acceptodds
Under review as a conference paper at ICLR 2027

Condition-Informed Prediction and Resource-Aware Selection for Inorganic Materials Synthesis Planning

Abstract

Materials synthesis planning (MSP) connects AI-driven materials discovery to experimental realization by proposing precursors, synthesis operations, and processing conditions for a target material. Existing approaches predict these components separately or in pairs and treat conditions only as a prediction target. Consequently, MSP remains centered on recipe prediction: it leaves unexplored how conditions can serve as process context for constructing synthesis routes, and it does not address which candidate routes should be executed under limited experimental resources. We propose CondMSP, which extends MSP from recipe prediction to execution-aware planning. CondMSP predicts temperature, time, and atmosphere by combining a joint predictor with variable-specific experts, and then uses these predictions both as process context, grounding the generation and retrieval of operation sequences, and as execution requirements, specifying which thermal stages can share a furnace run and at what cost. This specification is held fixed across all compared methods; within it, CondMSP ranks routes across precursor sets by a common utility and, under furnace budgets, jointly selects routes and allocates their thermal stages by constrained optimization. CondMSP improves condition and operation prediction and, under this shared constraint, raises reference-recipe recovery from 27.9% to 32.9% at three simulated furnace runs per target and from 34.7% to 40.0% at 30 trials, while a new evaluation set of 2,344 targets supports generalization beyond oxide-dominated training chemistry.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Materials Synthesis Planning, Inorganic Materials Synthesis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CdkWOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the MSP problem. It points out that existing methods determin different components in isolation, and limited experimental trial budgets and equipment availability require choosing which of the many possible routes to test. Correspondingly, this paper proposed CondMSP, which extends MSP from recipe prediction to execution-aware planning.

Critical strengths
  1. The studied problem is very important and interesting.
  2. The motivation of this paper is clear.
Critical weaknesses
  1. The background is not clear. The MSP problem is not a common machine learning task, the definition of the background in Introduction is not clear. Maybe the authors can provide a concrete example to describe the studied problem. This makes readers hard to understand the following text of this paper.
  2. No detailed related work discussion provided.
  3. The problem setting is hard to understand, especially in preliminaries. Are there train set and test set? What should a model learn on train set, and what it should do on test set? In other words, what is the complete data flow for a model? And how to evaluate them?
  4. The used LLM baselines are only light version of GPT, which can not reflect the strong planning ability of modern LLMs. What is the performance of using GPT 6, rather than using the mini version? And what about Claude, and other LLMs?
Questions for authors

Please see weaknesses above.

6 comments
AuthorsOpenReview ↗

Author Response (1/3)

We sincerely thank the reviewer for the helpful feedback on the clarity of our paper. We address each point below with clarifications and additional experiments.

[W1] Background and a concrete example of MSP.

We apologize for the lack of background, which made the paper difficult to read. Following the reviewer's suggestion, we describe the task with a concrete example. Materials synthesis planning (MSP) asks how a target inorganic material, specified only by its chemical composition, can be synthesized. In solid-state synthesis, a synthesis plan consists of the starting chemicals (precursors), the sequence of operations applied to them (e.g., mixing, heating, quenching), and the processing conditions (heating temperature, time, and atmosphere). A target can usually be synthesized from several precursor sets and procedures, and each attempt occupies a furnace for hours, so a planner must also decide which of the candidate routes to try within a limited budget. Table A illustrates the task with a target from our test set.

Table A. An example of the task (test target Ba₄Mn₃O₁₀, budget of three furnace runs of up to eight samples each)

StepWhat is decidedOutput of CondMSP for Ba₄Mn₃O₁₀
1. PrecursorsWhich starting chemicals to useRanked precursor sets: BaCO₃ + Mn₂O₃ (probability 0.40), BaCO₃ + MnO₂ (0.35), followed by sets with lower probability
2. ConditionsHow hot, how long, and in which atmosphere to fire each setBaCO₃ + Mn₂O₃: 1127 °C, 41.6 h, air. BaCO₃ + MnO₂: 1135 °C, 50.0 h, air
3. OperationsWhich sequence of operations to applyRanked sequences for each set, e.g., mixing → heating; mixing → heating → quenching; heating → quenching
4. SelectionWhich routes to try, and which share a furnace runTwo runs for routes of the BaCO₃ + Mn₂O₃ set and one run for routes of the BaCO₃ + MnO₂ set
OutcomeIs the reference route (the route reported in the literature) among those tried?Yes: the reference route, BaCO₃ + MnO₂ followed by heating and quenching, is in the third run

In this example, the precursor model ranks BaCO₃ + Mn₂O₃ above the reference precursor set BaCO₃ + MnO₂. CondMSP predicts the conditions for each set, generates operation sequences, and, with three furnace runs, tries routes of both sets, among which is the reference route. A pipeline that tries candidates only for the top-ranked precursor set, as MSP-LLM does, misses it. The example also shows the two roles of the predicted conditions in CondMSP: they serve as process context for operation prediction (step 3), and they determine which routes can share a furnace run (step 4), since routes with the same predicted conditions can be fired together.

