A Synthesizability-Guided Pipeline for Materials Discovery

Thorben Prein, Willis O'Leary, Aikaterini Flessa Savvidou, Elchaïma Bourneix, Joonatan E. M. Laulainen

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.

Introduction: Bridging the Synthesis Gap

Recent advancements in AI and high-throughput computing have populated databases with millions of theoretically "stable" inorganic crystals. However, a significant "synthesis gap" persists: the number of predicted materials far exceeds those known to exist experimentally. Traditionally, researchers rely on Density Functional Theory (DFT) to assess stability via the "convex hull" filter. While useful, DFT calculations typically reflect zero-Kelvin conditions and overlook kinetic barriers, temperature effects, and precursor availability. Consequently, many "stable" theoretical structures remain unsynthesizable, while some "unstable" ones are common. This work introduces a "synthesizability" filter to prioritize candidates that are practically viable for laboratory synthesis.

Methods: A Unified Prioritization Framework

The researchers developed a three-stage, synthesis-aware pipeline to bridge the gap between computational prediction and experimental realization.

Overview of the synthesizability-guided pipeline and the filtering process from millions of candidates to experimental targets.
  1. Synthesizability Modeling: The core of the pipeline is a dual-encoder model. A Composition Encoder (transformer-based) captures chemical signals like redox constraints, while a Structure Encoder (Graph Neural Network) evaluates local coordination and motif stability. These scores are combined using Borda fusion to rank candidates.
  2. Synthesis Planning: For top-ranked candidates, neural models (Retro-Rank-In and SyntMTE) suggest optimal precursors and furnace temperatures based on literature-mined recipes.
  3. Automated Execution: The final stage involves an automated solid-state laboratory that performs weighing, grinding, and heating, followed by automated X-ray diffraction (XRD) for product verification.

The pipeline was applied to a pool of 4.4 million structures from the Materials Project, GNoME, and Alexandria databases.

Key Results and Performance

The ensemble model demonstrated superior performance compared to traditional baselines, such as the 50 meV convex-hull heuristic and SynthNN.

Performance metrics comparing the proposed ensemble model against standard baselines.

The model achieved a precision of 0.764 and an F1 score of 0.789. In experimental trials, the team characterized 16 samples and successfully synthesized 7 targets (44%). Notably, this included the discovery of $Nd_3BTeO_9$, a novel compound with no previously reported crystal structure, and an unreported orthorhombic polymorph of $Eu_2WO_6$.

Discussion: Insights and Future Directions

The study revealed that while composition models offer high recall, structural data is crucial for precision, filtering out unrealistic geometries. Interestingly, several materials labeled as "theoretical" in databases were successfully synthesized, suggesting that AI can help correct omissions in current experimental records.

The primary advantage of this pipeline is speed; the transition from millions of candidates to physical products took only 72 hours. However, limitations remain. The model may exhibit "Positive-Unlabeled" bias, potentially struggling with exotic structures requiring non-standard synthesis. Currently, the pipeline is optimized for solid-state synthesis and does not yet cover hydrothermal or high-pressure methods. Future work will focus on expanding the chemical space and synthesis techniques covered by the model.

Summary updated

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.
BibTeX
@inproceedings{prein2025synthesizability,
  author = {Thorben Prein and Willis O'Leary and Aikaterini Flessa Savvidou and Elchaïma Bourneix and Joonatan E. M. Laulainen},
  title = {{A Synthesizability-Guided Pipeline for Materials Discovery}},
  booktitle = {AI4Science @ NeurIPS},
  year = {2025},
  url = {https://openreview.net/forum?id=2w7VeQp1MX},
}
Figure