acceptodds
Under review as a conference paper at ICLR 2027

OneMat: Generative Optimization from Scientific Literature for Materials Design

Abstract

Materials design still relies heavily on costly physical experimentation and search over predefined candidate spaces, limiting both the efficiency and breadth of exploration. Recent generative approaches in materials science have advanced candidate-level design by proposing novel compositions and atomic structures, but executable synthesis procedures for their realization remain largely outside iterative generation and optimization. We introduce OneMat, a literature-driven framework for synthesis-oriented generative optimization, where LLMs generate and iteratively refine executable synthesis plans. To enable this paradigm, OneMat transforms scientific literature into a structured domain resource containing evidence-grounded synthesis knowledge, generation examples, and preference signals, and integrates generation with retrieval and evaluation in an iterative optimization loop. We evaluate OneMat on acidic oxygen evolution reaction (OER) catalyst design using a literature corpus of 5,642 papers across mining, synthesis-plan generation, optimization, and wet-lab validation. Evidence-first mining consistently reduces unsupported extraction across four LLMs, while literature-derived post-training improves synthesis-plan quality, increasing the evaluator score from 0.51 (Base) to 0.62 (SFT) and 0.69 (DPO), and enhancing closed-loop optimization. Finally, automated wet-lab experiments translate OneMat-generated synthesis plans into functional catalysts, with the best-performing design discovered through limited experimental exploration achieving an overpotential of 155 mV at 10 mA cm and only a 21 mV increase after 500 CV cycles. Together, these results establish a route from scientific literature to experimentally validated synthesis-oriented generative optimization for materials design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative Optimization, Materials Design, Large Language Models, Preference Learning, Scientific Literature Mining, Synthesis Planning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RftVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The proposed method, OneMat, is a literature-driven framework that treats executable synthesis plans as optimization targets. It mines 5,642 acidic OER records with evidence-first extraction, adapts reasoning model via SFT and DPO on literature-derived preference pairs, and refines synthesis plans through an evaluation-revision loop. Automated wet-lab validation showed the practical potential of the proposed framework.

Critical strengths
  1. This paper tackles an important problem in physical and chemical sciences. The proposed framework directly targeted the problem to generate executable synthesis plans.

  2. The authors demonstrated an end-to-end pipeline from literature mining to automated wet-lab synthesis.

Critical weaknesses
  1. Core results rely entirely on an LLM-based surrogate evaluator, not human expert evaluations or physically-grounded simulation. Furthermore, the configuration of the evaluator model is not clear. Therefore, it remains unclear whether the reported score improvements reflect genuinely more valid or better-performing synthesis plans.

  2. Wet-lab validation reports only the single best case. The number of executed designs, success rate, and failure cases are missing, so framework-level effectiveness remains unverified.

  3. The demonstration of an end-to-end pipeline from literature mining to automated wet-label validation is interesting. However, technical contribution is limited in terms of the criterion of ICLR, because the framework mainly combines existing techniques.

Questions for authors

See the Weaknesses section.

Reviewer_E5afOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a literature-based LLM agent for proposing and refining synthesis protocols, and the authors validate it on acidic OER (oxygen evolution reaction) catalyst design tasks with wet lab synthesis. The agent framework extracts structured synthesis information from literature by prompting LLMs to output evidence before the extracted information. The authors use the extracted information from literature to fine-tune a Qwen3-8B model that suggests synthesis routes using SFT and DPO, and also build a database for retrieval of synthesis information during inference. Experiments assess extraction quality, single shot generation, and iterative refinement, reporting improvements in evaluator (LLM-based and rule-based) scores following post-training. The authors also report wet lab synthesis and performance measurement results at the end.

Critical strengths

1. Practical synthesis planning with wet lab demonstration

The paper addresses the practically relevant task of synthesis protocols and demonstrates the execution of selected protocols through wet lab synthesis and electrochemical measurements.

2. Literature-derived supervision for synthesis generation and preference learning

Using literature-reported pairs of baseline and preferred catalysts, the authors construct DPO training data without additional human annotation.

Critical weaknesses

1. Final suggested synthesis protocol for wet lab validation almost same with the one in the DPO training data

The proposed and wet lab validated final synthesis recipe is almost same with the one in the CORE_OER dataset, that is used for DPO training of the Qwen model. The training data ( 718_merged.pdf of core_oer_comparative_records.json in SI) already shows 154 mV of overpotential, and has almost the same synthesis protocol. This raises the question of whether the LLM agent proposes new promising synthesis protocols or reproduces existing protocols. This limits what can be concluded from the wet lab validation. Reproducing an already documented high-performing recipe does not establish the added value of the proposed post-training and iterative refinement framework over literature retrieval alone.

2. Insufficient validation of the hybrid evaluator

The reported gains of SFT, DPO, iterative refinement in Figure 1 are measured by a hybrid evaluator that combines LLM-based judgements of activity, stability, novelty, cost, and justification with rule-based checks of synthesis feasibility and completeness. The authors state that its relative importance weights were calibrated against expert assessments and checked against wet lab results, but the detailed description of it and quantitative correlations or agreement statistics are absent. It is unclear whether higher scores actually mean better synthesis protocols or simply better alignment with the evaluation criteria. Together with the weakness 1, this makes it difficult to assess the effectiveness of the proposed method.

Questions for authors
  1. Was the CORE_ORE record 718_merged.pdf included in DPO training? Please provide the data split of DPO and SFT training.

  2. Please provide details of the calibration of the hybrid evaluator.

My main concern is what the framework demonstrates beyond retrieving known synthesis protocols. SFT and DPO improve scores from a hybrid evaluator, largely LLM-based, while the best wet lab result closely reproduces a recipe already present in the supplied dataset. What evidence distinguishes the benefit of the proposed post-training from literature retrieval alone?

Reviewer_3hg1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

OneMat studies literature-driven generation and optimization of executable material synthesis plans, with acidic oxygen evolution catalysts as its application. It first extracts source evidence and structured synthesis knowledge from a corpus of 5,642 papers. A generator is then trained through supervised fine-tuning and preference optimization using literature-derived comparisons, and iteratively revises plans using retrieval and a hybrid evaluator. Evaluation includes extraction-order ablations, held-out plan generation, iterative refinement and automated wet-lab validation. The reported evaluator score increases from 0.51 to 0.69; the best validated catalyst achieves 155 mV overpotential at 10 mA/cm², increasing by 21 mV after 500 CV cycles. Wet-lab measurements support validation and calibration, rather than providing online optimization feedback.

Critical strengths
  1. The synthesis-planning problem is important.
  2. Traceable evidence and within-paper experimental preferences are useful design choices.
  3. Extraction-order ablations and concrete experimental procedures strengthen the submission.
Critical weaknesses
  1. Most evaluation weight depends on LLM judgments, without numerical calibration or independent validation.
  2. Wet-lab reporting lacks attempted-design counts, failures, independent repeats, raw curves, characterization results and matched controls.
  3. Internal generator comparisons do not establish superiority over strong retrieval or equal-budget search.
  4. The stated Ru ratio needs reconciliation with precursor quantities.
Questions for authors
  1. What are the evaluator’s identity, calibration size and independent expert correlations?
  2. Can all experimental outcomes and matched controls be provided?
  3. Does the 3:2 ratio describe measured composition?
  4. Do improvements persist against equal-budget RAG, best-of-N and incumbent-preserving refinement?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.