In the revised manuscript, we have added this example as Section 2.4 (An Example of MSP), which presents the task itself, and we have added Section 3.4 (End-to-End Pipeline), which follows how CondMSP processes the same target step by step. Section 1 now points to the example when MSP is first defined, and Appendix G describes the basic background of solid-state synthesis and data-driven MSP and gives the full example as Table 23.

[W2] Related work.

A related-work section is provided in Appendix A of the submission. Because of the page limit, it was placed in the appendix rather than in the main text, which made it easy to miss. It discusses materials synthesis prediction, including thermodynamics-guided route ranking, text-mined recipe datasets, and the prediction of precursors, synthesis conditions, and operation sequences, as well as self-driving laboratories.

In brief, prior data-driven methods predict precursors [1, 2, 3], synthesis conditions [4, 5], or operation sequences [6], and MSP-LLM [7] connects precursor prediction with precursor-conditioned operation prediction. These methods output candidate recipes, but they neither use predicted conditions as context for operation prediction nor decide which candidates to try under a limited budget; CondMSP addresses both. Self-driving laboratories such as the A-Lab [8] execute candidate recipes in furnace batches, and CondMSP provides the per-target step of choosing which routes to try and which can share a furnace run. In the revised manuscript, we have added a related-work section to the main text as Section 5, which condenses Appendix A and refers to it for the extended discussion.

AuthorsOpenReview ↗

Author Response (2/3)

[W3] Training and test sets, data flow, and evaluation.

We apologize that the problem setting and the data flow were not clearly described in the submission. There are separate training, validation, and test sets. Following MSP-LLM [7], the main benchmark is built from text-mined solid-state synthesis recipes [9] and split into 8,295 training, 932 validation, and 918 test targets. No target is shared across the splits, and no test recipe comes from a paper that appears in the training set. The condition model additionally uses 27,110 training records annotated with at least one of temperature, time, or atmosphere. A second set of 2,344 new targets [10] is used only for evaluation. On the training set, each module of CondMSP learns one mapping. On the test set, the modules are applied in sequence, starting from the target composition alone (Table B).

Table B. Task and data flow

ModuleWhat it learns from the training setInput at test timeOutput at test time
Precursor prediction (from MSP-LLM)Target → precursor setTarget compositionTop-10 precursor sets with probabilities
Condition prediction(Target, precursor set) → maximum heating temperature, maximum heating time, and atmosphereTarget and each predicted precursor setPredicted conditions for each precursor set
Operation prediction (generation and retrieval)(Target, precursor set, conditions) → operation sequence; the retrieval index stores the training proceduresTarget, predicted precursor set, and predicted conditionsRanked operation sequences for each precursor set
Route selection (MILP)Nothing is learned; the route utility combines the precursor probability and the operation rankCandidate routes, their utilities, predicted conditions, and the budgetRoutes to try and their assignment to furnace runs

At test time, every module uses the predicted outputs of the preceding modules, and the reference route of the test target, i.e., the route reported in the literature, is used only for scoring. The validation set is used to select the combination thresholds of the condition model and to calibrate the fusion of generated and retrieved operation sequences.

The main metric, recipe recovery, is the fraction of test targets for which at least one selected route has exactly the reference precursor set and operation sequence. It is measured under three kinds of budgets: the number of routes tried (trials), the number of furnace runs, each holding up to eight samples with compatible predicted conditions, and the total furnace-hours, including a 2-h overhead per run. In addition, Table 1 (paper) evaluates condition prediction, and Table 2 (paper) evaluates operation prediction with the literature precursors given as input, so that each module can also be assessed on its own. In the revised manuscript, Section 3.4 (End-to-End Pipeline) describes this data flow, a condensed version of Table B appears as Table 22 in Appendix G, and the data splits and evaluation metrics are described in the Datasets paragraph of Section 4 and the Metrics paragraph of Section 4.3.

AuthorsOpenReview ↗

Author Response (3/3)

[W4] Comparison with stronger LLMs.

Following the reviewer's suggestion, we evaluated GPT-6.1-sol and Claude Opus 5.5 under the protocol of the GPT-5.4 mini baselines in Table 3 (paper): the same fixed 100-shot examples, 30 ranked candidates per response, and the same scorer and furnace planner on all 918 test targets. Table C reports PP→SOP prompting, in which a model first predicts a precursor set and then 30 operation sequences for it, because it gave higher recovery than end-to-end prompting for both models.

Table C. Recipe recovery (%) of frontier LLMs under trial and furnace budgets (918 test targets). The rows for GPT-5.4 mini (100-shot), MSP-LLM, and CondMSP are from Table 3 (paper).

