acceptodds
Under review as a conference paper at ICLR 2027

MIRA: Interpretable Multiscale Prediction of Molecular Mixture Properties

Abstract

The property of a molecular mixture depends on its constituents, their proportions, and how they behave together. However, existing methods suffer from two key limitations: (1) interaction learning approaches typically fuse all signals into a single latent embedding before prediction, limiting interpretability and obscuring the separation between component-level and mixed-level contributions; (2) physically motivated decompositions offer clearer structure but often depend on property-specific equations or assumptions, restricting their transferability. To address these limitations, we introduce Mixture Interaction and Residual Aggregation (MIRA), a unified property-agnostic framework for molecular mixture property prediction. MIRA performs prediction-level multi-scale decomposition by assigning explicit predictive roles to composition-weighted constituents, molecular interactions, and residual whole-mixture effects. This design separates component-wise, pairwise, and formulation-level contributions while preserving a common architecture across diverse properties. Across 11 CheMixHub tasks, MIRA achieves the lowest or tied-lowest mean test MAE at the reported precision, reducing five-fold MAE by up to 61.0%. It further generalizes to external mixture-property datasets, including surface tension and electrolyte-property prediction, demonstrating the broad applicability of its architectural design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
mixture property prediction, molecular machine learning, intermolecular interactions, graph neural networks, molecular representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_odtLOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This work presents MIRA for molecular mixture property prediction. Unlike previous work that relies on fully black-box prediction, MIRA decomposes the prediction into three complementary pathways corresponding to individual components, pairwise interactions, and the mixture as a whole, which potentially improves the interpretability of its predictions. On 11 CheMixHub tasks, MIRA achieves state-of-the-art performance, and the ablation studies further validate the distinct and complementary pathway behaviors across different properties.

Critical strengths
  • The experimental section is comprehensive and thorough. Through benchmarking against baseline methods and extensive ablation studies, the results largely demonstrate the superior performance of MIRA on molecular mixture property prediction and help explain the different roles played by the three individual modules for different target properties.
Critical weaknesses
  • The main novelty of this work, namely the explicit modeling of the contributions from different components using three separate modules, still lacks sufficient justification. From a methodological perspective, supervision is applied only to the final aggregated prediction, while the outputs of the individual modules are essentially unsupervised. Under this setting, explicit modeling does not necessarily offer an advantage over implicit modeling, much like how hand-crafted feature engineering has largely been replaced by transformer-based representations. Therefore, although the model demonstrates strong empirical performance, I am not convinced that the proposed method provides substantial insight into the task itself.
  • The readability of the paper could still be improved. (i) There are several minor errors and formatting inconsistencies. For example, the last row under Density in Table 1 is missing a closing parenthesis, and Table 4 uses inconsistent numbers of decimal places when reporting the results. (ii) In Section 4.4, the experimental settings and objectives of the individual ablation studies are not described clearly, which makes this section considerably difficult to follow.
Questions for authors
  1. Since the training objective only supervises the final prediction obtained by combining the outputs of all three pathways, the individual pathway outputs are essentially unsupervised. This raises a major concern about whether explicit modeling of these pathways is necessary. Could the authors provide additional clarification or experiments to verify that the outputs of the individual components are indeed physically interpretable?
  2. Both the Pairwise Correction and Whole-mixture Correction appear to consider only interactions between pairs of components. Is it possible that interactions involving three or more components exist and cannot be captured by such pairwise modeling? If so, would this introduce a systematic limitation in MIRA's formulation? The authors are encouraged to provide the necessary clarification or correct my understanding if this concern is unfounded.
  3. In Section 4.5, MIRA's generalization ability is evaluated only on novel combinations of molecules that were individually observed during training, rather than on mixtures containing previously unseen molecules. Since MIRA is built upon a shared pretrained CheMeleon encoder, it should in principle be able to make predictions for systems containing molecules that were not observed during training. Could the authors include additional experiments on mixtures containing unseen molecules to further evaluate the generalization ability of MIRA?
  4. Could the authors provide a comparison with the baselines in terms of model size (e.g., number of parameters) and the runtime required for evaluation?
  5. Could the authors provide an anonymous code repository or include the implementation in the supplementary materials? This would help improve the reproducibility of the work.
6 comments

Weaknesses 1

The main novelty of this work, namely the explicit modeling of the contributions from different components using three separate modules, still lacks sufficient justification. From a methodological perspective, supervision is applied only to the final aggregated prediction, while the outputs of the individual modules are essentially unsupervised. Under this setting, explicit modeling does not necessarily offer an advantage over implicit modeling, much like how hand-crafted feature engineering has largely been replaced by transformer-based representations. Therefore, although the model demonstrates strong empirical performance, I am not convinced that the proposed method provides substantial insight into the task itself.

