🔬 materials science

The Precursor Genome: A Pairwise Reaction Dataset for Solid-State Synthesis

This paper introduces the Precursor Genome, a FAIR-compliant dataset of 1,035 autonomously generated solid-state reactions with fully traceable experimental metadata and validated X-ray diffraction outcomes, designed to enable data-driven and machine-learning approaches in synthesis science.

Original authors: Lauren N. Walters, Matthew J. McDermott, Bernardus Rendy, Yuxing Fei, Kristin A. Persson, Gerbrand Ceder

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lauren N. Walters, Matthew J. McDermott, Bernardus Rendy, Yuxing Fei, Kristin A. Persson, Gerbrand Ceder

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to bake the perfect cookie, but instead of flour and sugar, you're mixing weird powders like rust, chalk, and metal oxides to create brand-new materials. For decades, scientists have been baking these "inorganic cookies" one by one in their kitchens, writing down only the recipes that worked. If a batch turned into a sad, unreacted lump, they usually threw it in the trash and forgot about it. This meant that if you wanted to learn how to bake better, you only had a list of "success stories," missing all the crucial lessons from the failures.

Enter the Precursor Genome, a massive new cookbook created by a team of scientists and a robot chef named A-Lab.

The Robot Chef and the Great Cookie Experiment

The A-Lab is a "self-driving laboratory," which is basically a kitchen run by a super-smart robot arm that never gets tired, never forgets a step, and never hides a failed batch. The researchers wanted to see what happens when you mix different pairs of starting powders (called precursors) together. They picked 46 different powders containing 39 different elements—think of them as the ultimate pantry of ingredients for solid-state chemistry.

They didn't just mix a few; they cooked up 1,035 unique pairwise reactions. That's over a thousand experiments! And here's the kicker: they didn't just look at the winners. They recorded every single outcome, including the ones that didn't work, the ones that turned into weird intermediate shapes, and the ones that just sat there doing nothing.

The "Black Box" of Failure

Usually, when a scientist tries to mix two powders and nothing happens, they might write, "It didn't work," and move on. But the A-Lab is different. It acts like a forensic detective. For every single reaction, it measured:

Exactly how hot the oven got and how long it stayed there (thermal profiles).
How much powder went in and how much came out (masses).
What the robot's "eyes" (X-ray diffraction scanners) saw inside the mixture.

They ran 1,351 X-ray scans to peek inside the mixtures. It's like taking an X-ray of a cookie to see if the chocolate chips actually melted or if they're just sitting there as hard rocks.

The Magic Decoder Ring (Dara)

Once the robot finished baking, it had a mountain of X-ray pictures. To make sense of them, they used a special software tool called Dara. Think of Dara as a super-fast translator that looks at the X-ray patterns and tries to guess, "Oh, this squiggly line means we made silver oxide, and this bump means some of the starting powder is still there."

The robot made 1,950 different guesses (refinement cases) about what was in the mixtures. But robots can be overconfident, so human experts stepped in to double-check the robot's homework. They graded every guess on a three-tier scale:

Excellent: The robot nailed it; the fit is perfect.
Good: The robot got it mostly right, maybe missing a tiny peak here or there.
Poor: The robot was way off, or a big chunk of the mixture is still a mystery.

The humans verified these grades, and if two experts disagreed, a third one acted as the referee to make the final call. This ensures that the data is trustworthy.

What Did They Find?

The results are organized into a giant, colorful map (Figure 1 in the paper) that shows exactly what happened for every pair of ingredients. They found four main types of outcomes:

Completely Reacted (Emerald): The powders mixed perfectly to form the new target material. No leftovers!
Partially Reacted (Blue): They made the target material, but some of the original powders were still hanging around.
Transformed (Amber): The powders changed, but not into the target. Maybe one turned into a different version of itself (like losing water or oxygen).
Unreacted (Gray): Nothing happened. The powders just sat there.
Physical Failure (Pink): Sometimes, the robot couldn't even handle the sample after heating because it turned into a gooey liquid or a sticky clay.

Why This Matters

This isn't just a list of recipes; it's a FAIR dataset (which means it's Findable, Accessible, Interoperable, and Reusable). The team didn't just keep this in a drawer; they put it all into a digital ledger (a structured JSON file) that anyone can download.

They explicitly ruled out the idea that we can just guess how materials will form based on old, messy literature. The paper argues that without this kind of clean, machine-readable data that includes the "failures," we can't teach computers to predict new materials. The paper doesn't claim to have solved the mystery of all materials; instead, it provides the benchmark—the gold standard test set—that future AI models need to learn from.

