acceptodds
Under review as a conference paper at ICLR 2027

Retro-Forge: A Multi-Step Pairwise Retrosynthesis Framework for Solid-State Materials Synthesis

Abstract

AI-driven materials discovery has made remarkable strides in generating stable, unique, and novel structures at unprecedented scale, yet a critical gap remains between generation and realization: without valid synthesis routes, proposed materials stay in silico. Existing approaches to precursor prediction (PP)—the first and most consequential step of materials synthesis planning (MSP)—formulate it as a single-step problem, overlooking the established domain knowledge that solid-state reactions proceed pairwise. We introduce Retro-Forge, a multi-step pairwise retrosynthesis framework that for the first time casts PP as a sequence of learnable pairwise reactions. A single-step pairwise reactant prediction model is trained on a curated pairwise reaction dataset, expanded through synthetic data augmentation to address data scarcity. This model is composed recursively within a tree search to produce complete synthesis routes, and further optimized for this task via multi-step tree-RL. We construct a clean benchmark of 857 novel targets verified absent from all training data, on which Retro-Forge achieves state-of-the-art performance. Beyond precursor prediction accuracy, we independently corroborate that the large majority of its predicted pairwise steps are stoichiometrically balanced and thermodynamically favorable under DFT-derived reaction energies—demonstrating that the multi-step pairwise formulation is not only learnable but also chemically valid.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Precursor Prediction, Pairwise Reaction, Retrosynthesis, Data Augmentation, Material Science

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_2og5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an approach to inorganic synthesis based on pairwise reactions, i.e. the observation that solid-state reactions happen with pairs of precursors. The authors use an LLM to augment the dataset with additional reactions, use another LLM as the single-step model, and fine-tune the entire system with RL on synthesis routes. The method is evaluated with traditional synthesis metrics (top-k accuracy and frontier top-k), in addition to plausibility assessment based on stoichiometry and thermodynamical behavior.

Critical strengths
  • The paper identifies pairwise reactions as an important inductive bias under-utilized by previous methods.
  • The authors create a benchmark dataset based on pairwise reactions augmented with AI-generated samples which opens the way for future methods in this same line of work.
Critical weaknesses
  • (W1) Novelty compared to organic synthesis planning: This is the biggest weakness of the work, and the main reason for my rejection. The paper seems to borrow the idea of "a search algorithm running a single-step model recursively to produce a synthesis tree with purchasable leaves" directly from organic synthesis. While I understand inorganic chemistry traditionally generates the synthesis plan as a set of precursors, the search + single-step model was already fairly established in organic synthesis and seems to be borrowed verbatim with minor engineering efforts. The tree-RL fine-tuning was also already covered in organic synthesis [1,2] and not cited in this work. I can change my score if the authors show (both theoretically and empirically) what scientific contributions emerged from adapting this approach to inorganic chemistry.

  • (W2) Missing empirical results: Empirically, the evaluation does not isolate the effect of neither pairwise reaction prediction nor the search algorithm. If these are signficant components, they should be evaluated through the following ablations: 1) use the same single-step model you put in search to generate the precursors in a single-shot, and 2) run a search-based approach where the single-step model is trained on reactions with an arbitrary number of precursors. Additionally, traditionally baselines are cited but never compared empirically including [3], [4] and [5] to only name a few.

  • (W3) Secondary presentation concerns: below is a list of things to improve in the presentation of the paper.

    • framing: this is related to the novelty issue. If the paper's main contribution is the new dataset, it should be reframed around a dataset and evaluation contribution. Right now it reads more as a method's paper with many missing elements.
    • illustrative figures missing: Figure (1) should use a larger font. The authors could also add extra figures illustrating routes and precursor sets recovered by different baselines.
    • rewriting some sections: sections 5.1. to 5.3. mix evaluation setup with result reporting, when section 4.4 is supposed to have already described the experimental setup. I suggest rewriting to have all metrics' definitions and justifications in one place, preferably before presenting the results.
Questions for authors
  • What is the use of frontier top-k, beyond being a more relaxed version of precursor topk?
  • What would you say are the key lessons the ML community can learn from adapting search-based synthesis planning to inorganic synthesis?

References

[1] Liu, Xue et al. Retrosynthetic Planning with Dual Value Networks. ICML 2023. https://proceedings.mlr.press/v202/liu23as/liu23as.pdf

[2] GRASP: Navigating Retrosynthetic Planning with Goal-driven Policy. NeurIPS 2022. https://papers.neurips.cc/paper_files/paper/2022/file/42beaab8aa8da1c77581609a61eced93-Paper-Conference.pdf

[3] Prein et al. Retro-Rank-In: A ranking-based approach for inorganic materials synthesis planning. 2025. https://arxiv.org/abs/2502.04289

[4] Aykol, M., Montoya, J. H., and Hummelshøj, J. Rational solid-state synthesis routes for inorganic materials. Journal of the American Chemical Society, 143(24):9244–9259, 2021. doi:10.1021/jacs.1c04888.

