pith. machine review for the scientific record. sign in

arxiv: 2604.14078 · v1 · submitted 2026-04-15 · ❄️ cond-mat.mtrl-sci

Recognition: unknown

Natural Language Embeddings of Synthesis and Testing conditions Enhance Glass Dissolution Prediction

Sajid Mannan , K. Sidharth Nambudiripad , Indrajeet Mandal , Nitya Nand Gosvami , N. M. Anoop Krishnan
Authors on Pith no claims yet

Pith reviewed 2026-05-10 12:50 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords glass dissolution predictionnatural language embeddingsmachine learningnuclear waste immobilizationout-of-distribution generalizationstructural descriptorsmaterial durabilitysynthesis conditions
0
0 comments X

The pith

Natural language embeddings of synthesis and testing conditions improve machine learning predictions of glass dissolution rates and enable generalization to new compositions.

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

The paper evaluates whether adding natural language embeddings that capture details from literature on how glass is made and tested can enhance machine learning models for predicting how quickly glass dissolves. On a dataset of about 700 points from published studies, the enhanced model beats standard machine learning approaches. By also converting composition data into structural descriptors, the model can predict dissolution for glasses that include entirely new chemical elements not present in the training set. This matters for finding better glasses to store nuclear waste safely over long periods, as current models struggle with the many variables involved.

Core claim

We show that machine learning models incorporating natural language embeddings of synthesis and testing conditions outperform classical ML models in predicting glass dissolution rates. Furthermore, transforming compositional features into structural descriptors and combining them with the NLP-derived features creates a model capable of extrapolating to out-of-distribution glass compositions containing new elements absent from the training data, as validated on a dataset with 34 components compared to 28 in training.

What carries the argument

Natural language embeddings capturing synthesis and testing conditions from literature, paired with structural descriptors derived from glass composition.

If this is right

  • The NLP-ML model outperforms classical ML on the curated dataset of approximately 700 datapoints.
  • The generalizable model extrapolates to out-of-distribution glass compositions containing new elements.
  • The integrated approach offers a pathway towards high-fidelity glass dissolution prediction.
  • This accelerates the discovery of novel glass compositions with tailored durability for nuclear waste management.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar text-embedding methods could improve predictive models for other material properties where experimental conditions are described in published papers.
  • The structural descriptor approach might allow screening of hypothetical glass compositions without requiring them in the original training data.

Load-bearing premise

The descriptions of synthesis and testing conditions in the literature are sufficiently detailed, unbiased, and accurately encoded by the natural language embeddings, allowing the structural descriptors to support true extrapolation without hidden biases or data issues.

What would settle it

Measuring the model's performance on a new collection of glass samples where synthesis conditions differ substantially from those in the training literature, or on compositions with elements beyond the tested set, and finding no accuracy gain or failed generalization would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.14078 by Indrajeet Mandal, K. Sidharth Nambudiripad, Nitya Nand Gosvami, N. M. Anoop Krishnan, Sajid Mannan.

Figure 1
Figure 1. Figure 1: Measured vs. Predicted values of dissolution rate in log(dissolution)(g/m2 /day) of (a) the standard vanilla model (b) the composition-based model with NLP features (c) the descriptor-based model with NLP features using XGBoost ML models for both the training and test datasets. The inset shows a histogram of errors, where the shaded region represents the 95% confidence interval. SiO 2 56.2 B 2 O 3 17.3 Na … view at source ↗
Figure 2
Figure 2. Figure 2: Plot displaying the percentage of explained variance with increasing features for Embedding 1 (dimension - [510, 768]) [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Plot displaying the percentage of explained variance with increasing features for Embedding 2 (dimension - [485, 768]) [PITH_FULL_IMAGE:figures/full_fig_p028_3.png] view at source ↗
Figure 21
Figure 21. Figure 21: Plot displaying the percentage of explained variance with increasing features for Embedding 20 (dimension - [352, 768]) [PITH_FULL_IMAGE:figures/full_fig_p037_21.png] view at source ↗
read the original abstract