The Takeaway

The Precursor Genome is like giving the world a complete, honest diary of a robot chef's first year in the kitchen. It shows 1,035 attempts, from the perfect cookies to the burnt lumps, with every single detail recorded. It proves that to teach a computer how to invent new materials, we need to show it the whole story, not just the highlight reel. And the best part? You can go download the whole cookbook right now and start cooking up your own predictions.

Technical Summary: The Precursor Genome

Problem Statement
Solid-state reactions remain the primary method for synthesizing inorganic materials, yet the field lacks a large-scale, machine-readable dataset that reports experimental protocols and outcomes with consistent provenance. Existing literature is biased toward successful routes, often omits negative trials, and fails to preserve the raw sample-handling logs and characterization data necessary for true reproducibility. This data gap obstructs the application of first-principles, data-driven, and machine-learning approaches to synthesis science. While computational methods can predict millions of thermodynamically stable compounds, their experimental realization remains a bottleneck due to the absence of systematic, high-fidelity synthesis data.

Methodology
To address this, the authors utilized the A-Lab, an autonomous self-driving laboratory, to generate the "Precursor Genome," a dataset of 1,035 pairwise solid-state reactions.

Experimental Design: The study involved 46 common inorganic precursors spanning 39 elements, with a focus on materials relevant to electrochemical energy storage. Reagents were sourced from Sigma-Aldrich (generally ≥99% purity). Six precursors were excluded due to hygroscopicity or safety concerns.
Synthesis Protocol: Reactions were performed in stoichiometric ratios targeting hypothetical binary products (1:1 metal cation ratio). The A-Lab executed a three-stage workflow: precursor preparation (dispensing, weighing, slurry mixing with ethanol, drying), heating, and sample recovery.
- Heating: The dwell temperature ( $T_t$ ) was determined by Tammann's rule ( $T_t \approx \frac{2}{3} \min(T_{m,1}, T_{m,2})$ ), rounded down to the nearest 100°C, with bounds between 200°C and 1,100°C. A default profile included heating to 300°C, ramping to the target temperature, a 1-hour hold, and uncontrolled cooling.
- Characterization: Recovered powders were analyzed via powder X-ray diffraction (XRD) using a Malvern Panalytical Aeris diffractometer (Cu K $\alpha$ , 10–100° 2 $\theta$ ). Precursors were also characterized via Scanning Electron Microscopy (SEM) and Energy-Dispersive X-ray Spectroscopy (EDS).
Data Processing & Validation:
- Automated Analysis: XRD patterns were processed using the Dara framework, an automated phase-identification and refinement pipeline built on the BGMN kernel. Dara ranks candidate phase sets from the Inorganic Crystal Structure Database (ICSD) and Crystallography Open Database (COD) based on goodness-of-fit metrics (Rwp).
- Human Verification: Every refinement was independently validated by human experts on a three-tier quality scale (1: excellent fit; 2: good fit with minor discrepancies; 3: poor fit). Disagreements were arbitrated by a third expert.
- Data Structure: The dataset is organized as a hierarchical, Pydantic-validated JSON ledger (ledger_precursor_genome.json). It links precursor properties, synthesis metadata (thermal profiles, masses), raw XRD scans (1,351 total), and 1,950 refinement cases.

Key Contributions and Results
The paper presents the Precursor Genome as a comprehensive, FAIR (Findable, Accessible, Interoperable, Reusable) benchmark for solid-state synthesis.

Dataset Scope: The dataset contains 1,035 reaction outcomes involving 46 precursors. It explicitly includes "negative" data (failed or slow reactions) and intermediate phase formations, captured via short 1-hour dwell times.
Reaction Outcomes: Expert-validated outcomes are categorized into four controlled vocabulary terms:
1. Completely reacted: Target phase formed with no detectable precursors.
2. Partially reacted: Target phase formed alongside residual precursors.
3. Transformed: Precursors decomposed or reacted to form variant phases, but the target phase did not form.
4. Unreacted: Only original precursor phases observed.
  (Physical failures, e.g., liquid samples, are also recorded).
Data Accessibility: The full dataset, including raw .xrdml files, serialized refinement objects (.pkl), and interactive plots, is available on Zenodo and GitHub under a CC BY 4.0 license. The data structure is defined by a Pydantic schema, ensuring machine-readability and traceability from precursor pair to final phase assignment.

Significance
The authors claim that the Precursor Genome establishes a necessary foundation for training and evaluating predictive models of solid-state reactivity. By providing a transparent, machine-readable map of synthetic accessibility that includes both positive and negative outcomes, the dataset aims to bridge the gap between computational predictions and experimental realization. It serves as a standardized benchmark to overcome the reporting biases of traditional literature, enabling the development of autonomous, data-driven workflows for materials discovery. The work does not propose new synthesis algorithms or specific future applications but rather provides the essential infrastructure (data and schema) required for the community to develop such tools.