Reply 1

We thank the reviewer for raising this point. Supervising only the summed prediction leaves the individual pathway outputs underdetermined. Explicit decomposition makes each additive output directly observable and lets us test how the fitted model relies on it by replacing that output while holding the rest of the model fixed.

To assess the accuracy trade-off, we compared MIRA with a feature-fusion baseline using the same inputs and encoder architecture across five matched folds (Rebuttal Table 6). Explicit decomposition achieved lower mean test MAE on all four properties, with reductions of approximately 0.5%-3.0% relative to feature fusion. It achieved lower MAE on four of five folds for mixing enthalpy and Polymer-MW, three of five for NIST-small viscosity, and all five for IL viscosity. As described in the reply to question 1, we used the separate outputs to examine which pathways the model relies on for different properties and how their predictions relate to measurable mixture behavior.

Rebuttal Table 6: Explicit Decomposition vs. Feature Fusion (Test MAE ↓\downarrow)

PropertyExplicit Decomposition MAEFeature Fusion MAEΔ\DeltaMAE (Fusion − Decomp.)Better Across Folds
Mixing enthalpy (ΔHmix\Delta H_{\text{mix}})0.153474 ± 0.0023120.154253 ± 0.002168+0.000779Decomposition (4/5)
NIST-small viscosity0.012710 ± 0.0002210.012930 ± 0.000295+0.000220Decomposition (3/5)
Polymer-MW0.178065 ± 0.0078350.183624 ± 0.004270+0.005559Decomposition (4/5)
IL viscosity0.036069 ± 0.0007980.036699 ± 0.001023+0.000630Decomposition (5/5)

Weaknesses 2

The readability of the paper could still be improved. (i) There are several minor errors and formatting inconsistencies. For example, the last row under Density in Table 1 is missing a closing parenthesis, and Table 4 uses inconsistent numbers of decimal places when reporting the results. (ii) In Section 4.4, the experimental settings and objectives of the individual ablation studies are not described clearly, which makes this section considerably difficult to follow.

Reply 2

Thank you for catching these presentation details. In the revised manuscript, we will fix the missing closing parenthesis in Main Paper Table 1, standardize decimal precision across Main Paper Table 4 (and all other tables), and carefully proofread the entire text. Additionally, we will restructure Section 4.4 by explicitly stating the objective, experimental setup (retrained models vs. frozen-model interventions), and key takeaway at the beginning of each subsection so that the diagnostic progression is easy to follow.

Question 1

Could the authors provide additional clarification or experiments to verify that the outputs of the individual components are indeed physically interpretable?

Reply 1

We thank the reviewer for this suggestion. We analyzed how the fitted models use the three pathways and compared their outputs with physical reference data. The component pathway tracks the composition-weighted pure-liquid reference, the pairwise pathway helps predict symmetric deviations from it, and the whole-mixture pathway captures part of the composition asymmetry. Our reply to the first point of Weakness 1 explains the rationale for explicit modeling.

1. Pathway allocation in fitted models

Without separate physical labels, pathway outputs represent predictive attributions. Across independently fitted models, the dominant pathway agrees on 15/20 mixing-enthalpy systems, 16/20 NIST-small viscosity systems, and 17/20 surface-tension systems (Rebuttal Table 7). The component pathway has the largest magnitude for NIST-small viscosity, and the two correction pathways dominate for mixing enthalpy. For mixing enthalpy and NIST-small viscosity, replacing the largest-magnitude pathway output with its training-fold mean in the frozen model caused the largest test MAE increase (Rebuttal Table 8).

