An Open-Source Data Infrastructure for Alloy Synthesis Research: Stage 3 — Retrieval-Augmented Literature Advice and CALPHAD Feasibility Screening

Stages 1 and 2 of this project built a database, a composition-screening front end, andvalidated characterization pipelines for a single alloy-synthesis laboratory. Stage 3 adds thetwo pre-synthesis advisory components promised in the Stage 1 paper: a retrieval-augmentedliterature advisor and a CALPHAD-based thermodynamic module. The advisor retrievespassages from a locally indexed corpus of the laboratory’s materials-science books (72530 textchunks from 96 books in three topical collections) and generates guidance with a locally hostedlanguage model, isolated from the main application behind a JSON subprocess boundary. Abenchmark of 39 questions with known answer passages showed that the default embeddingmodel found the right passage among the top five for only 10% of the questions, because itreads only the first 256 tokens of each chunk and was trained on English only; embeddingthe same chunks with a long-context, multilingual model raised this to 51% and reducedthe wrong or missing answers from 19 to 7. Every answer whose context held the rightpassage was correct. Asking the Russian-source questions in English exposed a languagegap: from the advisor’s English queries the Russian books were practically unreachable. ARussian version of the query, searched among the Russian passages, now adds one passagefrom them; with it, five of seven such questions were answered correctly, four of them from theRussian source. The thermodynamic module selects and merges published binary assessmentsfrom the open SGTE collection, computes ternary isothermal sections on a 19701-pointcomposition grid with pycalphad, and serves them through a nearest-neighbour index toa web advisor, so that a phase query no longer waits for a live equilibrium calculation ofseveral seconds per temperature. Because a description merged from binaries cannot containa genuinely ternary phase, the module also includes a feasibility screen that asks how stablesuch a phase would have to be to appear. For the Nd2Fe14B composition, the screen placesthat threshold at a formation energy of about−7.0 to−9.2 kJmol−1 atom−1 between 800and 1500K, 0.2–0.8 kJmol−1 atom−1 more negative than with the binaries of two full ternaryassessments. Combined with the SGTE binaries, the Nd2Fe14B phase of either assessmentcomes out stable over most of its known range, but by at most 0.6 kJmol−1 atom−1. As inthe earlier papers, we emphasize architecture, validation and the engineering defects foundwhile building the system. The most consequential was in the merging of binary descriptions,which silently removed iron from the liquid and the iron solid solutions and counted sharedparameters twice; every thermodynamic result looked plausible until the merged descriptionwas checked against the binary files it was built from.