acceptodds
Under review as a conference paper at ICLR 2027

Learning Material Decisions from Heterogeneous Evidence

Abstract

Materials decisions combine incomplete composition ranges, property tables, response curves, processing histories, and text. We propose a query-conditioned framework that retains each observation's value type, role, conditions, and provenance, learns to reconstruct whole withheld evidence views, and applies typed constraints to the completed responses. Existing corpus observations provide supervision without new manual labels for every task. The implemented interface composes trained tree-based, neural, and donor specialists for property completion, joint requirement selection, and response-to-region inference; a fully shared heterogeneous encoder remains an extension. The organizing principle is decision sufficiency: a unique material state is unnecessary when compatible states support the same action. The decision layer preserves joint constraints, nonzero composition or process ranges, and unresolved applicability, distinguishing source-supported conclusions from model-conditional recommendations without autoregressive text generation. Evaluations show that complementary evidence improves completion against strong context-only controls and that learned functional responses support non-zero operating windows. Useful coverage and requirement violations expose tradeoffs that regression error alone misses. Matched controls separate these gains from general parameter transfer and show that joint-decision training does not improve every seed. The result connects fragmented materials knowledge to evaluated decisions over ranges rather than isolated property predictions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
materials prediction, applications

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ciL4OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Authors proposed a query-conditioned framework when the available data is incomplete. The paper argues that real world materials data doesn’t sit in a typical value but within a range and under specific temperature or other variables. The method supports various characteristics appear in material science data, while preserve its associated condition under such heterogeneous representation. The method leverage masking to learn the missing material properties and able to provide constraint to support the downstream engineering application.

Critical strengths
  • S1: The author designed the learning problem by considering the nature of the material science domain problem, which often have a variable collection of observations. The proposed query-conditioned framework works well on the heterogeneous material science data.
Critical weaknesses
  • W1: The authors acknowledged the full objective in Equation (3) is an extension, which is not being tested for evaluation. However, it’s unclear whether the proposed joint optimization is beneficial for downstream tasks.
Questions for authors
  • The experiment suggests that pretraining doesn’t help much. Do authors have intuition on how to train a better representation that transfer to downstream tasks?
1 comment
AuthorsOpenReview ↗

Response to Reviewer ciL4

Thank you for recognizing the importance of retaining conditions and ranges when learning from materials evidence. Combining that evidence improves completion: on material-group splits, the joint predictor reduces MAE from 104.39 MPa for context-only boosting to 69.58 MPa, while chemistry/process-only boosting obtains 132.64 MPa. The learned hidden curves also support evaluated operating windows. These gains come from the implemented specialists; the benefit of learning one shared representation remains unresolved.

The joint-training experiment adds joint-requirement supervision to a matched neural regressor. Its cost improves from 0.33640 to 0.32512 and from 0.33642 to 0.32364 in two seeds, and worsens from 0.30535 to 0.31800 in the third. Boosting remains stronger on this cost. This demonstrates feasibility and occasional gains, but no consistent downstream advantage. The complete objective in Equation (3) has not been jointly optimized across all evidence types.

For transfer, our working hypothesis is that the auxiliary task needs to match the downstream question more closely. The current test uses a small MLP, two pretraining passes, and adaptation with only 64 labeled target-bearing groups per fold. Its pretrained, scratch, and shuffled-label variants are essentially tied, so we cannot infer that the representation has learned transferable material structure. A more informative next test would train on whole withheld response curves and condition-dependent property views, retain the observation roles and conditions in the representation, and balance the auxiliary tasks so that common scalar properties do not dominate. Conditioned-set models such as Conditional Neural Processes offer a relevant architecture and comparison for that experiment.

A test of these training hypotheses should retain the scratch and shuffled-label controls, include strong tabular baselines, and hold out downstream property families or sources as well as material groups. That design would separate better supervised fitting from a benefit due to transferable representations.

Reviewer_NjPuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a query-conditioned framework for learning material decisions from heterogeneous evidence. Each observation carries a value type, operator, units, conditions, role, and provenance. The learning objective hides an entire evidence view and reconstructs it from the remaining collection. A decision layer checks predicted responses against typed requirements, preserves joint constraints, and distinguishes source-supported from model-conditional conclusions. A response-to-region operator inverts predicted functional responses into admissible operating windows through thresholding, reserve margins, and erosion. The implemented system composes tree-based, neural, and donor specialists under a common contract rather than training a single shared encoder. Evaluations on low-alloy steel data show that complementary evidence reduces completion error against chemistry-only and context-only controls.