Rebuttal Table 7. Relative Pathway Magnitudes and Dominant-Pathway Agreement

PropertyComponentPairwiseWhole-mixtureAgreement
Mixing enthalpy12.0%30.1%57.9%15/20
NIST-small viscosity58.3%13.1%28.6%16/20
Surface tension37.6%10.6%51.8%17/20

Rebuttal Table 8. Test MAE Increase under Frozen Pathway Mean Replacement

PropertyComponentPairwiseWhole-mixture
IL conductivity+0.218+0.125+1.901
IL viscosity+0.895+0.214+0.745
NIST-small viscosity+0.201+0.026+0.076
Mixing enthalpy+0.007+0.128+0.293

2. Component pathway

For 1,593 viscosity measurements from 188 held-out binary systems, we constructed a composition-weighted reference from the two pure-liquid log viscosities measured at the same temperature. For mole fraction xx of B,

Lref(x)=(1−x)ln⁡ηApure+xln⁡ηBpure.L_{\mathrm{ref}}(x)=(1-x)\ln\eta_A^{\mathrm{pure}}+x\ln\eta_B^{\mathrm{pure}}.

The component output correlated with this reference (r=0.947r=0.947), supporting its role as a composition-linear starting point. The full model fit the measured pure-liquid viscosities more accurately than the component pathway alone (MAE 0.0273 vs. 0.1526).

3. Pairwise pathway

We illustrate the pairwise pathway using methanol (A) and dimethyl carbonate (B) at 307.15 K. The measured deviation is δ(x)=ln⁡ηmix(x)−Lref(x)\delta(x)=\ln\eta_{\mathrm{mix}}(x)-L_{\mathrm{ref}}(x). We average the deviations at complementary compositions as [δ(x)+δ(1−x)]/2[\delta(x)+\delta(1-x)]/2. For predicted deviations, the reference uses the model’s pure-liquid predictions.

Rebuttal Table 9. Symmetric Log-Viscosity Deviations with Fixed and Composition-Dependent Pairwise Outputs

A:B compositionsOriginal pairwise outputFixed at meanMeasured average deviationPrediction (fixed)Prediction (original)
90:10 and 10:90+0.00074+0.00740−0.02725−0.01788−0.02453
80:20 and 20:80+0.00525+0.00740−0.03940−0.03475−0.03690
70:30 and 30:70+0.01016+0.00740−0.04635−0.04784−0.04507
60:40 and 40:60+0.01344+0.00740−0.04855−0.05650−0.05045

With model weights and other outputs fixed, restoring the original pairwise output preserved the mean prediction and lowered absolute error at all four composition pairs (0.00586 to 0.00210 overall). Across 167 held-out NIST-small binary systems, it reduced curve-shape MAE by 30.6% after subtracting each curve’s own mean. The pairwise pathway helps MIRA capture composition-dependent symmetric deviations.

4. Whole-mixture pathway

We used molecular-dynamics mixing enthalpies for 716 held-out binary systems from CheMixHub Miscible Solvents. At each composition, we subtracted the linear interpolation between the model’s two pure-component predictions. With the fitted model fixed, including the pairwise term lowered MAE at all four compositions (20:80, 80:20, 40:60, and 60:40) and for the averages over complementary compositions (0.294 to 0.205 kJ/mol).

When the proportions are reversed, 4xAxB4x_Ax_B and (xA−xB)2(x_A-x_B)^2 stay the same. The pairwise term therefore cancels in the difference between the two adjusted predictions, leaving only the change in the whole-mixture term. For each complementary pair, we compared this predicted change with the difference between the two reference mixing enthalpies. The resulting MAE was 0.278 kJ/mol, compared with 0.386 kJ/mol for a baseline predicting zero change. This shows that the whole-mixture term captures part of the composition asymmetry.

I thank the authors for their rebuttal. While their responses have addressed some of my concerns, the following issues remain unresolved:

  1. My primary concern, as stated in Weakness 1, is that the proposed explicit decomposition of component contributions may be unnecessary, as implicit modeling could already be sufficient. The additional experiments provided in Reply 1, in fact, seem to further support my concern. Across all four properties, the proposed explicit modeling approach achieves only marginal and statistically unsubstantiated improvements in MAE over the feature fusion baseline, which could simply be attributed to random variation or insufficient training. This fundamentally challenges the central motivation and necessity of the proposed method.