Long-term chemical durability of glass, crucial for immobilizing nuclear waste, is governed by glass properties such as composition, surface geometry, as well as external factors like thermodynamic conditions and surrounding medium. Despite decades of research, there are no models that account for these intrinsic and extrinsic factors to predict the dissolution rates of glass compositions. To address this challenge, we evaluate the role of natural language embeddings capturing the synthesis and testing conditions in enhancing the predictability of glass dissolution. Evaluating the approach on hand-curated ~700 datapoints extracted from the literature, we reveal that the machine learning (ML) model including natural language embeddings (NLP-ML) outperforms classical ML model in predicting glass dissolution rate. Furthermore, we developed a generalizable ML model by transforming the compositional features to structural descriptors of glass alongside NLP-derived features, enabling extrapolation capability to glass compositions with completely new elements absent in the training data. Evaluating this model on a completely new dataset of glass compositions 34 chemical components in contrast to the training dataset that had only 28 components, we demonstrate that the model indeed exhibits generalizability to glass compositions that are out-of-distribution. Altogether, this integrated approach offers a pathway towards high-fidelity glass dissolution prediction and accelerate the discovery of novel glass compositions with tailored durability for sustainable nuclear waste management.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript claims that incorporating natural language embeddings of synthesis and testing conditions into ML models improves prediction of glass dissolution rates compared to classical ML approaches. On a hand-curated set of ~700 literature datapoints, the NLP-augmented model is said to outperform baselines. By transforming compositional features into structural descriptors plus NLP features, the approach is further claimed to enable out-of-distribution generalization to glass compositions containing 6 entirely new elements (34 total components versus 28 in the training set), supporting extrapolation for novel durable glasses in nuclear waste applications.

Significance. If the quantitative outperformance and OOD results can be substantiated with full metrics and descriptor transparency, the work could meaningfully advance predictive tools for long-term glass durability by integrating extrinsic factors from literature text and attempting element-agnostic extrapolation. The use of real literature data and the ambition to move beyond in-distribution fitting are strengths that align with practical needs in materials design for sustainable nuclear waste management.

major comments (3)
  1. [Abstract] Abstract: the central claims of NLP-ML outperformance and successful OOD generalization to compositions with 6 new elements are asserted without any quantitative metrics, error bars, validation protocol, R² values, or data-split details, which are load-bearing for evaluating whether the headline results hold.
  2. [OOD generalization section] The OOD generalization section (and associated methods): the transformation of composition to structural descriptors is presented as enabling extrapolation beyond the original 28 elements, yet no equation, algorithm, reference, or pseudocode is supplied for deriving these descriptors (e.g., coordination numbers or bond lengths), leaving open the possibility of implicit element-specific leakage from tabulated properties or training-set models.
  3. [Results on the new 34-component dataset] Results on the new 34-component dataset: the claim that the model 'exhibits generalizability' to out-of-distribution glasses requires an ablation or control showing performance when new-element identities are masked in the descriptor pipeline; without it, the independence assumption central to the extrapolation result remains unverified.
minor comments (2)
  1. [Data description] Provide a clear table or figure caption listing all 34 vs. 28 components and the exact train/test split sizes to allow readers to assess the OOD claim directly.
  2. [Methods] Clarify the embedding model (e.g., specific pre-trained NLP model and version) and any hyperparameter tuning protocol used for the NLP-ML versus classical ML comparison.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments, which have helped us identify areas where the manuscript can be strengthened for clarity and rigor. We address each major comment below and outline the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims of NLP-ML outperformance and successful OOD generalization to compositions with 6 new elements are asserted without any quantitative metrics, error bars, validation protocol, R² values, or data-split details, which are load-bearing for evaluating whether the headline results hold.