[5] Szymanski, N. J., Nevatia, P., Bartel, C. J., Zeng, Y., and Ceder, G. Autonomous and dynamic precursor selection for solid-state materials synthesis (ARROWS3). Nature Communications, 14:6956, 2023. doi:10.1038/s41467-023-42329-9.

Reviewer_d4wgOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Retro-Forge formulates precursor prediction for solid-state inorganic synthesis as recursive pairwise reactant prediction. Given a target composition, an LLM predicts two reactants, and a tree-search procedure recursively decomposes intermediate compositions until the leaves belong to a precursor catalog. The framework combines a literature-derived two-reactant dataset, LLM-generated augmentation, staged fine-tuning, and tree-level reinforcement learning using precursor-set matching, stoichiometric balance, and diversity rewards. Candidate precursor sets are subsequently aggregated and reranked, with optional fusion of two model backbones.

Validation includes single-step prediction, multi-step precursor-set recovery, component ablations, computational chemistry checks, and a preliminary temperature-prediction study. On the newly constructed 857-target ICSD-2025 benchmark, the fused model reports 58.7% leaf Top-1 accuracy, compared with 52.9% for MSP-LLM; the corresponding Top-10 results are 67.6% and 66.9%. The evidence supports improved precursor-set recommendation, while validation of the predicted intermediate pathways remains less conclusive.

Critical strengths

The framework converts variable-size precursor-set prediction into a reusable two-reactant prediction task and optimizes the composed predictions through tree-level rewards. This is a coherent contribution: the training objective addresses whether local predictions combine into a successful precursor recommendation, rather than treating single-step accuracy as sufficient.

Critical weaknesses

Precursor-set accuracy does not establish the validity of the predicted intermediate pathways:

Tier 1 is obtained by retaining literature recipes with exactly two solid reactants, which does not itself establish that these records correspond to directly observed reaction steps. Likewise, the matching component of tree-RL supervises the terminal precursor set, not intermediate identities or reaction order. Consequently, distinct trees with the same leaves may receive identical matching rewards despite differing substantially in pathway validity.

There is a concrete example of this distinction in Figure 1 and Section 3: the proposed step LiVO3+Fe2O3→LiFeV2O7\mathrm{LiVO_3 + Fe_2O_3 \rightarrow LiFeV_2O_7} cannot conserve both Li and V, because their ratio is 1:1 in the only Li/V-containing reactant but 1:2 in the product. This is not resolved by adjusting coefficients or adding the permitted gaseous species.

The broader chemical checks also require a more qualified interpretation. The NNLS criterion excludes O, C, H, and N, while only 21-24% of step-targets have database formation energies; the approximately 85% computable coverage includes convex-hull fallback estimates. The reported reliable-DFT subset is helpful, but these checks do not by themselves validate the specific intermediate phases and reaction ordering. The evidence is therefore stronger for precursor recommendation than for chemically faithful pathway reconstruction.

The independence of model selection from the final test benchmark is unclear:

Appendix D states that its ablations use the 857-target benchmark, while Appendix D.4 describes selecting the final reward weights using these experiments and the “best checkpoint.” It is therefore unclear whether reward weights and checkpoints were selected on a separate validation set. This is not evidence of training-set contamination, but it leaves the independence of the reported test evaluation unresolved. Additionally, the ablations and final models use different versions of reward supervision and rerankers, complicating the transfer of ablation conclusions to the final system.

Questions for authors

Which dataset and metric were used to select the reward weights and the step-1,800 checkpoint? Were any model-selection decisions based on the 857-target benchmark? Please also clarify which ablations use the same filtered supervision and reranker-training protocol as the final models.

Does the implementation reject routes that reach the depth limit with non-catalog leaves? If it differs from Algorithm 1, please specify the actual completion criterion. If it follows the pseudocode, how frequently does this case occur, and what is its effect on the reported results?

Reviewer_1L3nOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Retro-Forge, a multi-step retrosynthesis framework based on pairwise reactions for precursor prediction (PP) in solid-state inorganic materials synthesis. Conventional approaches typically predict a complete set of precursors directly from the chemical composition of a target material. In contrast, this study formulates precursor prediction as a sequence of pairwise reactant prediction (PRP) tasks, based on the chemical insight that solid-state reactions proceed through successive pairwise reactions involving intermediate compounds.

The proposed framework expands approximately 8,000 pairwise reactions extracted from existing inorganic materials synthesis databases to approximately 120,000 reactions through LLM-based synthetic data augmentation. Reactant prediction models based on Qwen and LLaMA are then fine-tuned using a staged training strategy. The trained models are integrated into an AND-OR tree search algorithm that recursively decomposes a target material into intermediate compounds and starting materials, thereby generating multi-step synthesis routes. In addition, the authors introduce a reinforcement learning approach (Tree-RL) that optimizes the quality of entire synthesis routes rather than individual reaction predictions. The reward function incorporates agreement with ground-truth precursor sets, stoichiometric consistency, and diversity among generated routes.