  2. Perhaps due to an oversight, I could not find the authors' responses to Questions 2–5.

In summary, although the paper is clearly structured and the experiments are relatively comprehensive, I find its motivation insufficiently justified and lacking in novelty. The proposed approach offers limited new insights into the field and, in my view, does not meet the standards of ICLR. Therefore, I remain inclined to recommend rejection.

Reviewer_7kvpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper provides a new model, MIRA, for molecular mixture properties prediction. Different from previous models, this work explicitly splits the representation into three levels: monomer, 2-body, and whole-mixture, providing better interpretability. The interpretability is also verified in the Experiment section. Overall, this is an interesting research about mixture representation.

Critical strengths
  1. The separation into monomer, pairwise, and whole-mixture is useful. This enables us to conduct more research on the mechanism of mixture systems.
  2. The experimental effect is good, especially the pathway analysis in sec 4.4.
Critical weaknesses
  1. The Method section is hard to follow. After reading the paper, I basically understood the main idea of this article. But after reading the Methods section, I was even more puzzled because of the use of symbols and the disordered narration. Please clarify the symbols γ\gamma and β\beta in the text and reduce some redundant formulas, such as eq (7) and (8).
  2. A little overclaim of contribution. The pairwise representation is also at the microscale, rather than the mesoscale. So the "multiscale decomposition" is a little overclaim.
Questions for authors

I'm a bit confused about why BAMBOO-Mixer was chosen as a baseline. The core of the BAMBOO-Mixer model lies in its generative part rather than its predictive part.

3 comments

Weaknesses 1

The Method section is hard to follow. After reading the paper, I basically understood the main idea of this article. But after reading the Methods section, I was even more puzzled because of the use of symbols and the disordered narration. Please clarify the symbols γ\gamma and β\beta in the text and reduce some redundant formulas, such as eq (7) and (8).

Reply 1

We thank the reviewer for this helpful suggestion. In Feature-wise Linear Modulation (FiLM), γ(c)\gamma(c) scales each feature and β(c)\beta(c) shifts it. For a feature vector hh, the modulated vector is γ(c)⊙h+β(c)\gamma(c)\odot h+\beta(c), where ⊙\odot denotes element-wise multiplication and both vectors are learned functions of the measurement conditions cc. The component pathway uses γcomp(c),βcomp(c)∈Rd\gamma_{\mathrm{comp}}(c),\beta_{\mathrm{comp}}(c)\in\mathbb{R}^{d} (Eq. 6), and the whole-mixture pathway uses γmix(c),βmix(c)∈R2d\gamma_{\mathrm{mix}}(c),\beta_{\mathrm{mix}}(c)\in\mathbb{R}^{2d} (Eq. 12). When measurement conditions are absent, FiLM leaves the features unchanged by using unit scales and zero shifts.

In the revised manuscript, we will define these vectors where they first appear and reorganize Section 3 to introduce the dual-view molecular encoding, explain the three scalar pathways, and show how their outputs are summed. We will also merge Eqs. (7) and (8) into a single definition of the symmetric pairwise feature bijb_{ij} and add a notation table to the appendix.


Weaknesses 2

A little overclaim of contribution. The pairwise representation is also at the microscale, rather than the mesoscale. So the "multiscale decomposition" is a little overclaim.

Reply 2

We thank the reviewer for this suggestion. We agree that "multiscale" may suggest distinct physical length scales, which is not our intended meaning. We have therefore revised the terminology to "prediction-level decomposition" and clarified that the component, pairwise, and whole-mixture pathways operate on different scopes of molecular information. Each pathway produces a separate scalar output, and the three outputs are combined only at the final readout.


Questions 1

I'm a bit confused about why BAMBOO-Mixer was chosen as a baseline. The core of the BAMBOO-Mixer model lies in its generative part rather than its predictive part.

Reply 1

We sincerely thank the reviewer for this constructive suggestion. BAMBOO-Mixer comprises both a forward property predictor and an inverse generative module. We included it because its forward predictor provides strong published benchmarks on the exact same CheMixHub tasks and five-fold splits used in our evaluation (Main Paper Tables 1, 7, and 8 report the authors' CheMixHub adaptation, which omits the electrolyte-specific empirical module and molecular pretraining, while the generative module is outside our scope). Across the ten CheMixHub tasks with published BAMBOO-Mixer results, its forward predictor outperformed the best model in the original CheMixHub benchmark on seven tasks and tied on one at the reported precision, making it a highly competitive baseline.