Critical strengths
  • The problem formulation is well-motivated. Material selection deals with composition ranges, spec minima, and temperature curves, not clean scalar targets. The authors preserve this by typing each observation with its operator and role, so a normative bound like “≥ 485 MPa” is not treated as a specimen measurement during training. This distinction is often ignored in materials ML work.

  • The applicability result (Table 3, Figure 4) is the most convincing part of the paper to me. Simple lookup misclassifies 8,227 unknown-applicability queries as PASS, and the typed pipeline brings that to zero. This alone justifies the typed evidence representation.

  • The leakage controls are more careful than typical. Target masking removes both endpoints, notes, and availability flags together. Groups are closed over normalized grades, chemical fingerprints, and spec families before splitting (Sec. 4.1, A.2). I appreciate that the authors are explicit about material splits and spec-family splits being separate axes.

Critical weaknesses
  • The abstract and Sec. 3 are built around a shared heterogeneous encoder (Eq. 1) and a joint objective (Eq. 3), but both are called “extensions” multiple times and remain unvalidated. What actually runs is LightGBM, ExtraTrees, nearest-neighbor donors, and a two-layer MLP, glued together by a common interface. The gap between the proposed method and the implemented system is large. I think the paper should either implement the proposed architecture or honestly reframe itself as an integration study.
  • The learning components that are implemented mostly do not work. Pretraining, scratch, and shuffled auxiliary labels give nearly identical errors, 164.218 / 164.261 / 164.224, all worse than Ridge at 134.6, Table 7. Joint decision loss improves two of three seeds but regresses on the third, and all neural models stay weaker than boosting. Worse, the trivial reject-all strategy scores 0.29092, lower than every trained model on the reported cost metric, which makes me question the metric design. Evidence acquisition ties its location-only control at 24 vs 25 errors, Table 6. Together with weakness 1, the paper neither implements its proposed method nor shows learning helps through alternatives.
  • The headline completion gain in Sec. 4.1 is that adding YS and elongation improves UTS prediction, 69.6 vs 132.6 MPa. But YS and UTS are strongly correlated by metallurgical first principles so this is largely expected. On the no-context subset the joint model is actually worse, 263.95 vs 254.06. Independent endpoint regressors invert predicted bounds in 564 and 522 cases out of 7,797, and no consistency mechanism is proposed.
  • The applicability experiment in Table 5 pits a 4-bit quantized, single-step, zero-shot DiffusionGemma against a fully supervised specialist. This tells us little beyond the obvious supervision gap. Also, TabM, CNP, VAEAC, EDDI, and Multi-CrabNet are all discussed in related work but none appear as baselines. The authors acknowledge this but for a venue like ICLR I find the omission hard to accept.
  • The writing is a serious obstacle. Nearly every affirmative statement is immediately followed by “not X, does not establish Y, does not imply Z.” This pattern is not limited to the discussion, it is everywhere, and after reading a full subsection I often cannot tell what the authors are actually claiming. Key experimental details live almost entirely in the appendix while the main text is filled with abstract framework language, “evaluation contract,” “integration fidelity,” “execution replay.” The paper reads like it was written to preempt every possible objection rather than to communicate a result.
Questions for authors

N.A.

1 comment
AuthorsOpenReview ↗

Response to Reviewer NjPu

Thank you for recognizing the typed representation, applicability result, and leakage controls. The implemented system combines complementary observations to complete material responses and evaluate requirements or operating ranges, while preserving the distinction between measurements, normative bounds, and unresolved conditions. Its strongest evidence comes from the supervised applicability comparison and the context-only completion controls. The shared heterogeneous encoder is an extension.

Table 5 includes a direct-classification control separately trained on the same applicability task using the source and query. Across two specialist seeds, extraction followed by exact condition execution reaches 99.46-99.82% accuracy, versus 91.20-91.56% for direct classification; false applicability decisions fall from 18-40 to zero. On unseen clauses, accuracy is 96.97-99.49% versus 87.37-90.40%. This is the relevant supervised comparison for the typed route. The DiffusionGemma result measures zero-shot readiness; it cannot establish architectural superiority under equal supervision or compute.

