A Synthesizability-Guided Pipeline for Materials Discovery

TL;DR

The paper tackles the bottleneck of translating predicted inorganic crystal structures into experimentally realizable materials by introducing a unified synthesizability score that fuses composition and crystal-structure signals. It builds a dual-encoder model that outputs $s_c(x_c)$ and $s_s(x_s)$, which are combined via a rank-average ensemble to prioritize candidates for synthesis planning and execution, followed by two-stage precursor planning and calcination-temperature prediction. Evaluated on ~4.4 million structures, the approach identifies highly synthesizable candidates and guides automated solid-state synthesis experiments, yielding 7 successful syntheses out of 16 targets (44%). This work demonstrates that integrating synthesis-aware screening with experimental workflows can reduce time-to-discovery and reveals practical gaps in current materials databases, highlighting the method's potential to accelerate discovery of novel compounds.

Abstract

Computational materials discovery relies on the generation of plausible crystal structures. The plausibility is typically judged through density functional theory methods which, while typically accurate at zero Kelvin, often favor low-energy structures that are not experimentally accessible. We develop a combined compositional and structural synthesizability score which provides an accurate way of predicting which compounds can actually be synthesized in a laboratory. We use it to evaluate non-synthesized structures from the Materials Project, GNoME, and Alexandria, and identified several hundred highly synthesizable candidates. We then predict synthesis pathways, conduct corresponding experiments, and characterize the products across 16 targets, successfully synthesizing 7 of 16. The entire experimental process was completed in only three days. Our results highlight omissions in lists of known synthesized structures, deliver insights into the practical utility of current materials databases, and showcase the central role synthesizability prediction can play in materials discovery.

Figures (15)

• Figure 1: Our pipeline for selecting synthesizable candidates, predicting their recipe, carrying out the synthesis, and characterizing their structure.
• Figure 2: Test-set performance of the compositional, structural, and ensemble models. (a) Precision, recall, and F1 scores for identifying the target "theoretical" labels in the materials project. (b) ROC curves (AUC shown in the legend). The convex-hull baseline classifies compounds within $50\,\mathrm{meV}$ above the hull as synthesizable.
• Figure 3: a. Identified highly synthesizable ($\mathrm{RankAvg}>0.95$) targets and the results of performed experiments. For Eu$_2$WO$_6$, a non-target polymorph was synthesized that had the same space group as a known structure but with significantly different lattice parameters. For Sr$_4$Al$_6$MoO$_{16}$, a compound with the target composition was synthesized. However, due to low purity we could not rule out formation of one of the three known polymorphs. Several compounds, while listed as theoretical in Materials Project, do have similar structures reported, mostly in the ICSD. b. XRD scan, fit, and difference for Nd$_3$BTeO$_9$. A second unknown phase is present and accounts for <5% of total intensity. c. Refined hexagonal unit cell for Nd$_3$BTeO$_9$. The unit cell is hexagonal with the lattice parameters $a=8.761$ Å and $c=5.553$ Å
• Figure 4: Joint principal component analyses on MTEncoder embeddings shown in the same latent space for a. Materials Project, b. GNoME, and c. Alexandria. Top row: density of structures by embedding. Middle row: average synthesizability by embedding. Bottom row: distribution of rank-average synthesizabilities within each dataset.
• Figure 5: Synthesizability by compound class. Computed from entries in the Materials Project.