.
  2. [Methods, "Parallel Partial Gaussian Process Surrogate"] The number of Monte Carlo samples drawn from the predictive distribution for the majority-vote correction is not stated. Please specify the number of samples (or the rule used to choose it), since the reported accuracy of the sample-corrected predictions may depend on it.
  3. [Methods, "Tasks and Data"] The boundary cases N_c=0 or N_u=0 are excluded from the surrogate domain, but the 10x10x10 grid over alpha_A, alpha_B described in the comparison section can include the endpoints 0 and 1. Please state explicitly how such test points are handled when computing the reported accuracies.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the PPGP surrogate is fitted to form-factor entries, and phase labels are always validated against direct RPA ground truth.

full rationale

The paper's derivation chain is self-contained. The RPA inverse structure factor is computed exactly as S^-1(k) = G^-1(k) + U(k), and phase labels are determined by the determinant condition |S^-1(k*)| = 0. The surrogate model is trained only on the logarithm of the form-factor matrix entries G_11, G_12, and G_22 as functions of the architecture variables (N_c, N_u), not on phase labels. Predicted form-factor entries are then plugged back into the same RPA equations, and all reported accuracies are evaluated against full RPA simulations on held-out test points. No fitted parameter is renamed as a prediction, and no phase label is used as training output. The self-citations to Gu and Berger (2016), RobustGaSP, and robust Gaussian process emulation concern established peer-reviewed statistical methodology for computer-model emulation; they do not import the phase-prediction result, do not forbid alternative approaches through a uniqueness theorem, and do not define the target in terms of the surrogate. The only notable gap is the unspecified conversion from continuous (r, alpha_A, alpha_B) to integer bead counts, which is a potential implementation or correctness risk rather than a circular reduction, because both the surrogate and the RPA ground truth share the same forward map. Thus there is no step in which the derivation reduces by construction to its own inputs.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities. Its central claim rests on standard polymer field theory (RPA and Gaussian chains) and on a smoothness and stationarity assumption about the form factor surface. The main fitted quantities are the GP hyperparameters, which are internal to the surrogate and not presented as physical constants.

free parameters (1)
  • PPGP range parameters gamma_1, gamma_2 and nugget eta = not reported in main text, estimated by profile likelihood
    These Gaussian process hyperparameters are fit to the 50 training samples and control the smoothness and noise of the surrogate. The reported accuracy depends on their fitted values.
assumptions (5)
  • domain assumption RPA linear stability analysis about a disordered state correctly identifies phase class (DIS, MACRO, MICRO) from the determinant of S^-1(k).
    All ground-truth phase labels come from this criterion. If RPA spinodals are poor approximations to real phase behavior, the surrogate inherits that error. Invoked in Methods, Phase Behavior from RPA.
  • domain assumption The Gaussian chain model with evenly distributed charges and the double-sum form factor expressions in Equations (3)-(5) are the correct single-chain statistics.
    The surrogate is trained to reproduce these form factors; any error in the physical model is inherited by the surrogate.
  • domain assumption The m=200 discrete Fourier basis grid is fine enough to locate the spinodal instability k* (including k*=0 for macro).
    Phase labels are decided by the sign of the determinant at discrete k. An instability between grid points could be mislabeled in both the direct simulation and the surrogate prediction.
  • standard math A stationary Matern 5/2 Gaussian process with constant mean and shared nugget, trained on 50 maximin points, adequately represents the log-Gamma surface over the whole (N_c, N_u) domain.
    This is the central modeling assumption. It is empirically checked by hold-out NRMSE and phase accuracy, but it is not guaranteed by the physics.
  • standard math The predictive normal distribution on log-Gamma can be sampled and used for majority-vote correction of non-positive-definite Gamma predictions.
    The uncertainty correction assumes the GP predictive distribution is accurate enough that a majority vote over valid samples is representative of the true phase.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast phase prediction of charged polymer blends by white-box machine learning surrogates." pith.science (2026). https://pith.science/paper/XIXLGJGV

@misc{pith2026250907164,
  author       = {Pith},
  title        = {Pith review of: Fast phase prediction of charged polymer blends by white-box machine learning surrogates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XIXLGJGV}},
  note         = {Machine review of arXiv:2509.07164}
}
read the original abstract