Context-only boosting directly addresses the YS-UTS correlation concern because it already receives the YS and elongation information. Adding chemistry/process information reduces its MAE from 104.39 to 69.58 MPa on material-group splits and from 100.11 to 75.30 MPa on specification-family splits. The joint model improves over context-only boosting on 70% of held-out groups on each axis. The gain therefore goes beyond adding a known correlated property to a chemistry-only model. Section 4.2 independently finds that full-corpus training improves both boosting and neural predictors over filtered and size-matched pools, with the target cohort excluded. These gains concern supervised learning from complementary evidence under the stated grouped splits. The limited-label parameter-transfer test addresses a separate question.

Table 7 does not show useful parameter transfer, the decision loss has no consistent advantage across seeds, and the acquisition experiment does not establish a benefit beyond knowing the measurement location. Reject-all obtains the lowest stated scalar decision cost while accepting nothing. Evaluating practical utility therefore requires accepted coverage together with false acceptance and false rejection, as the operating-window experiment does through useful and violating width. The scalar cost alone cannot establish useful decision improvement.

On the no-context specification subset, the joint model is weaker, so the aggregate improvement does not extend to every missingness pattern. The interval-containment rule in Appendix A.5 requires ordered endpoints, so a reversed prediction fails that condition. This safeguards that decision, while the underlying independent regressors still produce inconsistent predictions. An ordered-output model would need its own evaluation to establish a solution to the prediction problem.

The baselines you identify address several parts of the problem. TabM is a strong neural tabular comparator; CNP and VAEAC are relevant controls for conditioning on observed subsets. EDDI is particularly relevant to the acquisition claim, which our present result does not establish. Multi-CrabNet offers a composition/task-conditioned materials baseline and would need adaptation to consume observed companion properties. Those comparisons remain to be evaluated for the model-learning claims. The existing matched boosting, context-only, and supervised applicability controls isolate narrower benefits of complementary evidence and typed execution.

Your presentation concern is fair. The implemented comparisons and the hidden-curve-to-window result should carry the main argument, with the shared architecture presented as future work and the limitations collected in one place.

Reviewer_GH5UOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper asks how incomplete materials data can support engineering decisions. For example, partial information about a steel can be used to predict a missing strength property or a strength–temperature curve, which can then be checked against requirements to propose an operating range. The system retains the meaning and conditions of each observation, combines specialized predictors with explicit checks, and evaluates the resulting decisions. Experiments show that related-property information improves prediction and that the most accurate predictor need not produce the best operating range. A fully shared model remains a proposed extension.

Critical strengths
  • S1. Related-property information improves prediction. Combining chemistry/process inputs with related properties outperforms matched single-source controls in the main strength-prediction task (Table 1). This is a useful empirical result.
  • S2. Evaluation considers engineering decisions. The operating-window study measures both useful and requirement-violating ranges, showing why a lower prediction error need not imply a better decision (Table 4).
Critical weaknesses
  • W1. The framework’s distinct contribution remains unclear. The experiments establish benefits from additional property information, specific predictors, and applicability checks, but the insight gained from integrating them is less clear. The shared encoder and full objective are explicitly extensions, so the contribution rests on the implemented system. The prediction-to-window route is quantitatively evaluated, yet the experiments do not clearly isolate what the integrated design adds beyond its component predictors and decision operators.
  • W2. The presentation makes the overall experimental argument difficult to follow. Additional-input comparisons (Section 4.1), parameter-transfer tests (Appendix C), and operating-window evaluations (Section 4.6) address different questions. The paper distinguishes these goals and connects prediction to operating-window decisions, but the reader still has to piece together how the findings support its central contribution. A clearer synthesis should identify which conclusions concern individual components and what the experiments collectively establish about the integrated framework.
  • W3. Decision-focused alternatives are insufficiently assessed. The prediction–decision mismatch is well established (e.g., Elmachtoub and Grigas, 2022, cited in the paper). Section 4.5 adds a decision loss for paired requirements, improving two of three seeds, while both training arms remain worse than full-evidence boosting on the stated cost. This narrow experiment leaves open whether directly optimizing utility would improve the main operating-window task. That task evaluates fitted curve predictors on one handbook cohort without a matched decision-focused comparator. The results characterize useful coverage and violations, but leave the comparative value of the proposed route against training directly for window utility unresolved.
Questions for authors
  • Q1. Which implemented component provides the main contribution, and which experiment best demonstrates its benefit over a simpler approach using the same evidence? Please explain how the other experiments support that contribution.
  • Q2. Do you have a matched decision-focused comparison for the operating-window task, using the same inputs and grouped splits and training to improve useful coverage while controlling violations? If available, please report the tradeoff between useful and violating width, so the comparison reflects both utility and constraint satisfaction.
  • Q3. The disrupted-context control shows that linked properties are useful. What evidence, if any, distinguishes transferable material relationships from source/specification conventions, and how does this affect the scope of the conclusions?
  • Q4. The reproducibility statement links a code archive. Which main results can be rerun from that archive without licensed raw data, and which require additional source access or source-version information?
2 comments
AuthorsOpenReview ↗