Method1 trial10 trials30 trials1 run3 runs10 runs48 h
GPT-5.4 mini (100-shot)7.1912.6918.0812.6414.8118.0815.31
GPT-6.1-sol (100-shot)6.7516.8827.2315.4720.8127.2322.00
Claude Opus 5.5 (100-shot)6.8617.2127.8916.9922.2227.8923.09
MSP-LLM (LLaMA-3.1-8B)7.8122.5534.6820.0427.8934.6827.92
CondMSP (LLaMA-3.1-8B)8.4626.1840.0121.7132.9042.8531.99

As shown in Table C, these models are much stronger than GPT-5.4 mini, recovering 27.23% and 27.89% of the targets at 30 trials compared with 18.08%, but they remain below CondMSP at every budget: at 30 trials and with three furnace runs, Claude Opus 5.5 reaches 27.89% and 22.22%, compared with 40.01% and 32.90% for CondMSP. They are also below MSP-LLM, which fine-tunes an 8B open model on the same training data. Strong general-purpose LLMs therefore improve substantially over the light versions, but a planner trained for this task remains clearly better. In the revised manuscript, we have added these results to Appendix I (Table 26) and referred to them in Section 4.3.

[1] He, Tanjin, et al. "Precursor recommendation for inorganic synthesis by machine learning materials similarity from scientific literature." Science advances 9.23 (2023): eadg8180.

[2] Noh, Heewoong, et al. "Retrieval-retro: retrieval-based inorganic retrosynthesis with expert knowledge." Advances in Neural Information Processing Systems 37 (2024): 25375-25400.

[3] Prein, Thorben, et al. "Retro-rank-in: a ranking-based approach for inorganic materials synthesis planning." arXiv preprint arXiv:2502.04289 (2025).

[4] Huo, Haoyan, et al. "Machine-learning rationalization and prediction of solid-state synthesis conditions." Chemistry of Materials 34.16 (2022): 7323.

[5] Karpovich, Christopher, et al. "Interpretable machine learning enabled inorganic reaction classification and synthesis condition prediction." Chemistry of Materials 35.3 (2023): 1062-1079.

[6] Na, Gyoung S. "Artificial intelligence for learning material synthesis processes of thermoelectric materials." Chemistry of Materials 35.19 (2023): 8272-8280.

[7] Noh, Heewoong, et al. "MSP-LLM: A Unified Large Language Model Framework for Complete Material Synthesis Planning." arXiv preprint arXiv:2602.07543 (2026).

[8] Szymanski, Nathan J., et al. "An autonomous laboratory for the accelerated synthesis of inorganic materials." Nature 624.7990 (2023): 86-91.

[9] Kononova, Olga, et al. "Text-mined dataset of inorganic materials synthesis recipes." Scientific data 6.1 (2019): 203.

[10] Lee, Sanghoon, et al. "Text-mined dataset of solid-state syntheses with impurity phases using Large Language Model." Scientific Data 12.1 (2025): 1969.

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

The paper proposes CondMSP, an LLM-based system for planning the synthesis of inorganic materials under budget constraints. Instead of only predicting the precursors for the target material, it predicts the precursors, uses them to generate procedures followed by selecting the routes for the execution, all within the furnace budgets. The method combines precursor prediction, condition prediction, recipe-generation and retrieval along with resource optimization moving beyond traditional methods that target each of these parts in isolation. Across different experiments, CondMSP is seen to improve results when compared to the provided baselines.

Critical strengths
  • The paper addresses an existing bottleneck and goes beyond recipe prediction to solve a more practical problem of resource aware synthesis.

  • The paper integrates each component of the workflow to provide an end-to-end framework, which is more realistic than independently evaluating each component.

  • Takes into consideration the budgetary constraints faced for synthesis.

  • The paper is well-motivated and clear.

Critical weaknesses
  • The paper is very crammed. Tables are tiny and hard to read. The spacing between paragraphs is less. Its hard to read.

  • The paper does not perform any laboratory validation. While this might not be feasible, I recommend mentioning it as a limitation as it is an important point.

  • Are there any novel recoveries? Or is the focus only on recovery instead of novelty?

  • As the datasets are public, it could happen that the data might have appeared in the pretraining of the models. Thus, it is necessary to compare the method with bigger proprietary models such as GPT 5.6/ Claude Opus/ Gemini Pro, instead of comparing only with smaller GPT-5.4-mini and GPT-4o-mini.

  • The paper does not provide a RAG-like capability to the LLMs and compare it. This will help in identifying how much of the gains are because of the method vs having a richer data corpus provided to the models.

  • Are there cases when the outputs are not feasible or incorrect but get selected because of the budget constraints? What happens for incorrect outputs? How are they handled?

Questions for authors

See weaknesses

6 comments
AuthorsOpenReview ↗

Author Response (1/3)

We sincerely thank the reviewer for the positive assessment and constructive feedback. We address each point below with additional experiments and clarifications.

[W1] Dense layout and small tables.

We apologize for the inconvenience this caused and agree that the paper is too dense to read comfortably. To fit the page limit, we reduced the size of the tables and the spacing between paragraphs too far. In the revised manuscript, we have used the additional page to make the main text less dense: we have increased the spacing, moved parts of the main-text tables to the appendix (Appendix E.3), and made the remaining tables easier to read.

[W2] Laboratory validation as a limitation.