To evaluate the proposed framework, the authors construct a benchmark comprising 857 novel inorganic materials reported in 2025. They compare Retro-Forge with existing precursor prediction methods, conduct ablation studies to assess the contributions of data augmentation and Tree-RL, and evaluate the stoichiometric consistency and DFT-based thermodynamic feasibility of the predicted reaction pathways. The results demonstrate that Retro-Forge outperforms existing methods in precursor prediction and that both data augmentation and route-level reinforcement learning contribute to improving the prediction accuracy and chemical consistency of multi-step synthesis routes.

Critical strengths
  1. Chemistry-informed formulation and route-level optimization: A major strength of this study lies in its integration of domain-specific chemical knowledge with multi-step retrosynthesis planning and reinforcement learning. By formulating precursor prediction as a sequence of pairwise reactions, the proposed framework explicitly models intermediate compounds and reaction pathways, rather than directly predicting a precursor set from the target composition. Furthermore, the introduction of tree-based reinforcement learning (Tree-RL) enables optimization of complete synthesis routes rather than individual reaction steps. Notably, the results demonstrate that Tree-RL improves multi-step precursor prediction accuracy and stoichiometric consistency while largely preserving single-step prediction performance. This finding highlights the importance of route-level optimization in multi-step synthesis planning.

  2. Evaluation on novel materials and chemical validity: Another strength of this study is the construction of a benchmark comprising 857 novel inorganic materials reported after the pretraining cutoffs of the models, providing a more rigorous assessment of generalization beyond memorized synthesis recipes. Furthermore, the authors evaluate not only precursor prediction accuracy but also the stoichiometric consistency and DFT-based thermodynamic feasibility of the predicted reaction pathways. This comprehensive evaluation strengthens the credibility of the proposed framework by assessing its predictions from both machine learning and chemical perspectives.

Critical weaknesses
  1. Validity of the pairwise reaction formulation and predicted synthesis pathways The central premise of this study is that solid-state synthesis can be represented as a sequence of pairwise reactions. While this chemistry-informed formulation is reasonable and potentially useful, it remains unclear to what extent the predicted intermediates and reaction sequences reflect actual solid-state reaction mechanisms. In practice, solid-state synthesis involves complex processes, including diffusion, nucleation, phase transformations, and competing reactions, which may not be adequately captured by a sequence of pairwise transformations. The current evaluation primarily focuses on precursor-set accuracy, stoichiometric consistency, and thermodynamic feasibility, without directly validating the predicted intermediates or reaction sequences. I encourage the authors to evaluate the predicted pathways against experimentally characterized reaction mechanisms, even for a limited number of representative materials. Such an analysis would help establish whether Retro-Forge predicts physically meaningful synthesis pathways rather than merely generating stoichiometrically consistent decompositions.

  2. Limited thermodynamic feasibility of the predicted pathways Although the proposed framework achieves high stoichiometric consistency, the thermodynamic evaluation reveals a substantial gap between stoichiometric validity and thermodynamic feasibility. Specifically, approximately 70–73% of the evaluated reaction steps are thermodynamically favorable, while only 52–53% of the evaluated routes consist entirely of favorable steps. This suggests that stoichiometric consistency alone is insufficient to ensure chemically feasible synthesis pathways. Furthermore, the thermodynamic assessment is performed at a fixed temperature of 900°C, which may not reflect the actual synthesis conditions of individual materials. The authors should discuss these limitations more thoroughly and clarify the extent to which the predicted pathways can be regarded as chemically feasible. It would also be valuable to investigate whether incorporating reaction Gibbs energies into the reward function,

Questions for authors
  1. Validity of the predicted reaction pathways To what extent do the predicted intermediates and reaction sequences correspond to experimentally observed solid-state reaction mechanisms? Can the authors provide a quantitative or qualitative comparison with experimentally characterized reaction pathways for a representative set of materials? This would help clarify whether the proposed framework captures physically meaningful reaction sequences rather than merely generating stoichiometrically consistent decompositions.
  2. Thermodynamic feasibility of the predicted pathways The thermodynamic evaluation indicates that only approximately 52–53% of the evaluated routes consist entirely of thermodynamically favorable steps, despite their high stoichiometric consistency. What are the primary reasons for the thermodynamically unfavorable steps? Are these discrepancies mainly attributable to limitations of the pairwise reaction formulation, inaccuracies in the predicted intermediates, or the thermodynamic evaluation assumptions, including the fixed temperature of 900°C? Clarifying these factors would help assess the chemical reliability of the proposed synthesis pathways.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.