Compatibilized polymer blends are a complex, yet versatile and widespread category of material. When the components of a binary blend are immiscible, they are typically driven towards a macrophase-separated state, but with the introduction of electrostatic interactions, they can be either homogenized or shifted to microphase separation. However, both experimental and simulation approaches face significant challenges in efficiently exploring the vast design space of charge-compatibilized polymer blends, encompassing chemical interactions, architectural properties, and composition. In this work, we introduce a white-box machine learning approach integrated with polymer field theory to predict the phase behavior of these systems, which is significantly more accurate than conventional black-box machine learning approaches. The random phase approximation (RPA) calculation is used as a testbed to determine polymer phases. Instead of directly predicting the polymer phase output of RPA calculations from a large input space by a machine learning model, we build a parallel partial Gaussian process model to predict the most computationally intensive component of the RPA calculation that only involves polymer architecture parameters as inputs. This approach substantially reduces the computational cost of the RPA calculation across a vast input space with nearly 100% accuracy for out-of-sample prediction, enabling rapid screening of polymer blend charge-compatibilization designs. More broadly, the white-box machine learning strategy offers a promising approach for dramatic acceleration of polymer field-theoretic methods for mapping out polymer phase behavior.

Figures

Figures reproduced from arXiv: 2509.07164 by the authors.

Figure 1
Figure 1. Workflow of random phase approximation simulation from system parameters to [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. True, predicted and residual values of log [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Phase prediction accuracy of four models for varying training sizes on the uniformly [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Computational time in seconds between the PPGP prediction using 50 training [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Phase diagrams in the αA vs. αB plane generated using the PPGP white-box method versus full RPA simulations for varying the chain length ratio r (panels (a)–(b))and the chemical incompatibility NAχAB (panels (c)–(d)). Panel (e) compares the computation times required t…
Figure 6
Figure 6. Figure 6: Phase boundaries between compatibilized (disordered or microphase-separated) [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Panel (a) compares phase prediction accuracy on a 2 [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages

  1. [1]

    T.; Jegal, J.; Kim, H

    Jung, H.; Shin, G.; Kwak, H.; Hao, L. T.; Jegal, J.; Kim, H. J.; Jeon, H.; Park, J.; Oh, D. X. Review of polymer technologies for improving the recycling and upcycling efficiency of plastic waste. Chemosphere 2023, 320, 138089

    work page 2023

  2. [2]

    Flory, P. J. Thermodynamics of high polymer solutions. The Journal of chemical physics 1942, 10, 51--61

    work page 1942

  3. [3]

    J.; Delaney, K

    Grzetic, D. J.; Delaney, K. T.; Fredrickson, G. H. Electrostatic manipulation of phase behavior in immiscible charged polymer blends. Macromolecules 2021, 54, 2604--2616

    work page 2021

  4. [4]

    H.; Xie, S.; Edmund, J.; Le, M

    Fredrickson, G. H.; Xie, S.; Edmund, J.; Le, M. L.; Sun, D.; Grzetic, D. J.; Vigil, D. L.; Delaney, K. T.; Chabinyc, M. L.; Segalman, R. A. Ionic compatibilization of polymers. ACS polymers Au 2022, 2, 299--312

    work page 2022

  5. [5]

    M.; Yang, K.-C.; Sun, D.; Delaney, K

    Xie, S.; Karnaukh, K. M.; Yang, K.-C.; Sun, D.; Delaney, K. T.; Read de Alaniz, J.; Fredrickson, G. H.; Segalman, R. A. Compatibilization of polymer blends by ionic bonding. Macromolecules 2023, 56, 3617--3630

    work page 2023

  6. [6]

    M.; Xie, S.; Murphy, E

    Edmund, J.; Karnaukh, K. M.; Xie, S.; Murphy, E. A.; Abdilla, A.; Ino, E.; Read de Alaniz, J.; Hawker, C. J.; Segalman, R. A. Compatibilization of Immiscible Polymer Blends through Pendant Ionic Interactions. Macromolecules 2025, 58, 7425--7433

    work page 2025

  7. [7]

    Compatibilization of the polystyrene/poly (ethyl acrylate) and polystyrene/polysoprene systems through ionic interactions

    Eisenberg, A.; Smith, P.; Zhou, Z.-L. Compatibilization of the polystyrene/poly (ethyl acrylate) and polystyrene/polysoprene systems through ionic interactions. Polymer Engineering & Science 1982, 22, 1117--1122

    work page 1982

  8. [8]