We agree with the reviewer that the lack of laboratory validation is an important limitation. We did not carry out the proposed syntheses: all methods are evaluated retrospectively against routes reported in the literature, and the furnace costs come from a simulated resource model. A recovered route therefore matches a synthesis reported to produce the target, but this does not establish phase purity or synthesis success, and a route that differs from every literature route cannot be assessed without experiments. The submission describes this in the Limitations (Appendix F) and the Ethics Statement, but the main text neither states that no laboratory validation was performed nor refers to the Limitations. In the revised manuscript, we have stated in the evaluation setup of Section 4.3 that all methods are evaluated on literature-derived benchmarks without laboratory validation, referring to the Limitations (Appendix F), and we have stated this limitation explicitly at the beginning of the Limitations.

[W3] Novelty of the recovered routes.

As the reviewer suggested, we further analyzed whether the routes recovered by CondMSP are novel. All test targets are new to the model: no test target appears in the training set, and no test recipe comes from a paper that appears in the training set. We compared every recovered route with the 8,295 training recipes (Table A).

Table A. Novelty of the recovered routes relative to the training recipes (% of recovered routes; 3 seeds)

MethodBudgetRecovery(Precursor set, operation sequence) not in trainingPrecursor set not in training
MSP-LLM30 trials34.6891.853.5
CondMSP30 trials40.0190.953.4
CondMSP3 furnace runs32.9090.854.1
CondMSP10 furnace runs42.8591.554.7

As shown in Table A, about 91% of the routes CondMSP recovers are combinations of a precursor set and an operation sequence that never occur in the training recipes, and about 54% use a precursor set that never appears in training. CondMSP therefore recovers the literature routes mostly by composing new combinations of precursors and operations, rather than by retrieving stored recipes.

Whether CondMSP can also find routes that differ from every reported synthesis, such as an entirely new route for a known target, cannot be confirmed without experiments, as discussed in [W2]. We therefore use recovery of literature-reported routes as the measurable criterion. In the revised manuscript, we have added this analysis to Appendix K.3 (Table 34).

[W4, W5] Comparison with frontier LLMs, with and without retrieval.

Following the reviewer's suggestion, we compared CondMSP with two frontier models, Claude Opus 5.5 and GPT-6.1-sol, and also gave the LLMs retrieval over the same training corpus that CondMSP uses (Table B). All runs follow the protocol of the GPT-5.4 mini baselines in Table 3 (paper): the same fixed 100-shot examples, 30 ranked candidates per response, and the same scorer and furnace planner on all 918 test targets. We report PP→SOP prompting, in which a model first predicts a precursor set and then 30 operation sequences for it, because it gave higher recovery than end-to-end prompting for both models (27.89% vs. 26.80% and 27.23% vs. 24.73% at 30 trials). For retrieval, the random 100 in-context examples are replaced, for each test target, by the 100 training records whose elemental composition is most similar to the target. Only the training split is searched. We did not use web search, since it could return the source paper of a test recipe. We applied retrieval to GPT-5.4 mini, our original baseline, and to both frontier models.

AuthorsOpenReview ↗

Author Response (2/3)

Table B. Recipe recovery (%) of frontier LLMs with and without retrieval under trial and furnace budgets (918 test targets). The rows for GPT-5.4 mini (100-shot), MSP-LLM, and CondMSP are from Table 3 (paper).

Method1 trial10 trials30 trials1 run3 runs10 runs48 h
GPT-5.4 mini (100-shot)7.1912.6918.0812.6414.8118.0815.31
GPT-5.4 mini + retrieval7.6320.0428.5418.4122.8828.5423.42
Claude Opus 5.5 (100-shot)6.8617.2127.8916.9922.2227.8923.09
Claude Opus 5.5 + retrieval9.0425.2736.7122.8828.6536.7129.85
GPT-6.1-sol (100-shot)6.7516.8827.2315.4720.8127.2322.00
GPT-6.1-sol + retrieval8.1724.7335.0820.9228.1035.0828.43
MSP-LLM (LLaMA-3.1-8B)7.8122.5534.6820.0427.8934.6827.92
CondMSP w/o conditions (LLaMA-3.1-8B)8.2425.2737.8720.8432.5341.8331.15
CondMSP (LLaMA-3.1-8B)8.4626.1840.0121.7132.9042.8531.99

As shown in Table B, without retrieval the frontier models are much stronger than GPT-5.4 mini, but they remain below CondMSP at every budget: at 30 trials, Claude Opus 5.5 and GPT-6.1-sol recover 27.89% and 27.23% of the targets, compared with 40.01% for CondMSP, and with three furnace runs 22.22% and 20.81%, compared with 32.90%. They are also below MSP-LLM at every budget. As the reviewer notes, the public recipes may appear in the pretraining data of LLMs. Such exposure would benefit the frontier models as well, yet they remain below CondMSP. In addition, CondMSP and MSP-LLM use the same LLaMA-3.1-8B backbone and the same training data, so any exposure affects both equally and cannot explain the gain of CondMSP over MSP-LLM.

