A preprint describes PIRAG-LM, a language-model framework for end-to-end planning of synthesis routes for inorganic crystals. The system is presented as interpretable: it uses retrieved evidence to support a proposed route, rather than relying only on a model’s free-form generation. Its underlying SSKB contains 13,820 experimentally synthesized inorganic crystals, with literature-derived synthesis information organized as route-level records.
For a proposed material, the framework searches for related precedents across chemical, structural and thermodynamic spaces. It then uses a structured language-model step to assemble and score a route, including the synthesis method, precursors and operating conditions. In practical terms, the model is trying to connect a material’s composition and structure with routes that have already been used in the literature.
A time-separated test
To test how that process handled records separated in time, the retrospective evaluation used historical synthesis records and a 2021–2026 time-axis set. In that evaluation, the system correctly predicted the synthesis method for 41 of 47 materials. Specialized routes were the main source of method errors.
Strong results, with conditions attached
Precursor performance was assessed only after the method prediction was correct. In those 41 cases, at least 50% precursor coverage was reached in 36 cases, or 87.8%, while the core precursor set was usable in 37 cases, or 90.2%. The figures therefore describe the quality of precursor suggestions after a successful method step, not all 47 materials at once.
Condition estimates were reported in the eligible prediction subset. Temperature prediction had a mean absolute error of 91.8 kelvins—the average size of the miss—and pressure prediction had a mean absolute error of 11.47 gigapascals. The pressure result was dominated by ultrahigh-pressure hydrides, showing where a single average can conceal difficult specialist cases.
The benchmark numbers were sequential
A separate, strictly separated benchmark of 1,000 materials produced higher headline percentages. The method was correct for 940 materials, or 94.0%. Of those 940, 913 had usable precursor sets, or 97.1%; of those 913, conditions were correct for 879, or 96.2%. These are sequential conditional results: each later percentage starts with the cases that passed the previous stage.
Retrieval changed the scores
Performance also changed with the retrieval setup. In the reported ablation, the full CS+SS+TS framework—covering chemical, structural and thermodynamic spaces—reached 87.2% accuracy, compared with 72.1% for the LLM-only baseline. An enhanced configuration added pressure-dependent information and thin-film records and reached 91.4% accuracy. The study reports these as configuration comparisons, so they show different performance levels without establishing that one setup caused the difference.
From predictions to diffraction
The researchers then used the framework in prospective experiments. Measured powder X-ray diffraction, or PXRD, patterns were consistent with the calculated Bragg positions for every target, supporting formation of the target phases. That result tests whether the predicted work produced the expected crystal phase; it does not establish the functional properties of those compounds.
Interpretability was shown through an illustrative evidence chain. Retrieved records were connected to an 84-out-of-100 synthesizability assessment, proposed sol-solid conditions and identified experimental risks. The example provides a way to trace a recommendation back to supporting records, but it was not independent validation of every model decision.
What the results do not settle
The boundaries matter. Method errors were concentrated in specialized routes, pressure errors in ultrahigh-pressure hydrides, and the temperature estimate remained sensitive to precursor activation and high-pressure heating protocols. The precursor and condition results were also conditional on earlier stages succeeding. Together, those caveats mean the reported accuracy describes performance on the tested literature-derived records, not universal reliability across inorganic synthesis.
The document identifies itself as an arXiv v1 preprint dated 26 August 2026. It reports that the SSKB and curated route-level synthesis records, along with the PIRAG-LM code and analysis scripts, are openly available through the stated GitHub repository, with an online platform also provided.
Paper data and sources
Original title: Interpretable physics-informed retrieval-augmented generation language model for end-to-end inorganic crystal synthesis planning
Authors: Wei-Jian Jiang, Ye-Nan Sha, Hui Guo et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text