    The microstructure of block copolymers formed via ionic interactions

    Russell, T.; Jerome, R.; Charlier, P.; Foucart, M. The microstructure of block copolymers formed via ionic interactions. Macromolecules 1988, 21, 1709--1717

    work page 1988

Show all 41 references
  1. [9]

    R.; Ummadisetty, S.; Nykaza, J

    Zhang, L.; Kucera, L. R.; Ummadisetty, S.; Nykaza, J. R.; Elabd, Y. A.; Storey, R. F.; Cavicchi, K. A.; Weiss, R. Supramolecular multiblock polystyrene--polyisobutylene copolymers via ionic interactions. Macromolecules 2014, 47, 4387--4396

    2014

  2. [10]

    K.; Karnaukh, K

    Beech, H. K.; Karnaukh, K. M.; Miyamoto, M. E.; Chen, K.; Edmund, J.; de Alaniz, J. R.; Hawker, C. J.; Segalman, R. A. Electrostatic Compatibilization of Amorphous and Semicrystalline Immiscible Polymer Blends. ACS Macro Letters 2025, 14, 969--975

    2025

  3. [11]

    A.; Nealey, P

    Mysona, J. A.; Nealey, P. F.; de Pablo, J. J. Machine Learning Models and Dimensionality Reduction for Prediction of Polymer Properties. Macromolecules 2024,

    2024

  4. [12]

    Machine learning-assisted design of advanced polymeric materials

    Gao, L.; Lin, J.; Wang, L.; Du, L. Machine learning-assisted design of advanced polymeric materials. Accounts of Materials Research 2024, 5, 571--584

    2024

  5. [13]

    J.; Av-Ron, S

    Arora, A.; Lin, T.-S.; Rebello, N. J.; Av-Ron, S. H.; Mochigase, H.; Olsen, B. D. Random Forest Predictor for Diblock Copolymer Phase Behavior. ACS Macro Letters 2021, 10, 1339--1345

    2021

  6. [14]

    G.; Casukhela, R

    Ethier, J. G.; Casukhela, R. K.; Latimer, J. J.; Jacobsen, M. D.; Rasin, B.; Gupta, M. K.; Baldwin, L. A.; Vaia, R. A. Predicting phase behavior of linear polymers in solution using machine learning. Macromolecules 2022, 55, 2691--2702

    2022

  7. [15]

    G.; Audus, D

    Ethier, J. G.; Audus, D. J.; Ryan, D. C.; Vaia, R. A. Integrating theory with machine learning for predicting polymer solution phase behavior. Giant 2023, 15, 100171

    2023

  8. [16]

    R.; Brettmann, B

    Ethier, J.; Antoniuk, E. R.; Brettmann, B. Predicting polymer solubility from phase diagrams to compatibility: a perspective on challenges and opportunities. Soft Matter 2024, 20, 5652--5669

    2024

  9. [17]

    A.; Kohl, P

    Fang, X.; Murphy, E. A.; Kohl, P. A.; Li, Y.; Hawker, C. J.; Bates, C. M.; Gu, M. Universal Phase Identification of Block Copolymers From Physics-Informed Machine Learning. Journal of Polymer Science 2025,

    2025

  10. [18]

    G.; Jayaraman, A

    Wessels, M. G.; Jayaraman, A. Computational reverse-engineering analysis of scattering experiments (CREASE) on amphiphilic block polymer solutions: cylindrical and fibrillar assembly. Macromolecules 2021, 54, 783--796

    2021

  11. [19]

    M.; Patil, A.; Dhinojwala, A.; Jayaraman, A

    Heil, C. M.; Patil, A.; Dhinojwala, A.; Jayaraman, A. Computational reverse-engineering analysis for scattering experiments (CREASE) with machine learning enhancement to determine structure of nanoparticle mixtures and solutions. ACS Central Science 2022, 8, 996--1007

    2022

  12. [20]

    p (q) and s (q) crease

    Heil, C. M.; Ma, Y.; Bharti, B.; Jayaraman, A. Computational reverse-engineering analysis for scattering experiments for form factor and structure factor determination (“p (q) and s (q) crease”). JACS Au 2023, 3, 889--904

    2023

  13. [21]