Retrieval improves all LLMs substantially: at 30 trials, Claude Opus 5.5 improves from 27.89% to 36.71%, GPT-6.1-sol from 27.23% to 35.08%, and GPT-5.4 mini from 18.08% to 28.54%. A large part of the gap between our original LLM baselines and CondMSP therefore comes from access to relevant training data. Even with the same training corpus, however, CondMSP remains ahead, and its margin grows with the budget: at 30 trials it recovers 40.01%, compared with 36.71% and 35.08% for Claude Opus 5.5 and GPT-6.1-sol with retrieval, and with three furnace runs 32.90%, compared with 28.65% and 28.10%. Only at the smallest budgets, where a single route or a single furnace run is tried, is Claude Opus 5.5 with retrieval comparable to CondMSP.

The rows of Table B also separate the contributions at 30 trials. Retrieval adds 7.85–8.82 points to the frontier models; CondMSP without conditions, which uses the same training data, adds a further 1.16–2.79 points through its candidate construction and route selection; and the predicted conditions add another 2.14 points. Beyond access to the training data, both the candidate construction and selection of CondMSP and its use of predicted conditions therefore contribute to its advantage.

In the revised manuscript, we have added these results to Appendix I (Table 26) and referred to them in Section 4.3.

AuthorsOpenReview ↗

Author Response (3/3)

[W6] Infeasible or incorrect routes selected under the budget.

Outputs that cannot be parsed into a precursor set and an operation sequence are removed before scheduling (Appendix D.3). Among the remaining routes, the planner selects those with the highest utility within the budget. Under a furnace budget, it fills each furnace run with routes whose predicted conditions are compatible, so the number of selected routes grows with the budget, from 8.0 per target with one run to 57.5 with ten runs. Routes that differ from the literature route are therefore selected, and this is how the budget is used: about 75% of CondMSP's successful cases at three runs come from routes other than the top-ranked one (see our response to Reviewer 1Wge, W1). An incorrect route consumes one trial or one furnace slot and counts as a miss. Since the plan is made in a single shot, without experimental feedback, it is not revised afterwards. To check whether the routes added under larger budgets are chemically infeasible, we examined every selected route (Table C). We count a route as element-consistent when its precursors supply every element of the target other than O, C, H, and N, and introduce no metal or metalloid absent from the target.

Table C. Element-level validity of the selected routes

MethodBudgetRoutes per targetElement-consistentMissing target elementInvalid and incorrect
MSP-LLM3 runs16.299.410.590.59
MSP-LLM10 runs29.999.480.520.52
MSP-LLM30 trials30.099.460.540.54
CondMSP1 run8.099.350.650.65
CondMSP3 runs19.899.250.750.74
CondMSP10 runs57.598.640.951.25
CondMSP48 h26.598.960.900.95
CondMSP30 trials30.099.390.600.60

As shown in Table C, 98.64–99.39% of the routes CondMSP selects are element-consistent at every budget, and the share of routes that are both invalid and incorrect stays between 0.60% and 1.25% while the number of selected routes grows sevenfold from one to ten runs, close to that of MSP-LLM (0.52–0.59%). These routes occur for 9 of the 918 targets at three runs, mostly when an anion source is missing (e.g., no fluorine source for Sn₂TiNbO₆F; see our response to Reviewer 1Wge, W1). The routes added under larger budgets are therefore mostly chemically consistent alternatives to the literature route, not infeasible ones. In the revised manuscript, we have added this audit to Appendix K.2 (Table 33).

Reviewer_1WgeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CondMSP, a framework for inorganic materials synthesis planning that uses predicted synthesis conditions both as process context for synthesis operation prediction and as execution requirements for resource-aware route selection.

Critical strengths
  • The authors introduce conditions of inorganic synthesis methods into synthesis planning, which gives full experiment instructions for scientists.
  • The presentation is clear.
Critical weaknesses
  1. A worked analysis of representative successful and unsuccessful predictions would help establish whether the selected routes are chemically plausible.
  2. The practical significance of furnace capacity remains insufficiently. This is a meaningful lab scheduling constraint. However, the reported gains are not intended to improved phase-pure synthesis or functional performance.
  3. A synthesis route containing multiple thermal stages. However, CondMSP predicts a single condition tuple for each target-precursor pair. This loses the correspondence between individual operations and their processing requirements. Thus, the synthesis condition may not well used.
  4. Varying the number of precursor candidates and SOP candidates per precursor set may lead to different results. Sensitivity analysis on recipe recovery and oracle recovery rate are needed.
  5. Why choose the highest synthesis temperature as the prediction target. For luminescent materials, including perovskite systems, a reported temperature may include one or two synthesis temperatures above the best performance temperature. From Table 1, Karpovich uses regression approach averages reported conditions for the same target-precursor pair. The same processing method could be applied.
Questions for authors

See weaknesses.

7 comments
AuthorsOpenReview ↗

Author Response (1/5)

We sincerely thank the reviewer for the constructive and detailed feedback. We address each point below with additional experiments and analyses.

[W1] Worked analysis of successful and unsuccessful predictions.

As the reviewer suggested, we analyzed the successful and unsuccessful predictions on the test set and report representative cases below. A prediction is successful when one of the selected routes has the same precursor set and the same operation sequence as the literature route, i.e., when the literature route is recovered. All cases below use a budget of 3 furnace runs, and Table B also gives the share of each failure type at a budget of 30 trials.

