Paper: *Beyond Structure: Revolutionising Materials Discovery via AI-Driven Synthesis Protocol-Property Relationships* Author: Guillaume Lambard arXiv: 2605.00313 | 2026-04-29
1. The Materials Science Dilemma: "We Predicted the Structure but Can't Make It"
Current AI methods for discovering new materials:
- Predict crystal structures
- Compute properties
- Identify thousands of candidates
- But:
- Can they be synthesized?
- Nobody knows
- The "synthesizability gap"
- Structures are predicted
- Yet labs can't make them
- A structure-centric paradigm
- Focused only on atomic configurations
- Ignoring:
- How to synthesize?
- What conditions are needed?
- What are the steps?
- Predictions → unverifiable
- Not applicable in practice
- Not only predicting structures
- But also predicting synthesis methods
- Executable synthesis protocols
- From "what it is" to "how to make it"
- Not natural language
- But structured, executable formats
- Machine-comprehensible
- Machine-executable
- Like standardized "recipes"
- Generate synthesis protocols
- Given target properties, inverse design answers:
- "I want this property"
- "What synthesis method should I use?"
- Actionable plans
- Not structure-property
- But synthesis-property
- Directly linked
- Skipping the intermediate "structure" step
- More practical
- Traditional approach = a food app that tells you "this dish is delicious"
- But never says how to cook it
- You can see it but never taste it
- New approach = a food app that gives you the complete recipe
- Ingredients
- Steps
- Temperature
- Time
- You can actually make it
- Ideal structures are predicted
- But no one knows how to build them
- Labs can't realize them
- They stay on paper
- Can't directly guide experiments
- Extra work is required
- Inefficient
- Directly provides the synthesis method
- Labs can run it
- Verifiable
- From prediction to experiment
- One direct step
- Efficient
- Predict → synthesize → test → feedback
- A complete cycle
- Continuous improvement
- Knowledge → action
- Theory → practice
- Prediction → realization
The problem:
What's needed:
2. The Synthesis-First Paradigm: Three Pillars
The paper proposes a Synthesis-First paradigm:
Core idea:
> Treat executable synthesis protocols as the primary design variable, not just atomic configurations.
Three pillars:
1. Machine-readable representations of synthesis protocols
2. Generative and inverse-design models
3. Synthesis-property relationships
An analogy:
3. Why Synthesis Protocols Matter More Than Structures
Problems with the structure-centric view:
The synthesizability gap:
Poor practicality:
Advantages of Synthesis-First:
Executable:
Practical:
Closed loop:
5. A Feynman-Style Judgment: Knowing "What" Beats Not Knowing "How"
Feynman famously noted that knowing the name of something and truly understanding something are entirely different.
In materials science:
> "Predicting a material's crystal structure is 'knowing the name'; predicting how to synthesize it is 'true understanding'. The insight of the Synthesis-First paradigm is that the ultimate goal of science is not to describe the world but to change it — and 'how' is the key to changing it."
This reflects the essence of applied science:
6. Takeaways
If you work on AI for science or materials discovery, ask yourself:
1. "Are my predictions verifiable?" 2. "Is synthesis feasibility considered?" 3. "Is the path from prediction to experiment clear?" 4. "Are executable protocols more valuable than abstract structures?"
This paper reminds us: the endpoint of scientific discovery is not "knowing" but "doing."
As AI materials science shifts from "structure prediction" to "synthesis prediction," it moves from being a theorist to being an engineer. In the future of materials discovery, the best AI is not the one that predicts most accurately, but the one that best guides experiments.
In the laboratory of science, the most precious prophecy is an executable protocol.