    Block Copolymer Theory

    Helfand, E. Block Copolymer Theory. III. Statistical Mechanics of the Microdomain Structure. Macromolecules 1975, 8, 552--556

    1975

  14. [22]

    Theory of Microphase Separation in Block Copolymers

    Leibler, L. Theory of Microphase Separation in Block Copolymers. Macromolecules 1980, 13, 1602--1617

    1980

  15. [23]

    W.; Schick, M

    Matsen, M. W.; Schick, M. Stable and Unstable Phases of a Diblock Copolymer Melt. Physical Review Letters 1994, 72, 2660

    1994

  16. [24]

    U.; Chantawansri, T

    Vorselaars, B.; Kim, J. U.; Chantawansri, T. L.; Fredrickson, G. H.; Matsen, M. W. Self-Consistent Field Theory for Diblock Copolymers Grafted to a Sphere. Soft Matter 2011, 7, 5128--5137

    2011

  17. [25]

    A.; Hill, M.; Leverick, G

    Xie, T.; France-Lanord, A.; Wang, Y.; Lopez, J.; Stolberg, M. A.; Hill, M.; Leverick, G. M.; Gomez-Bombarelli, R.; Johnson, J. A.; Shao-Horn, Y.; others Accelerating amorphous polymer electrolyte screening by learning to reduce errors in molecular dynamics simulated properties...

    2022

  18. [26]

    G.; Paluch, A

    Ethier, J. G.; Paluch, A. S.; Varshney, V. Limitations of theory-informed machine learning algorithms for the prediction and exploration of molecular properties: Solvation free energy as a case study. APL Machine Learning 2025, 3

    2025

  19. [27]

    The equilibrium theory of inhomogeneous polymers; Oxford University Press, 2006

    Fredrickson, G. The equilibrium theory of inhomogeneous polymers; Oxford University Press, 2006

    2006

  20. [28]

    Random forests

    Breiman, L. Random forests. Machine learning 2001, 45, 5--32

    2001

  21. [29]

    Classification and Regression by randomForest

    Liaw, A.; Wiener, M. Classification and Regression by randomForest. R News 2002, 2, 18--22

    2002

  22. [30]

    Deep learning

    LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436--444

    2015

  23. [31]

    Gu, M.; Berger, J. O. Parallel partial G aussian process emulation for computer models with massive output. The Annals of Applied Statistics 2016, 10, 1317--1347

    2016

  24. [32]

    Gu, M.; Palomo, J.; Berger, J. O. RobustGaSP: Robust Gaussian Stochastic Process Emulation in R . The R Journal 2019, 11, 112--136

    2019

  25. [33]

    E.; Mark, J

    Mark, J. E.; Mark, J. E. Physical properties of polymers handbook; Springer, 2007; Vol. 1076

    2007

  26. [34]

    K.; Yao, S

    Lee, J. K.; Yao, S. X.; Li, G.; Jun, M. B.; Lee, P. C. Measurement methods for solubility and diffusivity of gases and supercritical fluids in polymers and its applications. Polymer reviews 2017, 57, 695--747

    2017

  27. [35]

    Rasmussen, C. E. Gaussian processes for machine learning; MIT Press, 2006

    2006

  28. [36]

    Gu, M.; Wang, X.; Berger, J. O. Robust G aussian stochastic process emulation. Annals of Statistics 2018, 46, 3038--3066

    2018

  29. [37]

    J.; Williams, B

    Santner, T. J.; Williams, B. J.; Notz, W. I. The design and analysis of computer experiments; Springer Science & Business Media, 2003

    2003

  30. [38]

    https://keras.io, 2015

    Chollet, F.; others Keras. https://keras.io, 2015

    2015

  31. [39]

    XGBoost: A scalable tree boosting system

    Chen, T.; Guestrin, C. XGBoost: A scalable tree boosting system . Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 2016; pp 785--794

    2016

  32. [40]

    Chen, T. et al. xgboost: Extreme Gradient Boosting. 2025; R package version 1.7.7.1

    2025

  33. [41]

    U ber die von der molekularkinetischen Theorie der W \

    Fang, X.; Gu, M.; Wu, J. Reliable Emulation of Complex Functionals by Active Learning with Error Control. The Journal of Chemical Physics 2022, 157, 214109 mcitethebibliography main.tex0000664000000000000000000034616015057600130011233 0ustar rootroot [journal=jacsat,manuscript...

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.