Successful cases. Table A shows one successful case for each of the two main ways in which the literature route is recovered. In the first, the top-ranked route is the literature route. In the second, the precursor set is ranked first but the top-ranked operation sequence is not the literature one, and the plan recovers the route because it keeps several operation sequences. The second way is the more frequent one: it accounts for 65.1% of the successful cases, against 25.4% for the top-ranked route (the remaining 9.5% come from a lower-ranked precursor set). Conditions do not enter the success criterion. CondMSP predicts one condition per target–precursor pair, the maximum heating temperature and time, and the furnace evaluation applies it to every thermal stage of the route (see our response to W3), so the tables list one predicted temperature and time per route.

Table A. Successful cases (literature route recovered)

TargetLiterature route (precursors; operations)Temperature / time: literature → CondMSPHow it was recovered
(Ba₀.₈₅Ca₀.₁₃Y₀.₀₂)(Zr₀.₁Ti₀.₉)O₃BaCO₃ + CaCO₃ + TiO₂ + Y₂O₃ + ZrO₂; mixing → drying → heating → mixing → shaping → sintering1150 °C / 6 h, then 1400 °C / 6 h → 1400 °C / 4.0 hTop-ranked route: first-ranked precursor set, first-ranked operation sequence.
BaCe₀.₈₅Y₀.₁₅O₃₋δBaCO₃ + CeO₂ + Y₂O₃; mixing → heating → shaping → sintering1250 °C / 12 h, then 1450 °C / 12 h → 1456 °C / 10.0 hFirst-ranked precursor set, lower-ranked operation sequence (rank 4 or 6, depending on the seed). The top-ranked sequence, mixing → heating, is not the literature one.

Unsuccessful cases. We grouped the unsuccessful cases into three types by how the closest selected route differs from the literature route (Table B). In F1 (46.1% of the unsuccessful cases) the precursors and the number of thermal stages are the same, and the routes differ in non-thermal operations, in their order, or in the type of a thermal stage. In F2 (29.3%) the precursors are the same and the number of thermal stages differs. In F3 (24.6%) no selected route uses the literature precursor set. At 30 trials the shares are 54.8%, 29.2%, and 16.0%; F3 falls because more precursor sets are selected.

Table B. Failure types

Failure typeDefinition3 furnace runs (n = 1,848)30 trials (n = 1,652)
F1Same precursors and same number of thermal stages. The routes differ in non-thermal operations (mixing, drying, shaping, quenching), in their order, or in the type of a thermal stage (heating, sintering, or annealing).46.1%54.8%
F2Same precursors, different number of thermal stages.29.3%29.2%
F3Different precursor set: no selected route uses the literature precursor set.24.6%16.0%

Table C gives case studies of each type. In F1 the selected route can describe the same experiment as the paper: for BaTi₀.₈₅Sn₀.₁₅O₃ the precursors are the same, both routes have a heating stage followed by a sintering stage, and the predicted maximum temperature is 47 °C above the final literature temperature; the match fails because the second mixing step comes after the first thermal stage instead of before it. In F2 the selected route fires the same precursors once at the final temperature and omits the pre-reaction stage. One-stage routes of this kind are reported: three papers in our corpus prepare the parent compound YMnO₃ in one thermal stage at 1300 °C. In F3 the outcome depends on the case. For Ca₃(VO₄)₂ the selected set replaces NH₄VO₃ with V₂O₅, which is what 7 other papers on this compound in our corpus use. For Sn₂TiNbO₆F the selected routes are wrong: none contains a fluorine source. Such routes are rare: 0.75% of the selected routes, in 9 of the 918 targets, lack a source for an element of the target.

AuthorsOpenReview ↗

Author Response (2/5)

Table C. Unsuccessful cases by failure type

TypeTargetLiterature route (precursors; operations)Closest route selected by CondMSPTemperature / time: literature → CondMSPWhat differs, and whether the selected route is chemically plausible
F1BaTi₀.₈₅Sn₀.₁₅O₃BaCO₃ + SnO₂ + TiO₂; mixing → drying → mixing → heating → shaping → sinteringSame precursors; mixing → drying → heating → mixing → shaping → sintering1100 °C / 2 h, then 1370 °C / 2 h → 1417 °C / 4.0 hTwo thermal stages in both, of the same types; the second mixing step comes after the first thermal stage instead of before it. Plausible: same precursors and the same heating-then-sintering process; regrinding the powder after the first heating, before shaping, is the usual ceramic procedure.
F2YMn₀.₉₅Co₀.₀₅O₃Co₃O₄ + MnO₂ + Y₂O₃; heating → mixing → sinteringSame precursors; mixing → sintering1100 °C / 24 h, then 1300 °C / 24 h → 1319 °C / 24.0 hTwo thermal stages in the literature, one in the selected route: the pre-reaction at 1100 °C and the intermediate mixing are omitted. Plausible: same precursors, fired once at the final temperature (within 19 °C) for the same time; three papers in our corpus prepare the parent compound YMnO₃ from Y₂O₃ and a manganese oxide with one thermal stage at 1300 °C.
F3Ca₃(VO₄)₂CaCO₃ + NH₄VO₃; mixing → sinteringCaCO₃ + V₂O₅; mixing → heating, and in two of the three seeds also mixing → sintering. The NH₄VO₃ set is ranked third and gets no furnace run.900 °C / 10 h → 900 °C / 6.0 hOne precursor differs (vanadium source). Plausible: NH₄VO₃ decomposes to V₂O₅ on heating, and 7 other papers on this compound in our corpus use CaCO₃ + V₂O₅.
F3Sn₂TiNbO₆FNb₂O₅ + SnF₂ + SnO + TiO₂; mixing → heating → mixingNb₂O₅ + SnO₂ + TiO₂ and Nb₂O₅ + SnO + TiO₂. The literature set is not among the 10 proposed sets; the third-ranked set contains SnF₂ but no SnO and gets no furnace run.850 °C / 12 h → 1111 °C / 6.8 h (SnO₂ set) and 1082 °C / 4.0 h (SnO set)Not plausible: no selected route contains a fluorine source.