Response to Reviewer GH5U

Thank you for the careful review. Our system carries typed, incomplete evidence through response completion to decisions over conditions or ranges. A predicted value, a source-supported bound, and an observation with unresolved applicability retain their different meanings throughout that route. The evaluations show both how complementary evidence improves completion and why the resulting operating ranges change how we judge a predictor. The system uses trained specialists; the shared encoder and full objective describe an extension.

Appendix B, Table 5 compares typed execution with a simpler task-trained classifier using the same evidence. A task-trained classifier predicts applicability from the source and query, whereas the specialist extracts the relevant fields and executes the typed conditions. Across two specialist seeds, the latter achieves 99.46-99.82% accuracy versus 91.20-91.56% for direct classification, with zero versus 18-40 false applicability decisions. On 198 unseen clauses, the corresponding accuracies are 96.97-99.49% and 87.37-90.40%. This tests the value of explicit decision structure more directly than the zero-shot language-model comparison. In the larger lookup experiment, 8,227 cases with unknown applicability become affirmative passes under lookup, compared with zero under typed execution.

Table 1 tests the completion step: joint chemistry/process and property-context boosting improves MAE from 104.39 to 69.58 MPa relative to context-only boosting on material-group splits, and from 100.11 to 75.30 MPa on specification-family splits. The hidden-curve experiment then supplies completed responses to an implemented window operator. At zero reserve, ExtraTrees provides useful/violating width fractions of 0.62244/0.03647, versus 0.59842/0.02443 for an additive donor and 0.65750/0.20916 for a constant predicted anchor. These comparisons show why maximizing coverage or minimizing pointwise error alone is insufficient. The transfer experiment separately asks whether auxiliary training helps a small neural model; it did not show a useful transfer benefit in that setting.

The window study implements and measures the coverage/violation tradeoff for wholly withheld material-response curves. A matched model trained directly for window utility has not been evaluated; the paired-requirement decision-loss experiment addresses a different task. The general prediction-decision mismatch is already established in decision-focused learning, including the work of Elmachtoub and Grigas. The comparative advantage over a directly optimized window policy remains open. That comparison should use the same component splits, inputs, and reserve settings and report both useful and violating width.

The specification-family holdout and exclusion of grade/specification identifiers reduce direct reuse of source labels. The disrupted-context control further shows that the correct association between properties matters. Neither separates physical coupling completely from conventions shared across specifications; the two split axes are separate, and the curve study comes from one handbook. The supported conclusion is transfer across the declared held-out material groups or specification families within these data, with source-level transfer requiring a separate test.

For reproducibility, Appendix D documents a replay route from stored out-of-group predictions through the query/window operator, including the 797-record primary 30-ksi query. The release description also lists saved splits, features, masks, models, and predictions. Rebuilding the licensed Granta/JAHM/ASME-derived datasets and retraining from raw sources requires additional source access and version information. A result-by-result artifact manifest is still needed to specify those dependencies and distinguish prediction replay from raw-data reconstruction.

Thank you for the clarifications. The applicability comparison is helpful, but the distinct contribution of the broader integrated framework beyond its individual components remains unclear. The requested decision-focused comparison for the operating-window task has not been evaluated, leaving the comparative value of the proposed route against directly optimizing window utility unresolved. My main concerns are therefore not sufficiently addressed, and I maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.