Thank you for the detailed responses and clarifications, which have addressed most of our concerns. Accordingly, I have raised my score.

We sincerely thank you for your positive feedback and thoughtful consideration of our responses. We greatly appreciate your constructive suggestions, which have helped improve our work.

Reviewer_qEcYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes to decompose the learning of the molecular mixture property to component contribution module, pairwise interactioin module, and Whole-mixture residual module, whose predictions are then then summed at the prediction level. This paper demonstrated similar or better performance than the baselines on eleven tasks from CheMixHub.

Critical strengths

Paper is very well-written and is fairly easy to read; The experimental results are extensive. The decomposition is well-motivated

Critical weaknesses
  • The ablation study in Table 3 seems to show that the whole-only variant performs very similar to the full MIRA on four of the five tasks; Only IL cond benefits from the decomposed structure. Are these differences significant across folds? If not, then can we draw statements in knowing when it's the best to use the whole-only variant? Also if they are not significantly different, the accuracy gains over Table 1 baselines may come mainly from the attention-based whole mixture module.

  • The high-level decomposition is well motivated, but the design choice of each module is not very well motivated. It is difficult to understand e.g., within the whole mixture module, which component is critical.

Questions for authors

It is unclear to me whether the additive decomposition is identifiable. Because the formulation is unconstrained, the different pathways could potentially redistribute contributions without changing the final prediction. How do the authors justify interpreting pathway magnitudes (Fig. 2a) as property-dependent allocation then? Similarly, while a composition-linear reference has a clear physical motivation for some targets (e.g., log viscosity), I am not convinced that this assumption is equally appropriate for all 11 tasks.

5 comments

Weaknesses 1

The ablation study in Table 3 seems to show that the whole-only variant performs very similar to the full MIRA on four of the five tasks; Only IL cond benefits from the decomposed structure. Are these differences significant across folds? If not, then can we draw statements in knowing when it's the best to use the whole-only variant? Also if they are not significantly different, the accuracy gains over Table 1 baselines may come mainly from the attention-based whole mixture module.

Reply 1

We sincerely appreciate the reviewer's professional summary and positive evaluation of our work. Regarding your question about differences significant across folds, our answer is the gains are consistent and statistically significant: Full MIRA outperforms Whole-only on all 5 matched folds across all three tasks (5/5 folds; paired tt-test p<0.001p < 0.001 in Rebuttal Table 1), reducing mean test MAE by 77.82% for IL conductivity, 19.42% for NIST-small viscosity, and 3.39% for mixing enthalpy (ΔHmix\Delta H_{\text{mix}}). We recommend Full MIRA whenever maximum accuracy, pathway interpretability, or pure-component and pairwise modeling are needed (such as for IL conductivity and viscosity). Conversely, Whole-only serves as a lighter alternative for global or excess properties where a pure-component linear baseline is less informative-most notably ΔHmix\Delta H_{\text{mix}}, where the ideal reference is zero and Whole-only incurs only a 3.39% MAE increase-provided explicit pathway attribution is unnecessary. Finally, the accuracy gains over Main Paper Table 1 baselines stem primarily from the pairwise summary inside the invariant branch rather than the self-attention branch (see Rebuttal Table 2 and Reply 2).

Rebuttal Table 1: Fold-by-Fold Test MAE (Full MIRA vs. Whole-only)