Bold in the literature route marks the precursors in which the selected routes depart from it.

In the revised manuscript, we have added more cases for each type, together with the failure-type statistics, to Appendix K.1 (Tables 29–32).

[W2] Practical significance of the gains under furnace constraints.

As the reviewer notes, the reported gains do not demonstrate improved phase purity or functional performance, since we did not carry out the proposed syntheses. They instead address a practical decision in synthesis planning: which candidate routes to try for a new target within a limited number of furnace runs.

Autonomous laboratories face this decision in practice. For each target, the A-Lab [1] first tests up to five recipes proposed by a model trained on syntheses from the literature, heating them in furnace batches of up to eight samples, and uses one fixed temperature per target so that these recipes can be heated together in a single furnace. Only 30% of the 353 recipes it tested produced their targets, yet 63% of the targets were eventually obtained by trying several candidates per target. Most of these, 30 of the 36, came from the initial literature-based recipes, tried before any active-learning optimization. Proposing several literature-based candidates per target, and choosing which of them to try within the limited furnace runs, therefore matters in practice.

Our evaluation measures how well a method makes this choice, without running experiments. Following prior MSP work on precursor prediction [2, 3], synthesis operation prediction [4], and synthesis planning [5], we take routes reported in the literature to produce the target as the reference for success. These studies evaluate a fixed number of top-ranked candidates, as if every candidate were an equally costly attempt. We keep this evaluation and report it as trial budgets. In solid-state synthesis, however, furnace use is a major cost, since each run occupies a furnace for hours. We therefore additionally evaluate each method under a furnace budget, in a setting similar to the A-Lab’s: candidate routes for one target share a furnace run of up to eight samples (C = 8) only when their predicted conditions are compatible. Every method, including all baselines, receives the same budget, counted as the total number of furnace runs or furnace-hours used for the target, regardless of how many furnaces are operated in parallel. Under this budget, a gain in recovery means that, within the same number of furnace runs, more targets have a literature-reported route among the routes tried, or equivalently, that fewer runs are needed to reach the same number of targets.

AuthorsOpenReview ↗

Author Response (3/5)

Two results in Table 3 (paper) show this. First, CondMSP as a whole, that is, its candidate routes together with its selection, recovers 32.90% of the targets with 3 furnace runs, compared with 27.89% for MSP-LLM with the same 3 runs, and this matches what MSP-LLM recovers with 5 runs (32.75%). Second, our selection method improves recovery even when the candidate routes are the same. MSP-LLM and the other baselines try candidate routes in the order they were generated, regardless of which routes could share a run, whereas CondMSP selects the routes to try and assigns them to shared furnace runs jointly, by solving an MILP that maximizes total route utility within the budget. With the candidate routes held fixed, this selection alone raises recovery by 2.6 to 5.8 points at 3, 5, and 10 runs and at 48 and 96 h, for MSP-LLM’s candidates as well as ours. Part of the gain under furnace budgets therefore comes from how the furnace runs are used, not only from the candidate routes themselves.

In the revised manuscript, we have stated in the evaluation setup of Section 4.3 that no laboratory validation was performed, and described the role of the furnace constraint and this reading of the gains in the Limitations (Appendix F) and in Appendix E.3.

[W3] Multiple thermal stages and a single condition tuple.

We agree with the reviewer that, for a route with multiple thermal stages, a single condition tuple per target–precursor pair loses the correspondence between individual operations and their processing requirements.

In the submission, we formulated SCP following [6]: the target is the representative heating temperature and time of each record, defined as the maximum over its heating operations as in their released implementation (please see our response to W5). In MSP, we then assigned this predicted condition to every thermal stage of a route, and applied the same furnace constraints to all compared methods for consistency. In our benchmark, 65.9% of recipes contain a single thermal stage, for which the predicted condition corresponds to that stage. For the remaining 34.1%, which contain two or more thermal stages, every stage receives the predicted maximum, so the predicted condition does not correspond to the stages below it.