    Authors: We agree that the abstract should include quantitative support for the central claims to allow readers to immediately assess the results. In the revised manuscript, we will expand the abstract to report the key performance metrics (including R², RMSE, and error bars where applicable), the cross-validation protocol, and the train/test split details for both the in-distribution NLP-ML comparison and the OOD evaluation on the 34-component dataset. revision: yes

  2. Referee: [OOD generalization section] The OOD generalization section (and associated methods): the transformation of composition to structural descriptors is presented as enabling extrapolation beyond the original 28 elements, yet no equation, algorithm, reference, or pseudocode is supplied for deriving these descriptors (e.g., coordination numbers or bond lengths), leaving open the possibility of implicit element-specific leakage from tabulated properties or training-set models.

    Authors: We appreciate this observation on transparency. The structural descriptors are computed from established, element-agnostic physical properties (average coordination numbers and bond lengths drawn from standard glass-science references such as the work of Greaves and others on network formers). To eliminate any ambiguity, we will add a dedicated subsection in Methods that provides the explicit equations, a step-by-step algorithm, and pseudocode for the composition-to-descriptor mapping. We will also state explicitly that no training-set-derived models or element-specific lookup tables from the original 28 elements are used, thereby confirming the absence of leakage. revision: yes

  3. Referee: [Results on the new 34-component dataset] Results on the new 34-component dataset: the claim that the model 'exhibits generalizability' to out-of-distribution glasses requires an ablation or control showing performance when new-element identities are masked in the descriptor pipeline; without it, the independence assumption central to the extrapolation result remains unverified.

    Authors: We agree that an explicit control would further strengthen the independence claim. Because the structural descriptors are constructed solely from generic coordination and bond-length statistics that do not encode element identity, the pipeline is already element-agnostic by construction. Nevertheless, we will add an ablation in the revised Results section in which the descriptors for the six new elements are replaced by the mean values computed from the training-set elements; the resulting drop in predictive performance will be reported to quantify the contribution of the new-element descriptors. This control will be described alongside the existing OOD results. revision: partial

Circularity Check

0 steps flagged

No circularity: standard ML fitting on external literature data with independent OOD test

full rationale

The paper trains supervised ML models on a hand-curated set of ~700 literature datapoints, using pre-trained NLP embeddings for synthesis/testing conditions plus compositional-to-structural descriptor transforms. Performance claims and OOD generalization are assessed via held-out evaluation on a separate dataset containing 6 new elements (34 vs 28 components). No equations, fitted parameters, or self-citations reduce the dissolution-rate target to a quantity defined by construction from the inputs; the results remain falsifiable against external benchmarks and do not rely on renaming or ansatz smuggling.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on the representativeness of the ~700 literature points, the assumption that text descriptions encode the relevant extrinsic factors, and the premise that structural descriptors permit element-agnostic extrapolation.

free parameters (2)
  • ML model hyperparameters
    Hyperparameters of the NLP-ML and structural-descriptor models are chosen or fitted to the training data.
  • Embedding model selection
    Choice of pre-trained language model used to generate the condition embeddings.
axioms (2)
  • domain assumption Literature text descriptions of synthesis and testing conditions contain sufficient predictive information when converted to embeddings.
    Invoked by the decision to augment compositional features with NLP embeddings.
  • domain assumption Structural descriptors derived from composition are element-independent and support extrapolation to unseen chemistries.
    Required for the claim of generalizability to glasses with six new components.

pith-pipeline@v0.9.0 · 5556 in / 1341 out tokens · 58089 ms · 2026-05-10T12:50:22.224713+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

10 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    For dissolution experiments, glass was dried at 60°C, hand-ground, and sieved to 50–125 μm, then ultrasonically cleaned in deionized water and acetone

    INTRODUCTION Effective immobilization of radioactive waste is a critical global concern. Glass has emerged as a highly effective material for immobilizing nuclear waste1 due to its high chemical and physical stability and ease of production. Moreover, the glass matrix can readily incorporate many elements in nuclear waste, and such a glass needs to exhibi...

    work page doi:10.1007/978-1-4613-9069-5 2016

  2. [2]

    Perret, D. et al. Thermodynamic stability of waste glasses compared to leaching behaviour. Applied Geochemistry 18, 1165–1184 (2003). 6. Iler, R. K. The Colloid Chemistry of Silica and Silicates. (Cornell University Press, Ithaca, N.Y., 1955). 7. Malow, G. & Ewing, R. C. Nuclear Waste Glasses and Volcanic Glasses: A Comparison of Their Stabilities. in Sci...

