Research map / Label-efficient and robust learning / Recognition and long-tail
Retrieval-Retro: Retrieval-based Inorganic Retrosynthesis with Expert Knowledge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Retrieval-Retro, a retrieval-augmented model for inorganic retrosynthesis (precursor prediction) that implicitly extracts precursor information from retrieved reference materials via attention, using two retrievers (MPC and a thermodynamics-based NRE retriever).
Problem
Predicting precursor sets for inorganic materials from composition alone, especially discovering novel synthesis recipes with precursors not seen in reference materials, and selecting reference materials with thermodynamically favorable precursors.
Approach
Instead of copying precursor sets of retrieved reference materials, the model implicitly extracts their synthesis information through self-attention (over references) and cross-attention (target-to-reference) layers; a Neural Reaction Energy retriever selects references by predicted reaction enthalpy (pretrained on Materials Project DFT energies, fine-tuned on experimental formation energies), com
Key results
- On the year split, Retrieval-Retro reaches 61.16% Top-1 and 69.45% Top-10 exact match accuracy, versus 60.01% and 67.19% for the strongest baseline (Graph Network + MPC).
- On new-precursor (new case) test materials, Top-1 accuracy improves from 17.63% (Graph Network + MPC) to 20.15%.
- On the random split, Retrieval-Retro reaches 79.81% Top-1 versus 79.09% for Graph Network + MPC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Retrieval-Retro instead of Graph Network + MPC | Inorganic retrosynthesis prediction · Year split | Top-1 accuracy: 60.01 -> 61.16 | Helped |
| Use Retrieval-Retro instead of Graph Network + MPC | Inorganic retrosynthesis prediction · Random split | Top-1 accuracy: 79.09 -> 79.81 | Helped |
| Implicitly integrate retrieved precursor information rather than directly using it | Graph Network + MPC · Year split, new case | Top-10 accuracy: 26.22 -> 30.15 | Helped |
| Add NRE retriever (ablation) | Retrieval-Retro with MPC · Year split, new case | Top-10 accuracy: 30.15 -> 31.56 | Helped |
| Add NRE retriever (ablation) | Retrieval-Retro with MPC · Year split, subset case | Top-1 accuracy: 65.07 -> 66.00 | Helped |
| Use both trained retrievers instead of random retrieval (ablation) | Retrieval-Retro · Year split | Top-1 accuracy: 58.42 -> 61.16 | Helped |
| Add NRE retriever (ablation) | MPC-only Retrieval-Retro · Year split | Top-1 accuracy: 59.96 -> 61.16 | Helped |
| Add MPC retriever (ablation) | NRE-only Retrieval-Retro · Year split | Top-1 accuracy: 60.28 -> 61.16 | Helped |
| Add self-attention to the reference representations (ablation) | Retrieval-Retro with cross-attention · Year split | Top-1 accuracy: 60.34 -> 61.16 | Helped |
| Pretrain the formation-energy predictor on DFT data before experimental fine-tuning (ablation) | NRE formation-energy predictor · Experimental formation energy data | MAE (eV/atom): 0.0761 -> 0.0643 | Helped |
| Add MPC retrieval with explicit precursor use (ablation) | Graph Network · Year split, subset case | Top-1 accuracy: 63.98 -> 65.01 | Helped |
| Add MPC retrieval with explicit precursor use (ablation) | Graph Network · Year split, new case | Top-10 accuracy: 27.93 -> 26.22 | Hurt |
Limitations
- Ignores synthesis conditions such as temperature, time, pressure, and pairwise precursor reactions.
- Reaction energy is estimated from composition-based formation energy predictions rather than measured data, and experimental formation energy data is limited (1,637 entries).
- No uncertainty estimation is provided.
Experimental setup
Datasets: text-mined inorganic synthesis recipes (SynthesisSimilarity) Materials Project DFT formation energies experimental formation energy data (qmpy thermodata)
Benchmarks: inorganic synthesis recipe dataset (year split) inorganic synthesis recipe dataset (random split)
Metrics: Top-K exact match accuracy macro recall micro recall
Baselines: Composition MLP MPC retrieval method ElemwiseRetro Roost CrabNet Graph Network Graph Network + MPC
Builds on: MPC retriever (retrieval-based inorganic retrosynthesis) Graph Network (MEGNet-style) ElemwiseRetro Roost
Closest papers by meaning
- RECIPE: Learning to Rank Complete Precursor Sets for Inorganic Retrosynthesis (NeurIPS 2026 · similarity 0.92)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.89)
- URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment (NeurIPS 2026 · similarity 0.86)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.86)
- Retro-fallback: retrosynthetic planning in an uncertain world (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.