Following the reviewer's feedback, we extended the condition model so that each thermal operation receives its own condition. Since CondMSP also generates the operation sequence, the thermal operations of each route are known once the route is generated. We therefore gave the operation sequence to the condition model as an additional input and trained it on the stage-level temperature and time labels of the training recipes, so that it predicts a (temperature, time) pair for each thermal operation. At test time, the inputs are the predicted precursors and the generated operation sequence, and the reference recipe is used only for scoring. Table D reports the condition error for test targets whose reference recipe contains multiple thermal stages, separately for the hottest stage and the other stages, averaged over three seeds. The errors are computed on generated routes whose sequence of thermal operations matches the reference, so that each predicted stage can be compared with its reference stage; such a route exists for 68% of these targets.

Table D. Condition error per thermal stage on routes with multiple thermal stages.

Single condition on every stage (submission)Per-operation conditions
Hottest stage: temperature MAE107.0 °C109.0 °C
Other stages: temperature MAE277.1 °C169.8 °C
Other stages: time MAE8.8 h7.8 h

As shown in Table D, the single condition is accurate on the hottest stage, on which it is defined (107.0 °C), but not on the other stages (277.1 °C), since their temperatures are lower than the maximum. With the per-operation conditions, the error on the other stages decreases to 169.8 °C in temperature and from 8.8 h to 7.8 h in time, while the error on the hottest stage remains similar (109.0 °C). For example, the reference recipe for Bi₀.₄₆Na₀.₄₇Ba₀.₀₆Yb₀.₀₁TiO₃ (a lead-free ceramic) heats at 850 °C for 2 h and then sinters at 1200 °C for 2 h. The submission assigns 1115 °C for 2 h to both steps, whereas the per-operation model predicts 861 °C for 2 h for the heating step and 1163 °C for 2 h for the sintering step. These results show that the predicted conditions correspond more accurately to the individual thermal stages.

With the per-operation conditions, each thermal stage can enter the furnace constraints with its own temperature and time, rather than with the route maximum repeated at every stage. We therefore repeated the furnace-budget experiment of Table 3 (paper) under these stage-level constraints, applying the same per-operation model to the routes of all compared methods (Table E).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal

We thank the reviewers for their time and constructive feedback. We appreciate that the reviewers recognized:

  1. An inorganic synthesis planning method for a key practical bottleneck (fUXT, CdkW). Beyond predicting recipes, deciding which synthesis routes to actually run is an important and practical bottleneck, and the paper proposes a clearly motivated method for this resource-aware synthesis planning.
  2. Synthesis conditions as part of the plan (1Wge). Predicted conditions are used both to generate synthesis operations and to select the routes to run, so that the plan gives full experimental instructions for scientists.
  3. An end-to-end framework under budget constraints (fUXT). Precursor prediction, condition prediction, operation generation, and route selection are integrated into a single framework that plans within trial and furnace budgets, rather than treating each component in isolation.

We summarize the main discussion points below:

  1. Practical significance of the gains (1Wge W2, fUXT W2). We acknowledge that the gains do not demonstrate phase purity or functional performance, since we did not carry out the syntheses, and have stated this as a limitation. The gains address which candidate routes to try for a new target within a limited number of furnace runs, a decision that autonomous laboratories such as the A-Lab face in practice. As in many prior MSP studies, we evaluate recovery of literature-reported routes, and we additionally evaluate it under furnace budgets, where CondMSP with 3 runs recovers as many targets as MSP-LLM with 5.
  2. Conditions for routes with multiple thermal stages (1Wge W3). In the submission, the predicted condition was applied to every thermal stage, consistently for all methods including the baselines. Following the reviewer's comment, we additionally predicted a separate condition for each thermal stage, which substantially reduces the error on the stages below the maximum temperature (277.1 → 169.8 °C), and re-evaluated all methods under these stage-level constraints: CondMSP remains the best.
  3. Choice of the maximum temperature (1Wge W5). We followed the released implementation of prior work using the same synthesis dataset. Retraining the full condition model on labels averaged over each target–precursor pair, as in Karpovich et al., gives similar errors and leaves the ranking of the methods unchanged.
  4. Frontier LLMs and retrieval (fUXT W4–W5, CdkW W4). Claude Opus 5.5 and GPT-6.1-sol remain below CondMSP at every budget. Given retrieval over the same training corpus, they improve substantially but are comparable to CondMSP only at the smallest budgets, so the gains of CondMSP do not come only from access to the data.
  5. Additional analyses (1Wge W1, W4; fUXT W3, W6). We added a worked analysis of successful and unsuccessful cases, a sensitivity analysis over the numbers of precursor and SOP candidates, and analyses of the novelty and element-level validity of the selected routes. Most failures use the literature precursors and differ only in their operations, often as chemically reasonable variants; CondMSP outperforms MSP-LLM in every candidate setting; most recovered routes are combinations not seen in training; and few selected routes are element-inconsistent.
  6. Presentation (CdkW W1–W3, fUXT W1). We added a concrete example of the task, a step-by-step description of the pipeline, a related-work section, and a data-flow table, and made the main text less dense.

All revisions are highlighted in blue in the revised paper.
We sincerely thank the reviewers again for their thoughtful comments, which helped us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.