FoldIL ConductivityNIST-small ViscosityΔHmix\Delta H_{\text{mix}}
Fold 00.041649 / 0.1956730.013007 / 0.0159170.150086 / 0.155380
Fold 10.047134 / 0.1996350.012462 / 0.0157370.157409 / 0.160799
Fold 20.041246 / 0.1919350.012840 / 0.0152690.153193 / 0.159200
Fold 30.042437 / 0.1922620.012663 / 0.0163340.152778 / 0.158514
Fold 40.044470 / 0.1987520.012641 / 0.0156910.155576 / 0.162096
Full MIRA (Mean ± SD)0.043387 ± 0.0024350.012723 ± 0.0002080.153808 ± 0.002801
Whole-only (Mean ± SD)0.195651 ± 0.0035630.015790 ± 0.0003860.159198 ± 0.002550
Paired Δ\DeltaMAE (Mean ± SD)+0.152264 ± 0.001978+0.003067 ± 0.000458+0.005389 ± 0.001203
Win Rate (pp-value)5/5 (p<0.001p < 0.001)5/5 (p<0.001p < 0.001)5/5 (p<0.001p < 0.001)
Mean MAE Reduction77.82%19.42%3.39%

Weaknesses 2

The high-level decomposition is well motivated, but the design choice of each module is not very well motivated. It is difficult to understand e.g., within the whole mixture module, which component is critical.

Reply 2

We thank the reviewer for this suggestion. Our rationale for the three design choices is as follows. The composition-linear reference provides a common starting point for comparing mixtures of the same components at different compositions under fixed measurement conditions. The pairwise correction reflects a simple physical boundary condition: an effect that depends on two components being present should vanish when either is absent. The whole-mixture correction gives the model the flexibility to capture departures from additivity that a symmetric pairwise term cannot express, such as a binary mixture responding differently depending on which component is in excess. Within the whole-mixture module, self-attention contextualizes each component using the full mixture, while the invariant summary captures the composition-weighted mean, the maximum, and pairwise features.

The ablations in Rebuttal Table 2 show that the internal pair summary is particularly important. For IL conductivity, removing this summary alone increased test MAE from 0.041849 to 0.163473, even with self-attention and the separate Pairwise Correction pathway still active. This increase was about nine times the increase from removing self-attention (MAE 0.055412). The separate Pairwise Correction therefore did not compensate for removing pairwise features from the whole-mixture representation. For mixing enthalpy ΔHmix\Delta H_{\text{mix}}, removing the pair summary also caused a larger MAE increase than removing self-attention, although both changes were small. For NIST-small, removing the entire invariant summary increased MAE from 0.013007 to 0.014881.

The Whole-only comparison is consistent with this result. For IL conductivity, the Table 3 Whole-only architecture, which uses self-attention without the internal pair summary, had a MAE of 0.195673. Whole-only (Full Module) brought this down to 0.045023. Adding the explicit Component and Pairwise pathways further reduced MAE to 0.041849 and exposed separate outputs for pathway-level analysis.

Rebuttal Table 2: Whole-Mixture Module Internal Ablations (Test MAE ↓\downarrow)

Model VariantComp. / PairwiseSelf-AttnPair SummaryIL Cond.NIST-smallΔHmix\Delta H_{\text{mix}}
Full MIRA✓\checkmark / ✓\checkmark✓\checkmark✓\checkmark0.0418490.0130070.150086
w/o Self-attention✓\checkmark / ✓\checkmark—✓\checkmark0.0554120.0121340.151032
Mean pooling (replacing Attn)✓\checkmark / ✓\checkmarkMean pool✓\checkmark0.0485740.0122340.150537
w/o Pair summary✓\checkmark / ✓\checkmark✓\checkmark—0.1634730.0137490.152278
w/o Invariant summary✓\checkmark / ✓\checkmark✓\checkmark—0.1971810.0148810.154138
w/o Component— / ✓\checkmark✓\checkmark✓\checkmark0.0438930.0136050.151698
Whole-only (Full Module)— / —✓\checkmark✓\checkmark0.0450230.0139170.151380
Whole-only (Main Paper Table 3)— / —✓\checkmark—0.1956730.0154120.158361

Questions 1

It is unclear to me whether the additive decomposition is identifiable. Because the formulation is unconstrained, the different pathways could potentially redistribute contributions without changing the final prediction. How do the authors justify interpreting pathway magnitudes (Fig. 2a) as property-dependent allocation then?