    work page doi:10.1007/978-1-4684-4040-9_40 2003

  3. [3]

    Deng, L. et al. Ion-exchange mechanisms and interfacial reaction kinetics during aqueous corrosion of sodium silicate glasses. npj Mater Degrad 5, 15 (2021). 17. Frugier, P. et al. SON68 nuclear glass dissolution kinetics: Current state of knowledge and basis of the new GRAAL model. Journal of Nuclear Materials 380, 8–21 (2008). 18. Frugier, P., Chave, T....

    2021

  4. [4]

    https://ceramics.onlinelibrary.wiley.com/doi/abs/10.1111/jace.18345

    Interpreting the optical properties of oxide glasses with machine learning and Shapely additive explanations - Zaki - 2022 - Journal of the American Ceramic Society - Wiley Online Library. https://ceramics.onlinelibrary.wiley.com/doi/abs/10.1111/jace.18345. 27. Bhattoo, R., Bishnoi, S., Zaki, M. & Krishnan, N. M. A. Understanding the compositional control...

    work page doi:10.1111/jace.18345 2022

  5. [5]

    Solution‐Controlled Dissolution of Supplementary Cementitious Material Glasses at pH 13: The Effect of Solution Composition on Glass Dissolution Rates

    Snellings, R. Solution‐Controlled Dissolution of Supplementary Cementitious Material Glasses at pH 13: The Effect of Solution Composition on Glass Dissolution Rates. J. Am. Ceram. Soc. 96, 2467–2475 (2013). 37. Sessegolo, L. et al. Alteration rate of medieval potash-lime silicate glass as a function of pH and temperature: A low pH-dependent dissolution. C...

    2013

  6. [6]

    Icenhower, J. P. et al. Experimentally determined dissolution kinetics of Na-rich borosilicate glass at far from equilibrium conditions: Implications for Transition State Theory. Geochimica et Cosmochimica Acta 72, 2767–2788 (2008). 46. Abraitis, P. K. et al. The kinetics and mechanisms of simulated British Magnox waste glass dissolution as a function of ...

    2008

  7. [7]

    Dove, P. M. & Crerar, D. A. Kinetics of quartz dissolution in electrolyte solutions using a hydrothermal mixed flow reactor. Geochimica et Cosmochimica Acta 54, 955–969 (1990). 55. Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Preprint at https://doi.org/10.48550/arXiv.1...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1810.04805 1990

  8. [8]

    Bishnoi, S., Badge, S., Jayadeva & Krishnan, N. M. A. Predicting oxide glass properties with low complexity neural network and physical and chemical descriptors. Journal of Non-Crystalline Solids 616, 122488 (2023). 64. Fournier, M., Frugier, P. & Gin, S. Resumption of Alteration at High Temperature and pH: Rates Measurements and Comparison with Initial R...

    work page Pith review doi:10.48550/arxiv.1907.10902 2023

  9. [9]

    Hunter, J. D. Matplotlib: A 2D Graphics Environment. Computing in Science & Engineering 9, 90–95 (2007). 74. Shrikumar, A., Greenside, P. & Kundaje, A. Learning Important Features Through Propagating Activation Differences. in Proceedings of the 34th International Conference on Machine Learning 3145–3153 (PMLR, 2017). 75. Chen, T. & Guestrin, C. XGBoost: ...

    work page doi:10.1145/2939672.2939785 2007

  10. [10]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    McInnes, L., Healy, J. & Melville, J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. Preprint at https://doi.org/10.48550/arXiv.1802.03426 (2020). 6. Lundberg, S. M. & Lee, S.-I. A Unified Approach to Interpreting Model Predictions. in Advances in Neural Information Processing Systems vol. 30 (Curran Associates, Inc., 2017). ...

    work page internal anchor Pith review doi:10.48550/arxiv.1802.03426 2020