Reply 1

We thank the reviewer for this suggestion. While the unconstrained sum does not enforce strict mathematical identifiability, independently trained models consistently converge to the same property-dependent pathway allocation that aligns with physical intuition and functional sensitivity. Across fits (Rebuttal Table 3), all models agree on the dominant pathway for 15/20 mixing-enthalpy systems, 16/20 NIST-small viscosity systems, and 17/20 surface-tension systems—matching physical expectations where viscosity is governed primarily by pure-component properties (58.3% Component) and excess mixing enthalpy (ΔHmix\Delta H_{\text{mix}}) is governed by interaction terms (57.9% Whole-mixture + 30.1% Pairwise). Furthermore, replacing each pathway's output with its training-fold mean while keeping the model frozen (Rebuttal Table 4) confirms that the dominant-magnitude pathway also causes the largest test MAE degradation when intervened upon (+0.293 for Whole-mixture on ΔHmix\Delta H_{\text{mix}}; +0.201 for Component on NIST-small).

Rebuttal Table 3: Relative Pathway Magnitudes and Dominant-Pathway Agreement

PropertyComponentPairwiseWhole-mixtureDominant-Pathway Agreement
Mixing enthalpy (ΔHmix\Delta H_{\text{mix}})12.0%30.1%57.9%15/20
NIST-small viscosity58.3%13.1%28.6%16/20
Surface tension37.6%10.6%51.8%17/20

Rebuttal Table 4: Test MAE Increase under Frozen Pathway Mean Replacement

PropertyComponent Δ\DeltaMAEPairwise Δ\DeltaMAEWhole-mixture Δ\DeltaMAE
IL conductivity+0.218+0.125+1.901
IL viscosity+0.895+0.214+0.745
NIST-small viscosity+0.201+0.026+0.076
Mixing enthalpy (ΔHmix\Delta H_{\text{mix}})+0.007+0.128+0.293

Questions 2

Similarly, while a composition-linear reference has a clear physical motivation for some targets (e.g., log viscosity), I am not convinced that this assumption is equally appropriate for all 11 tasks.

Reply 2

We thank the reviewer for raising this point. The physical motivation for a composition-linear reference varies across properties. In MIRA, it provides a composition-weighted starting point at fixed measurement conditions, and the pairwise and whole-mixture corrections allow the final prediction to depart from it. We evaluated this design on all 11 CheMixHub tasks by retraining MIRA with the component reference removed and the other two pathways retained, using the same data split for each pair of models (Rebuttal Table 5). On this matched fold, retaining the reference lowered test MAE on eight tasks. Removing it increased MAE by 1.07% to 7.22% on those tasks, with the largest increase on MON. While on tasks where a linear reference is less physically motivated (Drug Solution, Polymer-MW and Olfactory), the correction pathways naturally compensate with negligible impact (< 0.8%), supporting its use as a unified architectural prior.

Rebuttal Table 5: Component Reference Ablation (Test MAE ↓\downarrow)

TaskFull MIRAw/o ComponentMAE Change
Density0.0029730.003060+2.94%
ΔHmix\Delta H_{\text{mix}}0.1500860.151698+1.07%
ΔHvap\Delta H_{\text{vap}}0.0388280.039833+2.59%
Drug Sol.0.0122460.012194-0.42%
Polymer-MW0.1892660.188368-0.47%
IL Cond.0.0418490.043893+4.89%
IL Visc.0.0283290.029668+4.73%
MON3.5840083.842601+7.22%
NIST-full0.0266900.027068+1.41%
NIST-small0.0130070.013605+4.60%
Olfactory0.0849400.084276-0.78%

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We thank the reviewers for their constructive feedback and positive comments on the decomposition design, extensive experiments, and pathway analysis. We also appreciate their follow-up comments acknowledging our clarifications and the additional evidence from multicomponent and unseen-molecule evaluations.

During the rebuttal, we provided case studies, significance tests comparing Full MIRA with Whole-only, and ablation studies of the internal components of the whole-mixture module. We also clarified the physical motivations of the pathway designs and supplemented the pathway analysis with frozen-model interventions and comparisons with measured properties, showing how the pathways capture distinct composition-dependent behaviors. We further added multicomponent and molecule-disjoint evaluations, reported computational costs, and provided an anonymous implementation.

We will incorporate these results and clarifications into the main text and appendices, replace “multiscale” with “prediction-level decomposition,” and clarify the method description and experimental settings in the revised manuscript. We thank the reviewers again for their suggestions and the Area Chair for their time and coordination.

Best regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.